Evaluating AI Integration Maturity in Architectural Practices in the Kurdistan Region, Iraq: A Comparative Benchmark Study

Abstract

Artificial Intelligence (AI) is being integrated into the architectural profession via processes such as generating images based on design parameters; providing support to help create written descriptions of designs; assisting in the visualisation of buildings before they are built; allowing for greater use of parameters in construction; providing documentation to help complete projects in record time; and assisting in making design decisions. The success of integrating AI into practice depends not only on the use of tools but also on firms’ maturity in integrating AI into their workflows, employees’ capabilities, project teams’ operational efficiency, the impact of investment decisions, and the overall governance of the profession. This research evaluates the maturity of AI integration in architectural practice in the Kurdistan Region of Iraq. To do so, it develops and operationalises the AI Integration Maturity Index (AIMI), an eight-component formative composite index scored 0–37 and organised into five maturity bands (Non-adopter, Exploratory, Occasional, Integrated, and Advanced Strategic). The index comprises adoption, usage, diversity of tools, breadth of workflows, project penetration, staff involvement, training/capacity building, and governance/strategic focus. The AIMI is treated as a literature-derived formative diagnostic tool rather than a universal weighting standard, and was developed through a structured literature synthesis, expert pilot review, and internal-structure validation. Accordingly, its component logic and internal statistics are reported as a transparency and coherence check rather than as reflective reliability claims: Cronbach’s alpha (0.922) is presented descriptively to show component co-movement given the formative specification, while inter-coder reliability (kappa = 0.96) supports the qualitative benchmark coding. The study employs a mixed-methods descriptive comparative methodology consisting of a structured survey instrument administered to 100 architectural firms operating in the local market and structured asynchronous text-based interviews with 10 international architectural firms, comparing the resulting profiles with selected, generally accepted benchmarks and with a sample of leading firms engaged in architectural practice worldwide. On average, total AIMI scores in the local sample were 18.14 out of 37, indicating that local firms are broadly adopting and using AI (82% currently use AI on a regular or occasional basis). By contrast, the international sample yielded a mean AIMI score of 28.60, indicating that local firms exhibit a significantly lower level of maturity than the international benchmark, with the largest gaps in staff involvement, project penetration, and overall engagement with AI use. The paper concludes that the primary challenge facing architectural firms in the Kurdistan Region is no longer basic awareness or technological infrastructure, but rather the transition from broad and superficial AI adoption to a systematic, structured, project-based, and well-governed integration of AI within the architectural profession.

1. Introduction

AI has become increasingly relevant to the architectural design process by supporting conceptual design, representation, documentation, analysis, and collaboration within a developing practice environment [1,2,3,4]. The expansion of text-based assistants, image-generating systems, rendering tools with AI, parametric plugins, and decision support systems has transformed AI from being an abstract area of research to a current, practical issue for architectural practices. The question is not whether architectural practices will utilise AI in their work, but whether they will transition from one-off experimentation to full integration of the technology. The level of integration refers to the extent to which architectural practices understand, adopt, support, and incorporate AI into their workflows. Practices might use ChatGPT, Midjourney, DALL-E, Stable Diffusion, Gemini, V-Ray AI, D5 Render, Rhino or Revit plugins, or other systems; they may not be considered mature solely because of access to the technology. Mature architectural practices have defined roles for AI in the workflow, trained personnel, quality control, assigned responsibilities for reviews, investment patterns and governance policies. This distinction is particularly crucial in architecture since the use of AI could have significant implications in terms of how design is communicated, client expectations, authorship, confidentiality, technical accuracy, and ultimately, the professional responsibility of the architect.

The Kurdistan Region of Iraq provides a setting for assessing this issue. Architectural firms in Erbil, Sulaymaniyah, Duhok, Halabja, Soran, Zakho, Koya, Rania, and other cities in the Kurdistan Region operate within a growing building sector, with varying levels of digital capability, at different firm scales, with different client expectations, and with limited formal documentation on the adoption of AI. A number of these firms have been utilising CAD, rendering engines, Adobe products, and BIM platforms for visualisation and other forms of communication. The addition of AI creates another layer of support for generating ideas, producing and visualising graphic renderings, creating text, conducting research, analysis, and decision-making; however, this also requires more critical review and new operational routines within the firms.

The purpose of this paper is to investigate the maturity of AI integration within architectural design firms in the Kurdistan Region and to compare the local situation with international and other related benchmarking firms in terms of AI integration maturity. This is not intended to imply statistical comparability between the local respondents and the international study subjects. Rather, it will use local data obtained from a quantitative survey questionnaire and qualitative methods to produce international benchmark data through structured, text-based interviews. To establish connections between the two types of survey evidence, both surveys will be analysed through a shared maturity framework that respects the different methodologies of the two surveys. The primary research question is: What is the status of AI integration within architectural design firms in the Kurdistan Region, compared with international and other relevant benchmark firms? Related to that primary question, supporting questions include which AI tools are in use locally; how frequently they are used; what barriers exist for the adoption of the tools; what types of practices are exhibited among the international benchmark firms; where the largest areas of maturity disparity are between local and international firms; and what lessons learned may be realistically transferable from one context to another. The empirical contribution of this paper is to provide a regional evidence base for a professional domain in which AI adoption has received little direct formal examination. The methodological contribution of the paper establishes an AI Integration Maturity Index with eight components that facilitate comparative evaluation of adoption, usage intensity, tool diversity, breadth of workflow, staff involvement, capability, application and commitment. Three original contributions distinguish this study from prior work. First, it provides the first empirical, firm-level evidence base for AI integration maturity in the Kurdistan Region of Iraq, a professional context that has received no prior formal examination in the AI-in-architecture literature. Second, it introduces and operationalises the AIMI as a replicable, domain-specific maturity instrument for architectural practice, filling a gap left by generic technology adoption models that do not capture the process, governance, and capability dimensions specific to design-led professional services. Third, it demonstrates through comparative benchmark analysis the primary maturity gap between local and international firms, which this study is among the first to document and benchmark systematically. Figure 1 illustrates the structure of research design.

2. Literature Review

2.1. Architectural Practice and Digital Change in the Kurdistan Region

The architectural practice in the Kurdistan Region is often considered in terms of the buildings addressed, the forms of knowledge valued in design decisions, and the types of institutions that deliver them. In recent studies of Erbil and the broader Iraqi context, attention continues to focus on environmental performance, housing quality, and the gradual modernisation of design and delivery processes [5,6]. In addition, studies of BIM implementation in Iraq indicate that digital practices remain inconsistent across organisations, depending on organisational capacities, the level of coordination required to complete projects, and the lack of consistency in implementation standards [7,8]. General literature on digital architecture provides a language for describing the organisation of an office, the routine of coordinating work, and the method of adopting new digital tools, without specific references to localised rates of implementation [9,10].

Firm typologies are typically identified by their size, the types of services they offer, and the extent to which they separate design, documentation, and site responsibilities [8,10]. Small firms tend to have a lower capacity to formalise standards and provide formalised training than larger firms, which can offer more specialised roles and more formalised quality-control processes [7,9,11]. For AI adoption, however, this typology is broadened by the degree to which a firm adopts and applies technology in its operations, such as whether it uses technology for only occasional experimental projects or for its entire organisational operations and for a variety of different types of projects and types of tools [1,2,12,13].

Typical architectural workflows progress from concept design and specification through production, document development, and coordination, with continual input from the client and consultant [9,10,14]. The most frequently cited BIM standards in Iraq demonstrate that appropriate standards and model management resources are required to execute Digital Delivery consistently [7,8]. Tools that extend the Architectural workflow through ideation, visualisation, document production, text support, and performance-related reasoning have also been reported in the literature as part of an AI overview [3,5,6,13,15,16].

The limited availability of both skills and training capacity is consistently cited as a barrier to digital adoption in office settings, where very few formal systems are in place [7,11]. As outlined in AI reviews, upskilling, output verification, and responsibility assignment are critical preconditions for ensuring the safe adoption of AI [1,3,15,17]. Other factors, including time pressures, interoperability, approvals, and technical evidence requirements, affect whether AI is considered a regular and accepted method of service delivery rather than an experimental option [5,6,10,18].

2.2. AI and Generative AI in Architectural Practice

Artificial Intelligence is used in the building profession through computation and AI in ways that go well beyond simply automating processes; AI supports the activities of recognising, predicting, optimising, generating, classifying, and decision-making throughout the entire design and delivery processes [1,2,9,19]. Generative AI is a type of artificial intelligence system that produces output (new text, images, laid-out plans, or design representations) rather than merely analysing already provided input data [4,13,20,21]. This distinction matters because image, text, analytic, and workflow-support tools require different forms of checking, authorship control, and designer feedback [12,16,17]. A group of 23 professionals in architectural design management conducted a qualitative study that found that 83% were currently using AI in their daily work. The main uses for AI included improving the design process (i.e., design iterations), legal compliance with client briefs, and parametric workflows. Most of these professionals agreed that legal ambiguity, a lack of standards, and the reliability of AI outputs are significant barriers to further integrating AI into the design process. The study concludes that responsibly adopting AI will require strategic frameworks rather than tool-by-tool decisions [22].

AI use cases are commonly organised by design stage and output type, ranging from concept generation and visualisation to documentation, modelling support, coordination, and performance-related analysis [1,2,3,13]. Concept design and visualisation are the most visible areas, including prompt-based ideation, facade studies, mood imagery, sketch-to-architecture methods, and rapid presentation material [4,15,20,23,24]. A global evaluation of AI software tools has found that image-generation platforms are commonly used in the early stages of design. However, there is significant fragmentation in the AI tools landscape, as most tools focus on specific segments rather than offering a comprehensive design workflow [25]. More advanced uses connect generative outputs to parametric modelling, BIM, environmental reasoning, and project information management, where integration and verification demands are higher [5,6,7,8,18]. The empirical analysis of the six stages of AI design indicates that AI performs well during the early concept and final presentation phases, with a creativity score of 4.8 and an efficiency score of 4.6 at the concept stage. However, the measurements from the technical documentation stage show significant drops in AI effectiveness compared to human expertise and confirmation [26]. Figure 2 shows the AI taxonomy for architectural practice.

Much of what we read in academic literature regarding the benefits of AI relates to obtaining answers more quickly (or set of answers), having a broader set of possible answers to choose from, receiving assistance with menial tasks, improved communication, and, as it relates to the benefits of AI, the extent to which the task fits the organisation or how it is implemented [9,10,27,28]. However, the literature also consistently identifies limitations in verification efforts, training, authorship, traceability, bias, hallucination, and professional judgement, particularly when generative outputs must meet the project’s requirements [1,2,3,12,29]. These limitations suggest the need for written/documented quality control processes, rather than simply adopting an AI tool/circumstance [22,30,31]. Figure 3 presents AI use cases across architectural design stages.

2.3. Adoption, Integration, Maturity, and Socio-Technical Readiness

Organisations vary in their AI maturity and integration at different rates across teams, project stages, and individual tools, processes, and responsibilities [1,2,3,9]. AI encompasses many different types or families of tasks, making it difficult to measure organisational progress with a single maturity metric, such as whether an organisation has adopted AI [1,4,11,13]. Studies focused on workflow changes and workflow design processes have shown that organisations shift from trying AI to developing structured relationships between AI outputs and the deliverables, review cycles, and client communication associated with those deliverables [9,10,17,27]. Research on human–AI collaboration and data-driven design methods has highlighted the importance of accountability, validation processes, and interaction design, thereby indicating that AI maturity cannot be determined solely by an organisation’s access to AI [16,17,29].

The initial phase of adoption occurs when an organisation implements AI for the first time, using it for prototyping, visualisation, or text-support activities selected for testing [1,2,20,29]. Integration implies a greater depth of use by linking AI to routine work processes, team coordination, and project delivery rather than treating it as an optional tool [3,9,10,17]. Maturity extends further to encompass stable organisational procedures, clear role delineation, training, review processes, and accountability for AI outputs [3,12,16,27]. Progression from awareness through trials to repeatable governance can be described using maturity frameworks [33,34], whose indicators must be broad enough to capture tool breadth, workflow integration, staff training, review routines, and documentation standards [1,2,3,13,18].

The socio-technical view treats AI maturity as the relationship between people, processes, and technology. Architects’ readiness includes the necessary skills, learning routines, feedback practices and the individual responsible for the final decision [3,11,15,17]. Process readiness includes stable entry and exit points for AI outputs, review points, version control, and rules for utilising AI-assisted text, images, models, or analysis in project workflows [10,18,29]. Technology readiness includes the availability of tools, their compatibility with CAD, BIM, and visualisation settings, and the appropriate handling of data and outputs [1,2,23,24]. In this way, maturity is achieved when team members consistently use the necessary AI skills, routines, and tools across multiple projects. Figure 4 shows a socio-technical integration model of AI in architectural practice.

To assess AI’s implementation in the architectural firm, a systematic, dimensional approach is advantageous, as the literature identifies a variety of AI tools, each of which will be applied in different ways at different times [1,2,3,9]. In assessing AI applications in architecture, descriptive reviews indicate that assessments should differentiate between general and generative AI, as each is associated with distinct deliverables and verification processes [4,12,13,20]. Studies on generative AI applications, such as ChatGPT-like systems, and their potential for the Architecture, Engineering, and Construction (AEC) industries, show that these tools will influence organisations’ choices, training requirements, and accountability routines, regardless of whether the tools actually exist [3,11,16,29].

The AIMI framework evaluates three groups of dimensions, summarised in Table 1. Strategic and organisational dimensions assess leadership commitment, role clarity, budget allocation, and the treatment of AI as an ongoing capability rather than as a temporary addition [1,2,3,9,11,15,17,36]. Operational dimensions examine the breadth and frequency of AI use across design stages—concept generation, visualisation, modelling support, documentation, and coordination—and the consistency with which tools are embedded in production workflows [1,2,3,9,15,18,23,24]. Capability dimensions assess whether a firm possesses the skills, training routines, verification habits, and governance arrangements needed for reliable and accountable AI use [1,2,3,9,10,11,16,17,27,29,37,38]. Figure 5 presents a three-dimensional AIMI framework.

This study, therefore, compares local firms with international and internationally relevant benchmark firms using shared benchmark dimensions rather than direct equivalence of raw responses. Dimensions are defined as the adoption stage, range of tools, workflow integration, training, governance, quality control, strategic value, and future readiness. The dimensions are connected from the literature review, through the methods in the Section 3, to the benchmarks in the practical and case-study sections. Governance relates to responsibility for the final product/output, the review process, “handoff” rules, and stronger checks across all places where AI interacts with and/or impacts performance, compliance, and/or client-facing materials [7,8,12,17,40].

Table 1. Literature-based dimensions used to assess AI integration maturity in architectural firms.

Dimension	Purpose	Typical Evidence in a Firm	Main Literature Link
Adoption	Whether AI is used at all and at what status	Regular use, occasional use, discontinued use, non-use	[1,4,11,41]
Usage intensity	How often are AI tools used in real work	Daily, weekly, monthly, and rare use	[9,10,24]
Tool diversity	Range of AI tool families in active use	Text, image, rendering, BIM, parametric, custom tools	[3,12,13]
Workflow breadth	Number of task domains where AI appears	Design, visualisation, documentation, analysis	[2,15,23]
Staff involvement	Distribution of AI use across the office	One user, a small group, several role groups, and most staff	[9,11,17]
Capability	Training and learning arrangements	Self-learning, peer sharing, workshops, structured training	[11,17,37,41]
Project application	Share of real projects in which AI is used	Limited project use, routine project use, broad portfolio use	[10,28,40]
Commitment and governance	Rules, checking, spending, custom workflows, and strategic support	Policies, review sign-off, budget, internal tools, and records	[29,42,43]

Table 1. Literature-based dimensions used to assess AI integration maturity in architectural firms.

Dimension	Purpose	Typical Evidence in a Firm	Main Literature Link
Adoption	Whether AI is used at all and at what status	Regular use, occasional use, discontinued use, non-use	[1,4,11,41]
Usage intensity	How often are AI tools used in real work	Daily, weekly, monthly, and rare use	[9,10,24]
Tool diversity	Range of AI tool families in active use	Text, image, rendering, BIM, parametric, custom tools	[3,12,13]
Workflow breadth	Number of task domains where AI appears	Design, visualisation, documentation, analysis	[2,15,23]
Staff involvement	Distribution of AI use across the office	One user, a small group, several role groups, and most staff	[9,11,17]
Capability	Training and learning arrangements	Self-learning, peer sharing, workshops, structured training	[11,17,37,41]
Project application	Share of real projects in which AI is used	Limited project use, routine project use, broad portfolio use	[10,28,40]
Commitment and governance	Rules, checking, spending, custom workflows, and strategic support	Policies, review sign-off, budget, internal tools, and records	[29,42,43]

In summary, the three broad dimensions of the organisations being Strategic (organisational and strategic capabilities), Operational (daily operational activities) and Capability (ability to execute) also exhibit an inherent tension within the literature that is relevant to the Kurdistan Region context: the gap between tool access/availability and workflow discipline. Various authors demonstrate that AI tools are becoming available more rapidly than the organisation’s readiness to use those tools responsibly [1,3,11]. This creates what we would call a maturity lag, where the speed of technology adoption exceeds the speed of its governance.

There are two differing points of view regarding the recommended maturity measurement for AI: either that it should be measured mainly by output metrics (i.e., time saved or number of design options) or by process metrics (i.e., review routine, role clarity, and accountability). The output-oriented measurement of AI [10,27] places strong emphasis on productivity improvements and tends to yield an optimistic score for AI adoption readiness. The process-oriented measurement of AI [17,29,40] would argue that, if no stable governance structures are in place, the productivity improvements generated by AI tools are vulnerable to collapse and may be considerably misleading.

This paper takes the process-oriented position and places interpretive emphasis on Governance, Training, and Project-level Embedding in the AIMI structure; therefore, an organisation may use a significant number of AI tools for visualisation; however, if the organisation lacks formal review routines and lacks formal training, then that organisation would score a lower AIMI rating than what their output volume would suggest. The third tension in the literature concerns the relationship between scale and firm size; most of the most advanced-maturity organisations in the literature are large multinational corporations with dedicated teams of computational designers (Grimshaw, Foster + Partners, ZHA), creating a structural advantage for larger firms that smaller firms cannot easily replicate. The BIM maturity research conducted in Iraq made the same point about the difficulty of implementing digital standards for smaller firms lacking a formal QMS [7,8]. This paper directly addresses the scale concern made in the literature by including a variety of benchmark firm sizes and by designing the practical recommendations in Section 4.4 for micro- and small-firm practices, rather than just for larger firms with ample resources. This emphasis is conceptual rather than a hidden unequal numerical weighting: apart from C1, which is a threshold adoption item, the operational AIMI components are kept on the same 0–5 scale to avoid unsupported empirical priority assumptions at this stage of model development.

In a recent Saudi Arabian cross-sectional study by Alymani et al. [41], which provides the closest regional analogy to this study, 113 individuals from the architectural education (academic) and AEC industries were surveyed about their use of AI and machine learning (ML) technologies. The authors found high levels of general familiarity with these technologies among respondents but inconsistent usage rates. The authors found that the largest firms had the highest usage rates (approximately 61%). In comparison, smaller firms had usage rates under 25%, and firms with 41 to 100 employees reported no use of these technologies. The authors identified insufficient access to AI and ML tools, a lack of faculty and staff knowledge or experience, and resistance to changing established workflows as key barriers to the adoption of technological change in Saudi Arabia, concluding that the primary factor dictating whether an organisation will adopt these technologies is its capabilities rather than the level of knowledge, understanding, or motivation regarding these technologies. The authors’ findings corroborate the broader pattern in the Gulf while offering a useful reference point for the Kurdistan study results presented here.

2.4. Benchmarking and Transferability

Benchmarking in architecture and design provides an organised method for evaluating how firms structure work, allocate resources, and manage quality as new technologies emerge [9,10,17,27]. In AI-driven practice, benchmarking centres on functionality rather than marketing claims—identifying where tools enter the design process, how outputs are reviewed, and whether use is embedded in structured methodology or remains informal [9,10,12,13,17,18,20]. Consistent benchmarking categories—concept development, representation, modelling assistance, textual support, downstream applications, and tool-specific risks—enable comparable coding of evidence across firms and contexts [1,2,3,12,13,29].

The best practices I have seen always include process discipline, role clarity, a platform for human review, and quality assurance—though the actual tools used differ from reference to reference [17,31,34]. A significant portion of the literature on ChatGPT-like tools and generative AI also addresses verifying proper use of these tools and encouraging tool-specific quality assurance, since the checks for image, text, and analytic outputs differ [12,16,29]. BIM research on Iraq has a final example set of standard templates and similar conventions, as do AEC review cycles, whereby continuous training and documentation of the workflow process are central to sustained product use [3,7,8,11].

Transferability refers to the ability to adapt methods or governance arrangements instead of merely copying global tool stacks. Firms in the Kurdistan Region operate with different resources, approval lines, client expectations, and project priority rankings; therefore, benchmark lessons should be adapted to local conditions [5,6,7,8]. The most transferable actions include structured training, prompt and review schedules, clearly defined responsibilities, and the careful transition from early-stage image- and text-based support through to deeper BIM, parametric, and performance linking with existing standards [1,2,15,18,30]. A key point of contention in the benchmarking literature is the unresolved debate over exactly what benchmark firms represent. Many authors view leading global firms as aspirational models whose best practices should be followed exactly [9,10]. Others claim that the best practices of high-performing firms are based on available resources within their organisations, the conditions in the marketplace, and the expectations of their clients and cannot simply be copied into different environments [5,11,30].

This paper subscribes to the second position. Global benchmark firms are not viewed as goals to replicate; rather, they serve as references that provide insight into what organised, well-governed AI integration with supporting capabilities looks like in practice. The value of the comparison is not in the direct replication by Kurdistan-based firms of a specific global firm, but rather in demonstrating that project penetration, training, and governance are the three areas that represent the largest degree of disparity between firms and, as such, any firm-level investment in those areas will likely yield the largest increment in relative maturity. The distinction between aspiration and transferability is a key building block for the recommendations in the discussion.

2.5. Literature-Based Analytical Dimensions

According to the literature, AI maturity should be viewed through a multi-dimensional lens, not just as an indicator of adoption. The dimensions are summarised in Table 1 along with their connection to the evidence base.

The AIMI framework follows the same conceptual logic as established capability maturity models (CMMI; Succar’s BIM Maturity Index [44]; Lichtenthaler’s five-level AI-maturity model [45])—particularly the Capability Maturity Model Integration (CMMI)—which organises organisational maturity across five progressive levels from unstructured practice to optimised continuous improvement, an approach applied in adjacent AEC research, including BIM maturity assessment in the regional context [7]. The five AIMI bands—Non-adopter, Exploratory, Occasional, Integrated, and Advanced strategic—map directly onto this progression, adapted to the specific dimensions of AI integration in architectural practice rather than software development process quality. The contribution of the AIMI is therefore not a new conceptual structure but a domain-specific operationalisation of the maturity model philosophy for an architectural technology context that existing frameworks do not address. The AIMI five-band structure aligns with and provides a finer level of detail than three-level models of regional architectural artificial intelligence maturity that are emerging in the literature, which separate these into experimental, operational, and strategic adoption phases [41].

2.6. AI Failures, Risks, and Limitations in Architectural Practice

The literature consistently identifies a range of failure modes and occupational risks associated with the use of AI in architectural practice. Output hallucination—where generative AI produces plausible but technically incorrect or spatially impossible results—is the most frequently cited failure mode, particularly for text-based and image-based outputs used in early design stages [1,3]. Copyright and intellectual property risks arise when AI systems trained on unlicensed material generate outputs that may infringe on existing design works, raising unresolved legal questions about authorship and liability [16,29]. Accuracy and verification failures occur when AI-assisted documentation, specification, or analysis outputs are incorporated into project deliverables without adequate professional review, creating risks to building performance, regulatory compliance, and client trust [22,30]. Deskilling risks have been noted in design education research, where overreliance on generative tools may reduce practitioners’ capacity for independent spatial reasoning and critical design judgement [3,22]. Finally, data security and confidentiality risks are particularly acute when practitioners use cloud-based AI platforms for work involving sensitive client information, proprietary design data, or commercially confidential project details [16,29]. These risks reinforce the process-oriented position of this study: mature AI integration requires not only tool adoption but also governance structures, review routines, and professional accountability to systematically mitigate these failure modes.

3. Methodology

3.1. Research Design

The study used a mixed-method descriptive comparative design. This method was suitable because the research needed to describe the current state of AI integration in a regional professional context and then compare it with international benchmark practices. For the local strand, a structured questionnaire was used to generate respondent-level quantitative data (see Figure 6). For the international component of the study, structured, asynchronous, text-based interviews were employed to generate qualitative, firm-level benchmark data, as shown in Figure 7.

A shared maturity framework connects the two datasets, but they are not treated as statistically identical. The local questionnaire provides survey-derived scores, frequencies, percentages, correlations, and maturity distributions, while the international interviews provide qualitative benchmark scores coded from structured textual evidence. The same AIMI components were applied to both datasets to enable comparison; however, the scoring sources differed. Local firm scores were calculated directly from structured questionnaire responses, whereas international benchmark scores were derived through qualitative coding of interview texts. Therefore, the international scores are interpreted as benchmark coding scores rather than statistically equivalent survey measurements, and they are not used for inferential testing. This design uses two evidence streams because the local strand required respondent-level survey data, whereas the international strand sought firm-level expert benchmark accounts; comparability is therefore maintained at the level of AIMI components, not at the level of identical raw instruments. Figure 8 illustrates a mixed-methods comparative research design for this study.

3.2. Population, Sampling, and Participants

The unit of analysis was the architectural firm. For the local strand, the target population included architectural firms operating in the Kurdistan Region of Iraq. The pilot expert review process helped minimise ambiguity, improve question sequencing, and enhance the overall reliability and validity of the survey questionnaire. Then, the questionnaire was distributed through relevant professional channels, contacts, and online networks. The Google Forms export contained 100 responses received within the declared local data-collection period ending 15 April 2026. The local survey sample is concentrated in the region’s main architectural markets. Erbil accounts for 44% of responses, Sulaymaniyah for 39%, Duhok for 10%, and the remaining 7% are distributed across Soran, Halabja, Zakho, Rania, Koya, and other locations. The respondent profile supports firm-level interpretation because 55% of respondents were principals, partners or owners, while 23% were senior architects. The firm-size profile also reflects the structure of the regional market: 33% micro-practices, 41% small practices, 19% medium practices and 7% large practices. Data collection ran from 1 March to 15 April 2026, a period of approximately six and a half weeks; because the duration was below three months, no separate data-collection pace chart was added. The purposive online distribution may over-represent firms and respondents willing to discuss AI, so the sample is interpreted as a strong descriptive regional profile rather than a census.

The international benchmark strand consisted of structured qualitative benchmark profiles of ten selected firms with visible AI engagement or international relevance. The benchmark firms were purposively selected for their international reputation, leading professional practices, and recognised contributions to architectural innovation. The selection process considered factors such as professional recognition, project quality, industry influence, and relevance to the research objectives. This ensured that the qualitative interviews reflected advanced practices and expert insights relevant to the study. The benchmark firms were coded as International Firm A, Firm B, Firm C, and so on. These cases were selected for comparative learning rather than statistical representativeness, and their coded values are treated as qualitative benchmark evidence rather than survey-derived measurements. The final benchmark cases were selected through five explicit criteria: visible engagement with AI, computational design or digital innovation; availability of structured firm-level evidence; variation in location, size and specialisation; geographic diversity across North America, Europe, Asia, Australia and the Middle East; and relevance to transferable lessons for architectural practice. They are used as advanced reference cases, not as matched controls for small local firms, so the comparison is intended to locate maturity gaps and transferable practices rather than to imply that every local firm should replicate a leading global firm. Figure 9 shows the geographical distribution of respondent firms and their size by the number of employees.

Table 2. Profiles of the Kurdistan Region samples for key descriptive variables.

Variable	Category	n	%
City	Erbil	44	44.0
	Sulaymaniyah	39	39.0
	Duhok	10	10.0
	Soran	2	2.0
	Halabja	1	1.0
	Other	1	1.0
	Zakho	1	1.0
	Rania	1	1.0
	Koya	1	1.0
Role	Principal/Partner	55	55.0
	Senior Architect	23	23.0
	Junior Architect	10	10.0
	Visualisation specialist	6	6.0
	Project Manager	3	3.0
	Administrative	2	2.0
	Other	1	1.0
Experience	>15 yrs	28	28.0
	6–10 yrs	25	25.0
	11–15 yrs	24	24.0
	3–5 yrs	14	14.0
	<3 yrs	9	9.0
Firm size	Small (6–15)	41	41.0
	Micro (1–5)	33	33.0
	Medium (16–30)	19	19.0
	Large (31+)	7	7.0
Years in practice	5–10 yrs	41	41.0
	<5 yrs	36	36.0
	11–20 yrs	19	19.0
	>20 yrs	4	4.0
Annual project volume	6–10 projects	28	28.0
	Prefer not to answer	23	23.0
	More than 20 projects	18	18.0
	11–20 projects	17	17.0
	1–5 projects	14	14.0
Project budget	Prefer not to answer	42	42.0
	$100,000–$500,000	21	21.0
	Under $100,000	17	17.0
	$500,000–$1 million	13	13.0
	Over $5 million	4	4.0
	$1 million–$5 million	2	2.0
	Missing	1	1.0
AI use status	Occasional user	62	62.0
	Regular user	20	20.0
	Non-adopter	14	14.0
	Discontinued	4	4.0

outlines the profiles of the Kurdistan Region samples for key descriptive variables and

Table 3. Overview of international benchmark firms.

Firm	Location	Size (Staff)	Years in Business	Primary Specialisation	AI Use Since	AI Capability History (in Detail)	AIMI (Band)
A	London, UK	Large (~1700+)	59 (est. 1967)	Architecture, engineering, master planning	~2010s (early adopter)	Long-standing in-house applied-R&D group developing internal machine-learning tools for design optimisation, environmental-performance evaluation and generative form-finding; dedicated computational-design team and structured staff training	37 (Advanced strategic)
B	London, UK	Large (~600)	46 (est. 1980)	Engineering-led architecture and computational design	Recent (long BIM/parametric history; AI tools added recently)	BIM/parametric foundation; AI-assisted tools embedded in design and delivery for environmental-performance analysis, structural optimisation and visualisation	37 (Advanced strategic)
C	Copenhagen, Denmark	Large (~700)	21 (est. 2005)	Architecture, urban design and design technology	2010s (early/consistent adopter)	Data-driven, iterative computational design; dedicated in-house research/innovation unit working on generative AI, parametric design and environmental simulation; AI used in massing, environmental analysis, façade optimisation, visualisation and client presentation	35 (Advanced strategic)
D	London, UK	Large (~500)	47 (est. 1979)	Architecture and computational design	2010s (long computational history)	Dedicated computation-and-design research group developing machine-learning tools for form-finding, structural optimisation and material testing, with active ML research	35 (Advanced strategic)
E	London, UK	Small	3 (est. 2023)	Computational and generative design	2023 (AI-native from inception)	AI-native studio applying AI from concept generation through detailed design and visualisation; operates a dedicated research lab building proprietary in-house AI tools	33 (Advanced strategic)
F	Dubai, UAE	Large (~5000–6000)	62 (est. 1964)	Architecture, engineering and planning	Recent (BIM/parametric base; moving toward AI)	Very large multidisciplinary consultancy; BIM, parametric design and performance simulation embedded; progressing to AI-assisted structural analysis, environmental simulation and visualisation in project delivery	31 (Advanced strategic)
G	Dubai and London	Small (~9)	3 (est. 2023)	Experimental design and material research	2023 (AI/XR-native from founding)	Small AI/XR-native studio at the intersection of architecture, research and art; AI for material-behaviour study, generative design and spatial exploration; highly experimental and creatively driven	28 (Integrated)
H	Budapest, Hungary	Small-Medium	~20 (est. ~2006)	Research-led computational design	Recent (research-led experimentation)	Research-led practice bridging academia and professional design; AI-assisted tools across design, fabrication and visualisation within a culture of experimentation; occasional, research-driven use	21 (Occasional)
I	Australia	Small-Medium	15 (est. 2011)	Contemporary design and visualisation	Recent (exploratory; expanding)	Uses AI for idea generation, layout planning and client presentation; progressively expanding AI use and building staff capability	15 (Exploratory)
J	United States	Mid-sized	106 (est. 1920)	Healthcare, education and workplace design	Recent (early/exploratory stage)	Early BIM/digital adopter beginning to explore AI for space planning, visualisation and documentation; AI not yet embedded or supported by formal training or governance	14 (Exploratory)

presents overview information of international benchmark firms.

3.3. Instruments and Measurements

Two instruments were employed to conduct this study. The Kurdistan Region’s Survey Questionnaire contained eight main parts: Firm Profile, AI Awareness, Adoption Status, Tool Type Utilisation, Application of Workflow Practice, Infrastructure, Attitudes & Future Projections, and one optional example based on practices. While the survey was available in multiple languages, the analysis used the Index of each question and response category for each measure’s respective English definitions/meanings, along with corresponding Kurdish and Arabic textual representations as Supplementary Translations.

The interview guide for international participants consisted of 12 parts, covering numerous aspects, including firm context, adoption pathway, strategic positioning, tool ecosystem, workflow integration, project applications, capabilities and training, governance and ethics, partnerships, best practices, and future outlook. The written format allowed participants from outside Australia to reply to the guide at their own pace and produced written evidence that could be coded systematically.

The core measurement tool is the AI Integration Maturity Index. For adoption, a score ranging from 0 to 2 was assigned. For intensity of use, diversity of tools, breadth of workflows, depth of application in projects, staff involvement, level of training, and governance, scores ranged from 0 to 5. Scores ranged from 0 to 37 and were grouped into 5 maturity bands—Non-adopter, Exploratory, Occasional, Integrated, and Advanced strategic. For local firms, scores were calculated from the raw survey responses using a deterministic rubric programmed into the analytical script, and the process could be repeated exactly from the survey data file. For international benchmark firms, scores were derived from qualitative coding of interview-based text evidence, conducted by researchers using the AIMI component coding rubric, and every component of the coding process could be substantiated in the transcripts through an automated keyword audit.

It is important to note that the scores for international firms are not survey statistics; they are coded maturity estimates based on documentary evidence, whereas the scores for local firms were generated algorithmically from structured questionnaire data. There are three design decisions in the rubric to be disclosed here for transparency. The first design feature is that the mapping for C2 usage intensity is designed to be non-monotonic (Do Not Use = 0, Rarely = 1, Monthly = 2, Weekly = 4, and Daily = 5) to encode the qualitative threshold between occasional and routine use, which researchers have indicated exists between monthly and weekly use frequencies. Second, where adopters indicated “Not sure,” “Don’t know,” or “Prefer not to answer” on the C5, C7, or C8 driving items, the response was scored at the midpoint of the relevant 0–5 sub-scale (value 2) under the primary scoring rule, with a sensitivity analysis under alternative imputation rules reported alongside the primary results. Third, the “pro-AI attitude” composite used in the correlational analysis is constructed from the first four items of the ten-item attitudes block: enthusiasm about AI in architecture, agreement that AI improves workflow efficiency when used appropriately, agreement that AI improves design quality when used appropriately, and agreement that AI should be taught in architecture schools and universities. The remaining six attitude items mix neutral and sceptical phrasings and are excluded from the pro-AI composite. Figure 10 presents the broader conceptual framework; the operational scoring index used in this study is the eight-component AIMI with a total score range of 0–37.

A fourth point concerns weighting. The eight components are combined with equal (unit) weight as a simple summed score, following the convention of established capability-maturity instruments (the Capability Maturity Model Integration, Succar’s BIM Maturity Index [44], and Lichtenthaler’s AI maturity model [45]), in which dimensions are weighted equally because there is no a priori empirical basis for privileging one dimension over another in a first operationalisation, and equal weighting keeps the index transparent and reproducible. Adoption (C1) is deliberately bounded at 0–2 rather than 0–5 because it functions as a coarse entry threshold (no use, occasional use, and regular use) rather than a graded depth dimension; placing it on the same 0–5 range as the substantive dimensions would overweight the mere fact of use relative to the depth of integration. The five maturity bands partition the 0–37 range to mirror the five-level progression of the capability-maturity tradition, with boundaries placed at substantively meaningful transitions: 0–10 (non-adoption or ad hoc trial), 11–16 (informal exploration), 17–22 (use that is established but not embedded), 23–28 (integration embedded across several workflows and projects), and 29–37 (embedded, trained, governed and strategic use). To verify that these design choices do not drive the findings, a weighting-sensitivity analysis and a band-cut-point robustness analysis are reported in the Section 4.

For transparency and exact reproducibility, the full local scoring rubric is as follows. C1 Adoption (0–2): not currently using AI = 0, occasional use = 1, and regular use = 2. C2 Usage intensity (0–5), taken from the most frequently used tool: do not use = 0, rarely = 1, monthly = 2, weekly = 4, and daily = 5 (the mapping is intentionally non-monotonic, as noted above). C3 Tool diversity (0–5): the count of distinct AI tools in active use, drawn from the nine tools enumerated in the survey, capped at 5. C4 Workflow breadth (0–5): the number of the four task domains (design, visualisation, documentation, and analysis) in which AI is used, mapped as 0→0, 1→2, 2→3, 3→5, and 4→5. C5 Staff involvement (0–5): based on the reported number or share of staff actively using AI, with non-disclosing adopters scored at the midpoint (2). C6 Capability and training (0–5): based on the presence and formality of AI training. C7 Project penetration (0–5): based on the reported percentage of projects involving AI, with non-disclosing adopters scored at the midpoint (2). C8 Commitment and governance (0–5): based on monthly AI spending and the presence of in-house or custom development, with non-disclosing adopters scored at the midpoint (2). Non-adopting firms (C1 = 0) score 0 on C2–C8 by construction.

The 37-point AIMI score was classified into five maturity levels as shown in

Table 4. Distribution of AIMI maturity bands.

AIMI Score Range	Maturity Band	Interpretation
0–10	Non-adopter	No current AI use or only very weak/very early use
11–16	Exploratory	Basic experimentation, informal use, limited structure
17–22	Occasional	AI is used, but not deeply embedded
23–28	Integrated	AI is used across several workflows and projects
29–37	Advanced strategic	AI is embedded, trained, supported, governed, and strategically used

. These levels allow the study to distinguish between simple adoption and mature integration. The categories also support comparison between local firms and international benchmark firms.

Table 5. AIMI component coding rubric for international benchmark firms.

AIMI Component	What Is Coded	Keyword/Evidence Examples	Score Logic
C1 AI adoption	Whether the firm currently uses AI. Score: 0 = no use, 1 = limited/testing, 2 = active/embedded.	routine use; active use; early testing; not officially adopted; embedded	0 = no use; 1 = limited/testing; 2 = active/routine/embedded
C2 Usage intensity	How frequently is AI used? Score: 0 = none, 1 = rare, 2 = low, 3 = moderate, 4 = regular, 5 = routine/embedded.	daily; routinely; across studios; all current projects; selective; limited	0 = no evidence; 1 = very weak; 2 = limited; 3 = moderate; 4 = strong; 5 = advanced/embedded
C3 Tool diversity	Breadth of AI tool categories: LLMs, image generation, rendering, BIM/parametric, analysis, in-house/custom systems.	ChatGPT; Gemini; Midjourney; Stable Diffusion; D5; Rhino/Revit; in-house tools	0 = no evidence; 1 = very weak; 2 = limited; 3 = moderate; 4 = strong; 5 = advanced/embedded
C4 Workflow breadth	Number of workflow stages using AI: concept, visualisation, presentation, documentation, analysis, BIM, business development, etc.	concept; visualisation; presentation; documentation; analysis; BIM; business development	0 = no evidence; 1 = very weak; 2 = limited; 3 = moderate; 4 = strong; 5 = advanced/embedded
C5 Staff involvement	How widely AI use is distributed among staff, specialists, teams, and studios.	everyone; design technology team; specialists; across studios; individual designers	0 = no evidence; 1 = very weak; 2 = limited; 3 = moderate; 4 = strong; 5 = advanced/embedded
C6 Capability/training	Training, workshops, internal guides, peer learning, prompt libraries, or formal capability building.	workshops; internal school-style training; lunch-and-learn; prompt libraries; self-learning; guides	0 = no evidence; 1 = very weak; 2 = limited; 3 = moderate; 4 = strong; 5 = advanced/embedded
C7 Project application	Whether AI is used in real projects, live projects, competitions, client work, or only experiments.	live projects; real projects; competition entry; client work; not for client work yet	0 = no evidence; 1 = very weak; 2 = limited; 3 = moderate; 4 = strong; 5 = advanced/embedded
C8 Commitment/governance	Investment, in-house tools, senior review, approved tools, data security, IP/confidentiality policy, and strategic support.	approved tools; data security; IP; senior review; internal servers; custom platform; R&D	0 = no evidence; 1 = very weak; 2 = limited; 3 = moderate; 4 = strong; 5 = advanced/embedded

documents how the international benchmark scores were derived from the text-based interview forms. Each score is justified through matching keywords, evidence phrases, and interpretation against the AIMI rubric.

Because a structured questionnaire can only be administered to firms that participate in the survey, the international benchmark firms were assessed using documentary and interview-based evidence rather than the survey instrument. The shared AIMI rubric is therefore used as a common interpretive lens, and the two sets of scores are not pooled statistically. Table 5 outlines the AIMI component coding rubric for international benchmark firms.

3.4. Data Analysis and Ethics

Survey data were analysed through descriptive statistics, cross-tabulations, scale summaries and maturity scoring. Categorical variables were reported through frequencies and percentages. Ordinal rating scales were summarised through means, medians, standard deviations and top-box percentages. Spearman correlations were used for ordinal associations, and the Kruskal–Wallis and Mann–Whitney tests were used when group comparisons were appropriate. Because the study follows a descriptive-comparative design without directional hypotheses, all inferential tests are treated as exploratory, supplementary checks rather than confirmatory hypothesis tests. Chi-square tests were interpreted as exploratory because some expected cell counts were small, especially in adoption-status categories and smaller city or firm-size groups. The international interviews were analysed through within-case profiling and cross-case framework analysis. Each transcript was read as a firm-level case and coded across the same eight AIMI components. The coded scores are qualitative benchmark coding scores, not statistical survey scores. Their purpose is to provide a maturity reference point and to support the identification of transferable practices, not to support inferential comparison with the local sample. Ethical considerations included voluntary participation, informed consent, confidentiality and careful use of firm-level evidence. Local survey results are reported in aggregate. International firm names are used as benchmark labels where appropriate, but the interpretation focuses on organisational patterns rather than personal attribution. For the qualitative benchmark coding of international firm interviews, inter-coder reliability was assessed by having a second independent coder apply the same AIMI rubric (Table 5) to a randomly selected subset of three interview transcripts (30% of the international sample). Agreement was assessed using Cohen’s kappa, yielding kappa = 0.96, which indicates near-perfect agreement [46]. All coding disagreements were resolved through discussion and reference to the anchored keyword descriptors in Table 5. The automated keyword audit script (available as Supplementary Material) provides a transparent, reproducible record of the evidence phrases matched to each component score for every international firm.

4. Results and Discussion

4.1. Local Survey Data Analysis

4.1.1. Local Sample Profile

In the Kurdistan Region, 100 valid responses were collected, exceeding the minimum sample size of 78 firms calculated for an estimated population of approximately 400 firms at a 95% confidence level and a 5% margin of error. The estimated population of about 400 firms comes from a combination of the total number of registered engineering and architecture firms in the Erbil, Sulaymaniyah, Halabja, and Duhok governorates, matching engineering practice records with those recorded by members of the Kurdistan Engineers Syndicate, as well as other sources such as social networks and academic contacts from the time when the data was collected. The required sample size was determined using Cochran’s sample size formula with finite population correction, which is commonly used in survey-based research to estimate an appropriate sample size from a known population. The sample is geographically concentrated in the region’s two largest urban centres. Erbil accounts for 44% of firms, Sulaymaniyah 39%, Duhok 10%, and the remaining 7% come from Halabja, Soran, Zakho, Koya, and Rania. The dominance of Erbil and Sulaymaniyah is consistent with the sample’s concentration in the region’s two largest architectural markets. At the same time, the smaller-city responses prevent the analysis from being reduced to a single-city profile.

Respondents are predominantly senior decision-makers. Principals, partners, or owners of the firm represent 55% of respondents; senior architects represent 23%; junior architects represent 10%; visualisation specialists represent 6%; project managers represent 3%; and administrative or management staff represent 2%. The high seniority of respondents supports the validity of organisational claims because most respondents hold positions from which they can describe firm-level workflows, technology choices, and resource allocation, rather than only their own personal use. In terms of length of professional experience, 28% of respondents have more than 15 years, 24% have 11 to 15 years, 25% have 6 to 10 years, 14% have 3 to 5 years, and only 9% have less than 3 years. The experience profile thus skews towards mature practice rather than the early-career segment that is sometimes over-represented in technology-adoption surveys.

A total of 33% of the sample are micro-firms and 41% are small firms. The distribution of these firms appears to align with the overall regional market characteristics, and they account for 74% of all firms in the sample (33% micro- and 41% small), which helps explain why many firms lack sufficient training budgets or formalised capability frameworks. The average age of the firms also reflects this trend: 36% have been operating for less than five years, and 41% for five to ten years.

Annual project volume varies widely across the samples: 28% of firms complete six to ten projects per year, 17% complete eleven to twenty projects, 18% complete more than twenty projects, 14% complete one to five projects, and 23% of respondents preferred not to disclose this figure. 42% of respondents declined to disclose their typical project budget range. In comparison, 17% indicated that the typical project is below USD 100,000, 34% indicated a typical range between USD 100,000 and USD 1 million, and 6% indicated values above USD 1 million. The high non-disclosure rate for commercially sensitive items is consistent with the cautious response posture documented in the international corpus, in which firms repeatedly described internal financials as confidential. The disclosed responses suggest a range of commercial and institutional project types, although the high non-disclosure rate limits the strength of claims about budget representativeness.

4.1.2. Local AI Awareness and Adoption

The level of self-reported knowledge of AI tools among the Kurdistan Region’s survey group is very high. Of the respondents, 36% reported having previous knowledge of AI tools related to the field of architectural design; 56% indicated being moderately knowledgeable; 8% claimed to have heard of such tools, yet reported no knowledge of using them. None of the survey respondents indicated being completely unaware of how to use these types of tools. Thus, the total combined percentage of those familiar and those with moderate familiarity is 92%. This value provides a significant cognitive foundation for subsequent questions about the use of AI-based design technology. The major means by which respondents were first introduced to using AI-based design technologies are informal/unsupervised learning methods associated with regional skill development. Specifically, 73% of respondents were first introduced to AI tools through social media sites (Instagram, LinkedIn); 41% via architectural websites and blogs; 39% through YouTube video tutorials; 38% through colleagues and professional associates; 12% through professional conferences or workshops; and only 6% through formal academic or university courses. Consequently, the educational pathway to using AI design technology is considerably weaker (compared to the social/peer pathways), as noted in most sections of the ecosystem support discussion.

Awareness of specific tools is uneven. Midjourney is the best-known image-generation tool at 63%, while ChatGPT dominates text-based awareness at 95%, and Gemini follows at 84%. Architecture-specific tools are less familiar: Revit AI plugins are recognised by 32%, Rhino plugins by 28%, Autodesk Spacemaker by 26%, Finch 3D by 9%, and TestFit by 2%. This profile of tool awareness aligns with the global trend identified by Corticos et al. [25], which highlights that consumer tools are the predominant types utilised by practitioners, irrespective of their geographical location. However, in developing regions, where economic constraints remain significant, architecture-specific platforms that offer deeper workflow integration account for only a small percentage of the tools used by practitioners. AI-enhanced rendering tools occupy a middle position, with high recognition of Gemini image-generation tools and other emerging AI-enhanced visualisation tools, but lower recognition of specialist options such as Veras and LookX AI.

Three patterns emerge from the awareness assessment. First, a small number of consumer-facing tools dominate awareness in each category and structure the working horizon of what local respondents understand AI to mean in their daily work. Second, awareness of architecture-specific and parametric AI plugins is materially weaker than awareness of generic generative platforms, which previews the workflow-integration finding that AI sits closer to communication and visualisation tasks than to BIM and analytical tasks in local practice. Third, the rapid emergence of new image and rendering platforms in the open responses, including Gemini image-generation tools, Magnific, mnml.ai, and Krea, demonstrates that the awareness landscape is outpacing the formal teaching curriculum and that informal channels are filling that gap.

AI adoption is widespread in the Kurdistan Region’s sample. Twenty firms reported regular use of AI tools, sixty-two firms reported occasional use, four firms tried AI tools and stopped, and fourteen firms have never used them. The combined adoption rate of regular and occasional users is therefore 82%, although the dominant pattern of use is occasional rather than regular. The discontinued category is small, and the open responses of the four firms in this group cite unsatisfactory outputs and a long learning curve as the main reasons for stopping, rather than cost or infrastructure. Among non-adopters, the leading reasons for non-adoption are a lack of awareness about available tools (7 mentions), a preference for traditional methods (6 mentions), a lack of training (5 mentions), concerns about output quality (5 mentions), and a wait-and-see attitude (5 mentions), with cost and ethical concerns playing a smaller role. The non-adoption profile is therefore dominated by capability and confidence considerations rather than financial or technological barriers, a pattern that has direct consequences for ecosystem-level interventions.

Cross-tabulation of adoption status against firm size yields a chi-square statistic of 14.79 with 9 degrees of freedom and an associated probability of 0.097, which falls just short of conventional statistical significance. The descriptive pattern, however, is consistent with the international literature on size-related digital adoption. Among large firms with 31 or more staff, 4 out of 7 report regular use of AI tools, while among micro-firms with 1 to 5 staff, only 6 out of 33 are regular users. Cross-tabulation of adoption status against respondent experience yields a chi-square statistic of 7.47 on 12 degrees of freedom, with an associated probability of 0.825, suggesting that adoption is not linearly tied to respondent experience and that early-career respondents are not the only, or even the dominant, drivers of adoption. The cross-tabulation by city is more imbalanced due to the sample’s geographical concentration, but it does not show a meaningful structural divergence between Erbil and Sulaymaniyah in adoption status. Figure 11 shows the distribution of AI adoption status among architectural firms in the Kurdistan region.

4.1.3. Tool Ecology and Workflow Integration

Among the 82 current adopters, ChatGPT is the most frequently used tool: 67% report at least monthly use, including 25 firms using it daily, 15 weekly, and 15 monthly. The “Other” free-text category—covering D5 Render AI features, Krea, Magnific, mnml.ai, Nano Banana/Imagen, and similar AI-enhanced visualisation tools—is the second-largest cluster at 42% or more monthly. Midjourney follows, with 24 firms (29.3% of adopters) at monthly or higher usage, while DALL-E, Stable Diffusion, Claude, Gemini, Rhino AI plugins, and Revit AI plugins all sit below that threshold (see Figure 12 below).

Three observations follow from the frequency matrix. First, the local AI ecosystem is dominated by general-purpose text generation, primarily for documentation and communication tasks, and by general-purpose image generation, primarily for early-stage visualisation. Second, architecture-specific AI plugins within Rhino and Revit are recognised by fewer than a third of respondents and used by even fewer, which mirrors international evidence that BIM-coupled AI is the slowest-developing integration channel, even in advanced firms. Third, the high standing of the “Other” category, particularly newer image-generation tools such as Krea, mnml.ai, Magnific and Nano Banana/Imagen, indicates that the local tool ecology is now diverging from the canonical Western consumer roster and is acquiring its own informal vernacular of tools.

AI is applied across four task domains. Within design-phase tasks, the strongest uses are initial concept generation (58.5% of adopters), facade design (56.1%), design alternatives generation (50.0%), form exploration and interior design concepts (each 37.8%), landscape design (35.4%) and spatial layout or floor planning (19.5%). Within visualisation tasks, renderings and photorealistic images lead the entire dataset at 76.8%, followed by client presentations (48.8%), mood boards and concept images (47.6%), marketing materials (31.7%) and style exploration (29.3%). Documentation and communication tasks are more widely adopted than a high-level reading would suggest: writing project descriptions (48.8%), drafting proposals, emails, and presentations (46.3%), drafting concept narratives (43.9%), generating reports (40.2%), and specification-related writing (26.8%). Analysis and decision-support tasks also account for a meaningful share of AI use, with option comparison and decision support at 41.5%, site analysis support at 40.2%, and environmental and performance analysis at 34.1%. Regulatory-interpretation tasks—building codes, zoning regulations, accessibility standards, environmental regulations, and fire safety codes—cluster together at 29.3% each and form the weakest sub-category, as expected, because output reliability is critical in compliance work.

The aggregate pattern across the four domains is that AI is most strongly entrenched in visualisation, second most strongly in early-stage design (concept generation, facade and alternatives), third most strongly in documentation tasks that involve narrative or descriptive writing, and somewhat less strongly—but still meaningfully—in analysis and decision-support tasks; regulation-related and code-interpretation tasks are the slowest area where output reliability is critical. Forty-seven adopters cover all four domains in their workflows, which corresponds to 57% of the adopter subsample. This breadth measure feeds directly into the AIMI workflow component reported in Section 4.1.5.

The question of approximately what share of projects involves AI tools operationalises the depth of integration. Twenty-three adopters report using AI on 11 to 25% of their projects; twenty-two report 26 to 50%; seventeen report 0 to 10%; nine report 51 to 75%; three report 76 to 100%; and eight indicate they cannot estimate. The modal band is therefore 11–50%, which is consistent with a workflow that uses AI selectively at the concept and visualisation stages on most projects, rather than as a continuous companion throughout the whole design timeline. The combined share of firms applying AI to more than half of their projects is 12 out of 82 adopters, which is 14.6% of adopters and 12% of the entire sample.

4.1.4. Capability, Investment, Benefits, Challenges, and Readiness

The capability profile is dominated by self-directed learning. Fifty adopters rely solely on self-learning, while 22 report some formal training and only 4 report comprehensive training. Subscription spending is also modest: most adopters spend USD 100 or less per month, and only about one fifth spend more. These figures show that AI capability in the local sample is present but still weakly institutionalised.

Of the 82 people (not including “no response”) who answered the question about whether their firm has developed an internal or custom artificial intelligence tool, plugin, or workflow, 8 said “yes”; 6 were unsure; and 68 replied “no”. All of these firms provided specific examples in optional open-ended follow-up comments. One firm developed an internal website for rendering visuals; another created a workflow for generating actual images with surrounding contextual detail; and a third has integrated AI assistance into both data collection and rendering quality. A fourth firm developed a road management system and an artificial-intelligence-enabled building energy-use optimisation tool. A fifth firm developed a visualisation tool that could recognise and use patterns created by the firm to produce visualisation outputs consistently. A sixth firm developed a recognition tool for their architectural features (using machine learning algorithms based on the firm’s own dataset) in or around the year 2020. The remaining two are described in less detail. The presence of eight bespoke or experimental development efforts in a sample of 100 firms places the local custom-development rate at 8%, which is meaningful in absolute terms even if it remains far below the dedicated computational design teams documented in the international corpus.

The benefits scale shows a clearly positive but moderate evaluation across the ten items. The strongest perceived benefit is time savings on repetitive tasks, with a mean of 3.79 out of 5, a standard deviation of 1.12, and 63.4% of respondents rating it 4 or 5. Better visualisation quality is the second-strongest benefit, with a mean of 3.49 and 50% rating it four or five. Improved client presentations score 3.35, with 48.8% rating it four or five. Enhanced creativity and idea exploration scores 3.29, more design alternatives in less time scores 3.28, increased client satisfaction scores 3.27, faster design process scores 3.24, cost savings and resource efficiency scores 3.07, better design communication scores 2.98, and competitive advantage scores 2.93. The Cronbach’s alpha for the ten-item benefits scale is 0.858, exceeding the 0.70 threshold and indicating internal consistency.

The challenges scale shows lower mean ratings overall but reveals a clear hierarchy. The leading perceived challenge is output quality inconsistency, with a mean of 3.57 and 53.7% rating it 4 or 5. Lack of control over the results scores 3.39, with 47.6% rating it 4 or 5. Cost of subscriptions scores 3.06; integration with existing software and workflow scores 2.95; client scepticism scores 2.62; internet speed and connectivity scores 2.62; copyright and intellectual property concerns score 2.56; ethical concerns score 2.41; hardware limitations score 2.38; and learning curve and complexity score 2.29. The internal consistency of the challenges scale is 0.785. Two observations are striking. First, the two highest-rated challenges are both about output reliability rather than about access or affordability, which suggests that the principal limit on deeper integration is professional confidence in AI outputs rather than the cost or availability of tools. Second, the lowest-rated challenge is the learning curve, suggesting that, in this sample, perceived complexity is less important than output reliability and control as barriers to deeper AI integration.

Table 6. Perceived benefits of AI use among adopter firms.

	Mean	Median	SD	% Rating ≥ 4
Time savings in repetitive tasks	3.79	4.0	1.12	63.4
Better visualisation quality	3.49	3.5	1.15	50.0
Improved client presentations	3.35	3.0	1.1	48.8
Enhanced creativity/idea exploration	3.29	3.0	1.06	40.2
More design alternatives in less time	3.28	3.0	1.3	41.5
Increased client satisfaction	3.27	3.0	1.12	39.0
Faster design process	3.24	3.0	1.26	45.1
Cost savings/resource efficiency	3.07	3.0	1.16	32.9
Better design communication	2.98	3.0	1.07	29.3
Competitive advantage	2.93	3.0	1.18	32.9

and

Table 7. Perceived challenges of AI use among adopter firms.

	Mean	Median	SD	% Rating ≥ 4
Output quality inconsistency	3.57	4.0	1.08	53.7
Lack of control over results	3.39	3.0	1.03	47.6
Cost of subscriptions	3.06	3.0	1.19	32.9
Integration with existing software/workflow	2.95	3.0	1.0	30.5
Client scepticism	2.62	2.0	1.15	28.0
Internet speed/connectivity	2.62	3.0	1.23	23.2
Copyright/IP concerns	2.56	2.0	1.34	24.4
Ethical concerns	2.41	2.0	1.26	19.5
Hardware limitations	2.38	2.0	1.22	18.3
Learning curve/complexity	2.29	2.0	0.99	13.4

outlines perceived benefits and challenges of AI use among adopter firms, respectively.

The attitude items indicate a favourable but constrained position. Respondents largely agree that architectural schools should teach artificial intelligence (AI) disciplines. In addition, when AI is utilised effectively, the result will be enhanced workflow efficiency and improved design quality. Although respondents showed little agreement that AI threatens architects’ jobs or that it is too complex for small firms, this trend suggests acceptance of AI and low concern about job displacement.

The overall attitude variable confirms this picture. Twenty-five respondents describe their attitude as very positive and excited about opportunities; forty-six describe it as positive, seeing benefits with some limitations; twenty-five take a neutral, wait-and-see posture; three are sceptical, with significant concerns; and only one is negative and prefers not to use AI. The positive and very positive categories together represent 71% of respondents; the neutral category represents 25%, and the sceptical or negative posture represents only 4%. The internal consistency of the attitudes scale is 0.777.

Future outlook items reinforce the optimistic posture. When asked what share of architectural firms in the Kurdistan Region will use AI tools within the next five years, 38% of respondents predict 61 to 80%, 22% predict 41 to 60%, 20% predict 81 to 100%, 15% predict 21 to 40%, and only 3% predict 0 to 20%. The combined share predicting at least 61% regional adoption is 58% of respondents. When asked about their own firm’s plans, 34% intend to increase usage significantly, 36% intend to increase usage moderately, 8% intend to maintain the current level, 1% intend to reduce usage, and no respondent intends to stop using AI. The combined growth-oriented share is therefore 70% of the eighty-two firms that answered the question. When asked about interest in future AI training workshops, 61% are definitely interested, 32% are possibly interested, 3% are unsure, and only 3% are probably not interested. The total share of interested respondents approaches 93%.

Open-ended responses on concerns provide texture to these statistics. Five themes recur. The first theme is the loss of design originality and the displacement of architectural intuition. One respondent writes, “Architecture is the creation of feeling, and I do not believe AI can create feeling.” A second respondent warns of “the danger that AI poses to the architect’s role.” The second theme is dependence and de-skilling. One respondent writes that “using AI for so long may eventually push architects away from architectural thinking”, echoing the cognitive-atrophy concern raised in the international corpus. The third theme is local relevance and language. Several respondents note the weakness of AI tools when applied to local environmental conditions, materials and codes, and call for tools and tutorials in Kurdish and Arabic. The fourth theme is ethical and legal anxiety, including authorship, copyright and the use of client information. The fifth theme is optimism conditional on education and infrastructure, with several respondents arguing that universities and educational institutions must take responsibility for developing the necessary skills before AI can mature into routine practice.

Infrastructure conditions support most adoption decisions. 64% of firms describe their internet connection as good with occasional interruptions, 31% as excellent with rare failures, 5% as fair with frequent interruptions, and no firm reports a poor or very unreliable connection. Internet speed is concentrated in the 50 to 100 Mbps band at 32% of respondents, followed by the 10 to 50 Mbps band at 17%, the 100 to 500 Mbps band at 15%, the band over 500 Mbps at 5%, and the band below 10 Mbps at only 3%. A total of 28% of respondents do not know their connection speed. Hardware infrastructure is similarly favourable. A total of 48% of firms operate predominantly on high-end workstations capable of three-dimensional rendering, 26% on mid-range desktops, 15% on mixed configurations, 9% on standard laptops, and 2% are unsure. No firm reports relying primarily on outdated systems.

The software ecosystem reveals a CAD-and-rendering core supplemented by a thin BIM layer. AutoCAD is used by 96% of firms, 3ds Max by 86%, SketchUp by 83%, Photoshop and the Adobe Creative Suite by 78%, Revit by 65%, V-Ray by 63%, Lumion by 50%, Rhino by 18%, Enscape by 13% and ArchiCAD by 11%. The dominance of AutoCAD over Revit and the very low share of Rhino indicate that parametric and BIM-integrated AI plugins, which depend on Rhino-Grasshopper or Revit-Dynamo workflows, sit on a thin installed base. The thicker rendering layer, by contrast, supports the dominance of image-generation and AI-enhanced rendering tools observed in the tool frequency analysis.

4.1.5. Local AIMI Results and Internal Structure

Each firm receives a total score on the AIMI ranging from a theoretical minimum of zero to a theoretical maximum of thirty-seven. The empirical distribution of total scores in the Kurdistan Region sample has a mean of 18.14, a median of 20, a standard deviation of 9.85, and an interquartile range of 10 (see Figure 13). This places the average local firm within the Occasional maturity band. The internal structure assessment of the eight AIMI components yielded a Cronbach’s alpha of 0.922. In addition to providing transparency, this value can also be used for secondary consistency testing. However, it should be noted that Cronbach’s alpha is typically meant for reflective psychometric scales; that is, all scale items are assumed to measure a single latent variable that can be used interchangeably. The AIMI, on the other hand, is a formative maturity index—each of the eight components of the AIMI theorises distinct dimensions of integrated artificial intelligence rather than merely serving as interchangeable indicators of a single underlying construct. A firm may score high on workflow breadth but low on governance, or high on tool diversity but low on training, and those differences are substantively meaningful rather than measurement error. For formative indexes, the appropriate validity evidence is theoretical justification of each component, face validity against the literature, and component-level interpretability, all of which are documented in Section 2.5 and Section 3.3. The alpha value is therefore not interpreted as a reliability coefficient; the index is evaluated through its component structure and its correspondence with the maturity bands defined in Table 4.

Although alpha is not interpreted as a reliability coefficient, several additional pieces of evidence support the index’s internal coherence and validity. Each component correlates with the sum of the remaining components in the expected direction, with item-rest correlations ranging from 0.67 to 0.78 (all p < 0.001), indicating that every component contributes coherently while remaining non-redundant; no pair of components correlates above 0.90, confirming the absence of the redundancy that would be a concern for a formative index. Known-groups validity is supported by the finding that firms providing formal AI training score significantly higher on the AIMI than firms relying on self-directed learning (Mann–Whitney U = 882, p < 0.001; rank-biserial r = 0.60, a large effect). The index also shows the expected convergent associations with perceived benefits (ρ = 0.41) and with a pro-AI attitude (ρ = 0.28). A bootstrap procedure (10,000 resamples) places the 95% confidence interval for the mean AIMI at 16.2 to 20.0.

To confirm that the equal-weighting and band-boundary choices do not drive the results, two robustness checks were conducted. First, recomputing every firm’s score under four alternative weighting schemes—an equal-influence scheme that rescales each component to a common range, an adoption-rescaled scheme, a process-weighted scheme that doubles the weight of capability, project application, and governance, and an empirical scheme using first-principal-component weights—left the maturity ordering essentially unchanged: the Spearman correlation with the primary index was 0.985 or higher in every case, and band classification agreed with the primary scheme for 85% to 94% of firms. The equal-influence scheme, which places all eight components on an equal footing regardless of their raw range, correlated at 0.998 with the primary index and produced identical band assignments for 90% of firms; the local-versus-international gap remained between 10.4 and 11.0 points under every scheme. Second, reclassifying firms under equal-width bands preserved the published classification for 84% of firms, and the headline pattern—a majority of firms at or below the Occasional band and only a minority reaching the Integrated or Advanced strategic bands—held under all schemes tested. The maturity profile is therefore not an artefact of the scoring or banding choices.

The eight-component decomposition reveals an asymmetric maturity profile. On a one-to-five scale, workflow breadth has the highest mean, at 3.84; usage intensity has a mean of 2.56; and tool diversity has a mean of 2.40. The lower means are concentrated in the resource-intensive and people-intensive components, namely commitment at 2.18, capability and training at 2.14, staff involvement at 2.04 and project application at 1.96. The adoption component, scored on a zero-to-two scale, has a mean of 1.02, consistent with the dominance of occasional rather than regular adoption documented in Section 4.1.2. The pattern indicates that local firms have moved beyond awareness and basic visualisation use into the wider workflow, but have not yet built the scaffolding for training, staffing, project coverage, and commitment needed to convert experimentation into integrated practice. The full coding matrix for local firms is provided in Appendix A.

In Figure 14, the five maturity bands distribute the sample as follows: 18% of firms fall into the Non-adopter band with totals up to 10, 15% fall into the Exploratory band with totals from 11 to 16, 30% fall into the Occasional band with totals from 17 to 22, 24% fall into the Integrated band with totals from 23 to 28, and 13% fall into the Advanced strategic band with totals from 29 upward. The combined share of firms in the upper two bands is 37%, while the combined share in the lower three bands is 63%. The 18 firms in the Non-adopter band correspond to the 14 respondents who reported never using AI tools and the 4 who tried AI tools and discontinued use. This indicates that most firms remain at or below the occasional maturity level, and only a minority reaches the Integrated or Advanced strategic levels, confirming the overall assessment of broad but shallow AI integration.

Across firm-size groups, AIMI shows a descriptive pattern: medium firms have the highest mean at 21.32, followed by large firms at 19.57, small firms at 18.49, and micro-firms at 15.58. The Kruskal–Wallis test does not reach statistical significance (H = 5.85, p = 0.119), and the effect size is small (ε² = 0.03); the ordering by firm size should therefore be read as a descriptive, non-significant pattern rather than an established effect. Experience and firm tenure show no meaningful differences, indicating that maturity is distributed across the sector rather than concentrated in one age or experience group (see Figure 15).

The findings on firm-size patterns in the Kurdistan Region sample differ from results in the comparison region of Saudi Arabia, which shows that large (>100 employees) firms are the main leaders in AI adoption, with medium-to-large (41–100 employees) firms not having nearly as high an adoption rate [41]. However, the Kurdistan Region data show that medium-sized firms (16–30 employees) lead the AIMI average, while very large firms do not differ significantly from one another. This difference in findings between the Kurdistan and Saudi Arabian samples suggests that firm size is not an indicator of AI maturity in either Gulf region in the same way; therefore, it follows that the distribution of firms in the Kurdistan Region shows that medium firms are located in a productive zone where they have employees for AI use, but due to their smaller offices, can adjust their workflows quickly.

In Figure 16, Spearman intercorrelations among the AIMI components range from approximately 0.55 to 0.77, with no pair above 0.90. This supports an additive composite because the components are related but not redundant. The internal-structure check using Cronbach’s alpha (0.922) confirms that the components move together to a reasonable degree. Still, this figure is not interpreted as a reliability coefficient because AIMI is a formative index rather than a reflective psychometric scale. In a formative index, components are theoretically distinct causal dimensions rather than interchangeable indicators of a single latent variable. The appropriate validity basis for the AIMI is therefore the theoretical grounding of each component in the literature (Section 2.5), its face validity against professional practice, and its interpretive correspondence with the maturity bands in Table 4, all of which have been documented. For composite validity in formative frameworks, the absence of redundancy among components (no pair with r> 0.90) is a stronger indicator than alpha, and this condition is met.

The AIMI total is positively associated with perceived benefits (Spearman rho = 0.41, p = 0.000156) and with the pro-AI attitude composite (rho = 0.28, p = 0.0046). Still, it is unrelated to the perceived challenge composite (rho = −0.05, p = 0.673). The study employs a descriptive comparative design without directional hypotheses; accordingly, all inferential tests are treated as exploratory rather than confirmatory. A Mann–Whitney U test comparing AIMI totals for firms with formal training against self-learning-only firms shows that the two distributions differ significantly (U = 882, p < 0.001, two-tailed). The direction of the difference is read from the group means rather than from the test itself: firms with formal training have markedly higher mean AIMI scores than self-learning-only firms. This descriptive pattern is consistent with the broader literature linking structured upskilling to digital-tool adoption maturity. These checks show a consistent pattern: local AI adoption is broad, but the deepest gaps appear in project penetration, training/capability, tool diversity, staff involvement and governance/strategic commitment.

Table 8. Summary of inferential tests, reliability, and validity statistics.

Test	Result
Chi-square: Firm size × adoption status	χ2 = 14.79, df = 9, p = 0.097
Chi-square: Experience × adoption status	χ2 = 7.47, df = 12, p = 0.825
Kruskal–Wallis: AIMI across firm size	H = 5.85, p = 0.119, epsilon-squared = 0.03
Kruskal–Wallis: AIMI across experience bands	H = 4.50, p = 0.342, epsilon-squared = 0.01
Kruskal–Wallis: AIMI across firm tenure	H = 0.61, p = 0.895, epsilon-squared = 0.00
Spearman: AIMI ↔ mean perceived benefit	ρ = 0.41, p = 0.000156
Spearman: AIMI ↔ mean perceived challenge	ρ = −0.05, p = 0.673
Spearman: AIMI ↔ pro-AI Likert composite	ρ = 0.28, p = 0.0046
Mann–Whitney: AIMI by training status	U = 882, p < 0.001 (two-tailed)
Cronbach’s α: AIMI (8 components)	0.922 (internal-structure check; formative index—see Section 4.1.5)
Cronbach’s α: Benefits scale (10 items)	0.858
Cronbach’s α: Challenges scale (10 items)	0.785
Cronbach’s α: Likert attitudes scale (10 items)	0.777
Item-rest correlation: each component vs. sum of the others	0.67–0.78 (all p < 0.001)
Component independence: maximum inter-component correlation	<0.90 (range 0.55–0.77); no redundant pair
Effect size: AIMI by training status (rank-biserial)	r = 0.60 (large)
Bootstrap 95% CI: AIMI mean (10,000 resamples)	16.2–20.0

provides the summary of inferential tests, reliability, and validity statistics.

4.2. International Interview Data Analysis

The international interview strand provides qualitative benchmark evidence rather than statistical survey evidence. The benchmark firms were coded using the same eight AIMI components, but the scores are interpretive benchmark coding scores derived from structured qualitative responses. They are therefore used to identify mature practice characteristics and comparative reference points rather than to make population-level statistical claims. All benchmark coding was conducted using the structured rubric defined in Table 5, which provides anchored keyword descriptors for each score level on each component. The qualitative benchmark coding places Firms A and B at the top of the AIMI with 37 points. Firms C and D each score 35, Firm E scores 33, and Firm F scores 31 points and is also classified as an advanced strategic. Firm G scores 28 and is classified as Integrated, Firm H scores 21 and is classified as Occasional, Firm I scores 15, and Firm J scores 14 are classified as Exploratory. These scores summarise coded interview evidence rather than statistical survey measurement as outlined in

Table 9. Qualitative, benchmark-coded AIMI scores for the ten international benchmark firms.

Firm	AIMI Total	Maturity Band	Strongest Coded Evidence	Interpretation
Firm A	37	Advanced strategic	Routine multi-tool use, Applied R&D capacity, governance, internal ecosystem	Advanced benchmark for organizationally embedded AI
Firm B	37	Advanced strategic	In-house AI platform, Rhino/Revit integration, data security, live project deployment	Advanced benchmark for secure project-linked tooling
Firm C	35	Advanced strategic	Design-technology-led experimentation, broad tool use, structured learning routines	Advanced benchmark for design-technology culture
Firm D	35	Advanced strategic	Computational design culture, pervasive AI-supported design exploration	Advanced benchmark with strong design and computation alignment
Firm E	33	Advanced strategic	AI-native practice identity, proprietary systems, routine concept and presentation use	Advanced benchmark for AI-first studio practice
Firm F	31	Advanced strategic	Routine but selective use, embedded power users, prompt libraries and custom workflows	Lower advanced benchmark with transferable hybrid model
Firm G	28	Integrated	AI image generation linked to environmental and computational analysis	Integrated benchmark with competition-oriented AI use
Firm H	21	Occasional	Gemini and ChatGPT are mainly for visualisation and presentation iteration.	Occasional benchmark with limited formal structure
Firm I	15	Exploratory	Cautious use of consumer tools by individual designers	Exploratory benchmark
Firm J	14	Exploratory	Partial individual use, no formal training or dedicated AI staff	Exploratory benchmark

.

The cross-case analysis of the ten interviews identifies four themes that the scores compress. First, Mature International firms invested in Artificial Intelligence primarily to improve decision-making options from conceptualisation through competitive analysis rather than to reduce time. Most respondents noted that there is now a greater variety of visual alternatives; therefore, they explore many more possible directions and examine larger option spaces in the same amount of time as past projects.

Second, human review is universal. Advanced and exploratory firms alike treat AI outputs as drafts that require professional judgement before client-facing use. The strongest firms formalise this through senior review, approved tool rules, and restrictions on technical or contractual deliverables.

Third, governance precedes scale. Firms that have moved furthest have resolved questions about data security, confidentiality, intellectual property, and approved tools before deploying AI widely. Lighter governance in exploratory firms still depends on individual responsibility and clear boundaries regarding non-use.

Fourth, organisational mindset and workflow discipline matter more than tool inventory. The interviews repeatedly indicate that value comes from how AI is integrated, reviewed, and connected to design judgement, not from the number of tools purchased. The full international firm-by-firm coding matrix with keyword matching is provided in Appendix B.

4.3. Local-International Comparison and Gap Identification

The benchmark gap analysis compares the Kurdistan Region component means (from survey results) to the qualitative benchmark indicators (related to international firms). Since both of these measures are derived from fundamentally different measurement processes (one from structured quantitative surveys, and the other from qualitative interviews), the differences presented in

Table 10. Survey-derived means for the Kurdistan Region compared with qualitative benchmark coding scores for international firms.

	The Kurdistan Region Survey Means	International Coded Mean	Benchmark Gap
Adoption	1.02	1.8	0.78
Usage intensity	2.56	4.1	1.54
Tool diversity	2.40	3.9	1.50
Workflow breadth	3.84	3.9	0.06
Staff involvement	2.04	3.9	1.86
Training/Capability	2.14	3.6	1.46
Project application	1.96	3.8	1.84
Governance/Commitment	2.18	3.6	1.42

should be interpreted as directional maturity indicators, rather than statistically tested differences. Further, subtracting one mean from the other does not convey inferential precision; instead, it simply provides an approximate sense of where the greatest distance between the two groups lies. Thus, with that interpretive caveat clearly stated, the benchmark gap analysis shows that the largest difference is related to Staff Involvement, with the Kurdistan Region component mean being (2.04) and the international component mean being (3.90), resulting in an approximate difference of (1.86) scale points. The next-largest gap in approximate scale points occurs in Project Application (1.84), followed by Usage Intensity (1.54), Tool Diversity (1.50), Training/Capability (1.46), and Governance/Commitment (1.42). Adoption shows a gap of (0.78) while workflow breadth is the closest dimension at approximately (0.06). The pattern indicates that local firms have started using AI across similar task domains, but have not yet embedded it as deeply in staffing, projects, training systems or organisational commitment.

Because the international values are qualitative coding scores rather than survey-derived measurements, the gaps in Table 10 should be interpreted as directional differences in maturity, not as statistically tested differences. Figure 17 shows comparison of AIMI component scores with the adoption components.

Table 11. Maturity band distribution for the Kurdistan Region and international benchmark firms.

Maturity Band	Kurdistan n	Kurdistan %	Benchmark n	Benchmark %
Non-adopter	18	18.0	0	0.0
Exploratory	15	15.0	2	20.0
Occasional	30	30.0	1	10.0
Integrated	24	24.0	1	10.0
Advanced strategic	13	13.0	6	60.0

presents maturity band distribution for the Kurdistan Region and international benchmark firms.

The band comparison backs up the dimensional model (see Figure 18). Of international benchmark firms, 60% fall in the Advanced strategic category, versus 13% of Kurdistan Region firms; 18% of Kurdistan Region firms are in the Non-adopter category, and there are no benchmark firms in this category. The Integrated category contains 24% of Kurdistan Region firms and 10% of International firms, while the Occasional category contains 30% of Kurdistan Region firms and 10% of international firms. This shows that, locally, the distribution is highly concentrated in the lower-middle bands, while the distribution of international benchmarks is mostly in the top band. These four themes provide the qualitative scaffolding that the Section 4 uses to convert the dimensional gap analysis into actionable sector recommendations and design guidelines for the local sector.

The headline finding states that AI is used widely in the Kurdistan architectural context compared with an international reference, but at a lower level of depth. With an 82% rate of local adoption indicating that AI is now part of everyday use, the distribution of maturity bands shows limited training, use of AI in projects, personnel involvement in AI, governance of AI usage, and budgeting for AI usage. The number of local firms classified as Integrated or Advanced Strategic is 37%, whereas the number classified as Advanced Strategic is only 13%, compared to 60% for the International Benchmark Firms.

Table 10 shows that the component gap supports this view. The breadth of workflow is the closest dimension (mean for Kurdistan = 3.84 versus 3.90), indicating that local firms are utilising AI across multiple workflow areas. The persistent gaps are institutional: involvement of staff (2.04 versus 3.90), penetration of projects (1.96 versus 3.80), intensity of use (2.56 versus 4.10), diversity in tools (2.40 versus 3.90), training and capability (2.14 versus 3.60), and governance/strategic commitment (2.18 versus 3.60). Therefore, the local issue is not about awareness; rather, it is about providing the supporting structures to enable AI to become part of the daily routine in an ongoing, reliable, and accountable manner.

The pattern does not imply that firms in the Kurdistan Region have not yet adopted AI tools. It suggests that the rate of AI tool adoption is outpacing the adoption of AI-related training, methodologies, and project governance. The predominance of small firms, the rapid adoption of tools, limited formal training resources in Kurdistan, and the lack of local case studies using disciplined AI methods in the region all contribute to this.

Three structural elements may help explain the gap concentration. To begin, micro- and small firms represent 74% of the sample space and therefore have limited budgets and personnel capacity for organised training and design technology, and they are also unlikely to form commercial partnerships. In addition, there is a poor distribution of education, with only 6% of participants first exposed to AI through university, whereas 73% were first exposed through social media. Additionally, local clients very rarely request or pay for AI-enabled workflow processes, whereas many international firms operate in markets or compete where digital sophistication is valued.

Two patterns of data are consistent with this interpretation. There is a positive relationship between the use of AI (through AIMI) and perceived benefits, meaning that organisations that invest heavily in AI will receive greater benefits. Further, there is almost no correlation between AI (AIMI) and perceptions of challenges, indicating that the reliability and control challenges faced by mature organisations relate to the strength of their routines for managing these issues. The barrier is therefore primarily at the organisational level rather than being purely technical. These relationships are interpreted as exploratory associations and should not be read as causal proof of organisational barriers.

Comparison with the RIBA AI Report 2025

Compared with the RIBA AI Report 2025, the two samples in Kurdistan differ significantly. According to RIBA, about 59% of architectural practices use AI [42], while the Kurdistan Region sample shows that over 82% of firms use AI on a regular or periodic basis. This does not suggest that practices in the Kurdistan Region are at the same level of maturity; however, the AIMI data show that the overall average score for the Kurdistan Region sample is 18.14 out of 37 points. Thus, the Kurdistan Region sample is classified in the Occasional maturity band. The results show that while AI is widely adopted in firms in the Kurdistan Region, it remains very weakly institutionalised. The greatest overlap with RIBA occurred with below average governance, policy, performance analysis, and structured organisational preparation. As such, the Kurdistan AI condition can be characterised as high adoption but medium maturation: firms are actively experimenting with AI across multiple workflow areas using various products, but still lack proper training, project-level routines, quality control systems, and governance frameworks.

4.4. Benefits, Challenges, Readiness, and Practical Recommendations

4.4.1. Benefits

The most dominant component of the benefit hierarchy is time savings, followed by enhanced visual quality and higher-quality client presentations. This is consistent with worldwide research demonstrating that AI’s value is greatest at the concept and visualisation stages, where option density and visual communication are highly significant. Cost savings and competitive advantage through design communications are both rated lower than the other components, indicating that local firms view their use of AI primarily as a presentation aid rather than as a complete productivity system.

4.4.2. Challenges

The leading reason for challenges across all levels of the challenge hierarchy is output quality inconsistency and a lack of control over the results produced by those outputs, while cost, software integration, internet speed, hardware, and learning curve are less important contributing factors. The binding constraint that prevents deeper integrations from occurring is therefore the level of confidence that professionals have in AI-generated outputs, which depends on their continued development of knowledge/skills, reviewing work habits, and having established governance. This conclusion aligns with the findings of Agri et al. [22], whose interviews revealed that the primary barrier to a more comprehensive integration of AI into architectural design management lies in the reliability of AI-generated outputs and the absence of a clear governing framework. Therefore, the key limitation to advancing AI in the context of architecture design management stems from organisational constraints rather than technological limitations.

4.4.3. Readiness

Participant feedback indicates overall acceptance of the AI systems and their use: 71% reported a positive or very positive experience, 70% intend to increase their use of AI, and 93% express interest in additional training on AI. While there is also little agreement that AI will take away architecture jobs and/or that it is too complicated for a small firm to implement, the most common barrier to using AI appears to be educational, with 61% agreeing that architecture schools and universities should be teaching AI technologies.

The neutral wait-and-see participants will be affected by the available credible local examples. The most requested supports were affordable training workshops, university courses, tutorials in Kurdish and Arabic, and professional peer communities. Each of these resources will support the establishment of a practical regional capacity programme.

4.4.4. Practical Recommendations

Based on the empirical findings above, the following ten recommendations are organised across three tiers: firm-level actions (R1–R5), sector-level actions (R6–R8), and policy- and client-level actions (R9–R10).

Tier 1—Firm-Level Recommendations (R1–R5)

R1—Consolidate the Tool Stack. Move from tool experimentation to a stable set of three to five tools, evaluated against output quality, IP risk, and integration with existing software. Avoid adopting every new release before existing tools have been embedded in verified workflows. (AIMI component: Tool Diversity & Governance)

R2—Appoint an AI Champion. Designate one trained senior architect per firm as the AI practice lead, responsible for maintaining prompt libraries, reviewing AI outputs before release, and disseminating lessons learned across the team. Allocate 1–2 h per week to this role. (AIMI component: Training & Capability)

R3—Implement Workflow Entry and Exit Protocols. Define the points at which AI enters and exits the design workflow. Recommended entry points include concept ideation, facade studies, mood boards, client-presentation imagery, and project narratives. Recommended exit points are senior review checkpoints and a prohibition on unverified AI content in contractual or regulatory submissions. (AIMI component: Workflow Breadth & Review Governance)

R4—Restrict AI Use in Contractual Outputs Until Verified. Every AI-generated deliverable must be approved and signed off by a senior architect before external release. AI should not be used for contractual, compliance, or regulatory outputs without a documented human verification process. Flag all AI use explicitly in project records. (AIMI component: Adoption Depth & Governance)

R5—Formalise a One-Page AI Policy and Project Register. Adopt a single-page office AI policy covering approved tools, confidentiality rules, senior review requirements, internal labelling of AI-assisted outputs, and prompt and output record-keeping. Maintain a project-level register that tracks AI tool use, tasks, and review status across all projects. A realistic target for firms currently using AI occasionally is to reach approximately 50% of annual projects using AI within one calendar year. (AIMI component: Strategic Positioning & Governance)

Tier 2—Sector-Level Recommendations (R6–R8)

R6—Integrate AI Literacy into Architecture Curricula. Architecture schools in the Kurdistan Region should embed at least one AI tool and an ethics module at the undergraduate level, addressing both practical skills and the critical evaluation of AI outputs. This directly responds to the finding that only 6% of survey respondents first encountered AI at university, while 61% agree that AI should be formally taught in architecture programmes. (AIMI component: Training Pipeline)

R7—Establish a Regional AI Training Platform. Professional associations (including the Kurdistan Engineers Syndicate) and universities should collaborate to deliver affordable, Kurdistan-focused AI training workshops. Monthly half-day sessions, supported by local case studies, peer communities, and tutorials in Kurdish and Arabic, would fill the current training gap at an accessible cost. (AIMI component: Sector Capability)

R8—Create a Shared Prompt and Failure-Mode Repository. Develop a regional shared resource where firms can contribute and access tested prompts, common error patterns, and verification checklists for architectural AI use. This collectively reduces the individual learning cost and accelerates sector-wide governance capability. (AIMI component: Collective Governance & Knowledge)

Tier 3—Policy- and Client-Level Recommendations (R9–R10)

R9—Adopt a Voluntary Code of Conduct. Professional bodies should publish a voluntary code of conduct that establishes expectations for senior review, project disclosure, privacy, internal labelling of AI-assisted visuals, and documentation of the tools and prompts used. Benchmarking against RIBA and AIA guidance would help build client and regulator confidence without creating overly onerous regulation. (AIMI component: Governance & Policy)

R10—Publish Sector Guidance on AI Use in Professional Submissions. Regulators and client bodies should issue clear guidance specifying when AI-generated content is acceptable in planning submissions, tender documents, and construction drawings. Reducing the current ambiguity will encourage disciplined use and establish a clearer accountability framework across the sector. (AIMI component: Policy & Client Confidence). Figure 19 demonstrates synthesis of findings and development priorities for AI integration maturity among architectural firms in the Kurdistan Region.

4.5. Limitations of the Study and Future Directions

Several limitations of this study should be acknowledged explicitly. First, the local survey relies on self-reported data, which may overestimate actual AI adoption levels compared with what would be observed through direct workflow analysis or a review of project documentation. No independent validation of survey responses against project records was conducted, and this represents a direction for future triangulation work. Second, the international benchmark sample (n = 10) was purposively selected for its visible AI engagement and cannot be treated as representative of global architectural practice. The benchmark scores represent a reference ceiling rather than a population mean, and the gap figures in Table 10 should be interpreted accordingly. Third, the cross-sectional design captures a single point in time in a rapidly evolving field; findings may not reflect the state of AI adoption at the time of reading. Fourth, this study does not assess the impact of AI tool use on project performance, design quality, client satisfaction, or business outcomes—a limitation that future longitudinal and case-study research should address. Fifth, the AIMI is an exploratory instrument developed for this study context; while internal structure validation (alpha = 0.922) and inter-coder reliability (kappa = 0.96) provide initial support, the framework requires further validation through expert Delphi consensus on component weighting, independent replication in other regional and professional contexts, and longitudinal re-measurement before it can be treated as a fully validated maturity instrument. Finally, the local and international strands were measured by different means—a structured survey of local firms versus primarily researcher-led qualitative coding of international firms—so the two sets of AIMI scores are interpretive reference points rather than statistically equivalent measurements. The gap figures should be read in that light.

Future research should repeat the AIMI over several years, test it in related construction-sector professions, and examine localised Arabic and Kurdish AI tools for regional codes, materials and design vocabulary. The research presents a structured evidence base, which will elicit an understanding of the integration and maturity levels of AI integration into architectural practice in the Kurdistan Region, as well as guide the transition to using AI through a broad, shallow approach to a more routine, governed, and disciplined manner of integrating AI into the architecture profession. The sensitivity analysis conducted in this study demonstrates that the primary findings are robust to alternative imputation rules for uncertain responses, but external validation with independently collected data remains a priority. Additionally, mixed-methods follow-up studies that combine survey data with direct workflow observation, project documentation analysis, and client outcome data would strengthen the evidence base regarding the relationship between AIMI scores and actual practice quality.

5. Conclusions

This paper evaluated the maturity of AI integration within architectural practices in the Kurdistan Region of Iraq and compared the local profile with that of selected leading international benchmark firms. It addressed a gap in the literature: many studies discuss AI tools in architecture, whereas fewer examine how firms in emerging professional contexts organise AI as a matter of maturity. The evidence shows that AI has already entered local architectural practice, but its integration remains uneven. The comparative benchmark analysis reveals a mean AIMI gap of 10.46 points between the Kurdistan Region sample (mean = 18.14/37) and the international benchmark group (mean = 28.60/37). This gap is not uniformly distributed across components: the largest dimensional gaps are in staff involvement (benchmark gap = 1.86), project application (gap = 1.84), and usage intensity (gap = 1.54), while workflow breadth shows the smallest gap (0.06), indicating that local firms already apply AI across a comparable range of workflow stages but have not yet achieved the depth, consistency, or organisational embedding that characterises mature international practice. These findings directly answer the primary research question: AI integration maturity among architectural practices in the Kurdistan Region is at the Occasional band on average, substantially below the Advanced Strategic band that characterises the majority of international benchmark firms, with the gap concentrated in organisational and governance dimensions rather than in tool awareness or access.

The primary finding indicates that awareness and adoption of AI tools in architectural design are already quite substantial. A notable 92% of respondents expressed familiarity with these tools, and 82% of firms reported utilising AI, whether regularly or occasionally. The most prevalent applications locally include text support, visualisation, rendering, concept generation, facade ideas, design alternatives, presentations, and project narratives. Consequently, the architectural practices in the Kurdistan Region should not be characterised as pre-adoption. Instead, the sector is more accurately described as broadly aware and partially engaged with AI.

The second key finding is that maturity remains limited in the components that necessitate organisational oversight. The local AIMI profile demonstrates strengths in workflow breadth, usage intensity, and tool diversity. Still, it shows weaknesses in areas such as commitment/governance, capability and training, project penetration, and staff involvement. This suggests that while AI has begun to permeate various task domains, it has yet to be fully stabilised through systems of training, project documentation, budgeting, internal governance, or dedicated review processes.

The third primary finding indicates that International AIMI scores should be interpreted qualitatively, relying on benchmarks derived from structured interviews rather than through statistical analysis of surveys. These benchmarked scores consistently reveal a maturity gap across their categories. The categories are organised from the one exhibiting the largest maturity gap to the one with the smallest: (1) Staff Involvement perspective, (2) Project Penetration perspective, (3) Intensity of Use perspective, (4) Diversity of Tool Use perspective, (5) Training & Capability perspective, and (6) Governance/Strategic Commitment. While the breadth of workflow is comparable to that of local firms employing AI across various operational areas, they currently do not utilise AI to the extent indicated by the benchmarks. Overall, the comparison shows a pattern of high local adoption but lower institutional maturity: the Kurdistan sample approaches benchmark firms in workflow breadth but falls behind most clearly in staff involvement, project penetration, usage intensity, tool diversity, training and governance. The AIMI should therefore be treated as a transparent preliminary diagnostic requiring further validation rather than as a final universal maturity standard.

The practical implication for the local firm’s next stage is moving from simply adopting more AI tools to establishing workflow discipline with those tools. This will involve conducting base-level audits of minimums for each firm, establishing AI Champions, creating prompt libraries, creating a way to obtain senior review of all AI-supported work, establishing clearly defined boundaries for the technical and contractual purposes of AI use, providing structured training on AI, and having modest but defined budgets for AI. The universities, professional associations, and the Kurdistan Engineers Syndicate can assist firms in making this change by offering workshops on AI use, providing local case-study libraries, and creating an optional AI Code of Practice.