Subscribe
Share
Share
Embed
Reimagining AEC is Dr Sepehr Abrishami's independent podcast on how the built environment handles information. Each episode takes one subject from digital construction, from the ISO 19650 series and the common data environment to the golden thread, soft landings and the arrival of artificial intelligence in design and surveying practice, and tells it as a single argument built on the published standards, guidance and evidence. Sepehr is a researcher and former architect who leads the Centre for AI in Construction and teaches building information modelling at master's level. The episodes are written for postgraduate students and practitioners who want the reasoning behind the rules, not a summary of them.
A door schedule lists every door on a building with its type, size, fire rating and hardware. If my name, Dr Sepehr Abrishami, is new to you, the part of my work that matters here is my research on artificial intelligence, or AI. That research includes generative AI, which produces text, images and data from learned patterns, and practices already use AI to draft reports and project documents. Reimagining AEC is a podcast I produce outside my teaching, and here I read the RIBA AI Report 2026, from the Royal Institute of British Architects, or RIBA, on how architects are using AI.
Several architects interviewed for the report expect AI soon to produce schedules consistently and accurately. Some welcome the prospect, because it would release them for more creative work, while others see a cost. One architect places door schedules among the routine tasks through which an architect in training learns the job. So a practice deciding who compiles its next door schedule, an AI tool or its newest graduate, is choosing between time and training. The report does not make that choice for it. Behind the choice sits a second one, about who answers for the schedule if the tool gets a door wrong.
In the survey behind the report, nearly three quarters of respondents say their practice now uses AI on at least some projects. Yet among those whose practices use it, only 17 per cent agree that their designs are better because of it. This episode works inside that gap, and its question is where AI belongs in an architect's work, and who stays accountable for what it produces. The aim is that you can read the survey with its sample and limits in view, and attribute each essay's argument to the person who made it. From there you can judge where AI fits in a practice's workflows, in the stages of the RIBA Plan of Work 2020 and in a practice's governance. I begin with who answered the survey and what they report doing. Then I turn to what respondents believe AI is doing for them, and what worries them. The essays follow in two groups, four about the practice itself and five that look beyond it, and the door schedule supplies the answer at the end.
One fact about the report shapes everything else. It is an edited collection. Chris Williamson, President of RIBA, writes the foreword. Adrian Malleson, RIBA's Head of Research, reports the survey. Named contributors write the essays, and the report states that their contributions are the views of their authors and not the position of RIBA. So when I describe an argument, I give it to its author rather than to RIBA.
Who answered comes first. Malleson places a short note about the survey at the end of his 2026 chapter, and it deserves to be read before any percentage. Over 1,100 people responded between March and April 2026. Just over half, 51 per cent, came from small practices of one to nine staff, and 27 per cent from large practices of 50 or more. Architects were by far the largest group, and the Chartered Institute of Architectural Technologists, or CIAT, circulated the survey to its members. Around one respondent in ten was based outside the UK.
The limits of that sample fall under two concerns. The first is who answered. Respondents selected themselves, so the report says its results are best read as a good indication of AI and the profession, but not as definitive. People also answer about their practice, and the report does not say whether two respondents from one practice count twice. That point is my own reading, not the report's. The second concern is how the answers are counted. Not everyone answered every question. The questionnaire changed for 2026, so some comparisons with earlier years, including the uses of AI, are no longer like for like. And some sub-groups do not add up to the headline. The split by practice size reconciles with it. The split by location does not. With about nine respondents in ten in the UK, use of 85 per cent in the UK and 73 per cent elsewhere would imply a total near 84 per cent, well above the headline of 74. Most bands of experience also sit above it. The report does not explain these bases, so keep both concerns beside every figure that follows.
Within those limits, the trend is consistent. In 2024, 41 per cent of respondents said their practice used AI at least sometimes. The share was 59 per cent in 2025 and 74 per cent in 2026. Intensity rose faster. Use on every or most projects grew from 6 per cent of practices in 2024 to 31 per cent in 2026. Size divides the profession. More than 90 per cent of practices with 50 or more staff use AI, against 84 per cent of medium practices, with 10 to 49 staff, and 63 per cent of small ones. Large practices also use it more intensely, with 41 per cent using it on every or most projects, against 28 per cent of medium practices and 29 per cent of small ones. Experience, by contrast, barely matters. Only those with more than 30 years in the industry stand apart, at 59 per cent, so the report rejects the idea that AI is mainly for younger professionals. RIBA Chartered Practices also report higher use than practices that are not chartered, 76 per cent against 67.
Breadth, however, is not depth. Asked how far AI has gone in their practice, only 15 per cent say it is now embedded in day-to-day work. A further 17 per cent are formally trialling it, and 62 per cent are starting to explore. For most, the report concludes, adoption remains exploratory. Users expect that to change. Among them, 91 per cent expect their use to increase over the next two years. Investment plans follow, with 57 per cent planning to invest in AI tools and 50 per cent in training. Practices that do not use AI are a different population. Of these, 24 per cent will definitely not adopt it for project work in the next 12 months, and a further 41 per cent think adoption unlikely. Not every practice will adopt AI. For the door schedule, this means most practices are at the stage of trying such a task, not of building it into a standard workflow.
What practices use AI for falls into two groups, by how far each use reaches into technical design. The most reported support early design, the rules a design must meet and the running of the office. Early design stage visualisation leads, used by 56 per cent of practices that use AI, standards and regulatory compliance checking follows at 40 per cent, and practice management at 32 per cent. Uses at the technical end are reported less often. Specification and technical design reach 24 per cent, and building information modelling, or BIM, 10 per cent. On compliance checking, the report adds that AI is not a replacement for professional judgement and oversight. Now place those uses on the RIBA Plan of Work 2020, the RIBA framework that divides a project into stages from strategic definition through to use. The mapping is mine rather than the report's. The heaviest use clusters in Stage 1, Preparation and Briefing, and Stage 2, Concept Design. The door schedule belongs mainly to Stage 4, Technical Design, where reported use is lower. So the survey answers the first half of our question descriptively. AI has arrived mostly at the front of the project and in the office, and only lightly at the stage where a door schedule is made.
That is what practices do. What they believe AI is doing for them comes next, and the answers form a gradient from process to product. Among AI users, 73 per cent report improved productivity and 57 per cent a positive return on investment. Agreement falls as the claims move towards the building. About a quarter, 26 per cent, agree that AI has improved outcomes for building users or communities, and only 17 per cent that their designs are better. Malleson adds two cautions in his chapter. The productivity gains suggested in the early days of BIM were perhaps never fully realised. And AI companies are running very significant operating losses, so the future cost of AI is highly uncertain and unlikely to fall, and future returns may look different. Creativity draws the same restraint. Only 30 per cent agree that AI enhances creativity, while 77 per cent agree that AI can never replace human creativity. For a door schedule, a saving of time is easy to report. Whether the schedule is more correct is a separate claim, and the survey gives little support for it.
Concern is strongest where AI touches the standing of the profession and the people in it. On standing, 76 per cent agree that AI increases the risk of imitation, and a majority, 55 per cent, that it will enable people without sufficient professional qualification to design buildings. Chris Williamson's 2026 foreword names the same risk, of plausible-looking designs produced at speed by those without adequate qualification. Fees offer no relief, with only 7 per cent of AI users agreeing that AI has let them increase project fees. On people, 59 per cent expect AI to reduce staff across architecture, but only 30 per cent in their own organisation. And 61 per cent agree that AI will make it harder for early career professionals to gain the skills they need. Malleson explains why. AI is most readily used for set tasks done in a prescribed way, and those are often the tasks through which people first become professionally adept. He asks where future architects will come from. The door schedule is one of those tasks. A third concern sits outside the practice altogether, with 65 per cent agreeing that the energy AI consumes will set back global sustainability goals.
The profession expects change without expecting improvement. Some 93 per cent expect AI to have at least a moderate effect on practice within five years, including 45 per cent who expect a major effect. Yet only 17 per cent say AI makes them more optimistic about the future of architecture, and 49 per cent disagree. Governance has not kept pace with use. Around 19 per cent of practices have a documented AI policy, a slight increase on the year before, and a further 17 per cent are developing one. Size divides the profession again. Of large practices, 56 per cent have a policy, while 43 per cent of small practices have no plans to develop one. In his foreword, Chris Williamson commits RIBA to developing an AI Overlay to the RIBA Plan of Work and a policy for responsible use of AI during 2026. When I checked in November 2026, I could not find either published, so look on the RIBA website for their current status. Surveyors, by comparison, already work under a professional standard of 2025 from the Royal Institution of Chartered Surveyors, or RICS, on the responsible use of AI in surveying practice. It has been in effect since 9 March 2026. In a practice with no policy, nobody has decided whether AI may draft the door schedule, or who checks it. Use, in short, has outrun both confidence and governance.
The essays come next, starting with four that look at how AI enters the working life of a practice. Giuseppe Bono, formerly Head of Design Computation at TP Bennett, argues in his 2026 essay that discussion of AI in architecture dwells on its most visible outputs, which are aesthetic. He locates its larger value in the practice as a business, with fees, scheduling, compliance, tendering and contract administration. His examples share one principle, that AI is most useful on the repetitive, document-heavy work of running a practice. It can draft first versions of fee proposals from the practice's templates, and it can extract obligations, deadlines and responsibilities from contracts, provided qualified practitioners review its outputs. He describes agentic AI, meaning systems that carry out tasks through AI agents, as making work more consistent and more predictable, and calls its adoption as much an organisational transformation as a technical upgrade. His conditions are just as clear. Practices must invest in training, workflow redesign and governance, and stay alert to data security, accountability, deskilling and unrealistic expectations of speed. His essay presents prototypes and possibilities rather than measured results, although the survey's finding that 32 per cent of AI-using practices apply it to practice management suggests his direction is under way.
Bono looks at functions. Olivia Stobs-Stobart, Design Manager and AI Lead at Plan A Consultants, looks at people, and her 2026 essay states its position in its title, people first, technology second. Her lessons rest on two principles. The first is that context comes before tools. AI used without knowledge of the project, the client and the team produces outputs that may look coherent but miss the point, and in her words, integration is not installation. Plan A's own pilot brought together leadership, technical staff, sceptics and advocates, and found that junior professionals taught it the most. After launch it stepped back, so that AI did not become the drumbeat of every conversation. The second principle is to make the work explicit before automating it. Before automating its bid process, Plan A mapped a process that had lived in people's heads, and found that the mapping had value whether or not AI was used. She also asks practices to define success before they begin, with measures oriented around skill rather than time. For the door schedule, that means writing down how the practice compiles and checks one before any tool is chosen, and deciding which skills the task should still teach.
The third essay carries the voices of members themselves. Sharon Anderson, an independent researcher, reports qualitative research that RIBA conducted with its members in April 2026. Its stated aim was a deeper understanding of AI use rather than a quantitative assessment, and the essay does not give the number of participants, so treat it as testimony rather than measurement. Use was often limited to large language models for practice management, though several practices described wider use across the early stages of the Plan of Work. Members valued help with bids, planning applications, research and images. Among the risks they named, hallucination, meaning plausible output that is false, was the primary concern. One senior architect at a large practice warned that "if AI hallucinates garbage that is not spotted, lawsuits await". Mitigation centred on the human in the loop, thorough checking and a live AI policy, kept under review, and some members prefer not to adopt AI at all. Members also expect architects to become orchestrators of AI agents. Our running example comes from this essay. One architect described "many menial tasks, such as door schedules", along with contract reviews, as "really crucial to an architect in training".
The fourth essay gives a way of thinking about the information AI produces. Tom Holberton, Associate Professor at University College London, begins his 2026 essay with history. In 1962 RIBA surveyed practice to list the tasks that need to be done, and within a year the first Plan of Work gave those tasks sequential stages. Computation later moved attention from managing tasks to managing information. His concern is that generative models produce information by inference rather than calculation, filling in plausible detail from patterns. His response separates three ways of using AI by the control each leaves. Describing treats AI as a search space, prompted with words or images, so speed of iteration is traded against control of detail. Deriving transforms one representation into another, such as a massing into a render, so inferred versions multiply and knowing which to trust becomes hard. Doing generates actions that drive standard software, so its actions are legible and auditable even when its reasons are not, and the work can be supervised as a person's would be.
Applying the three modes to the door schedule is my illustration, not Holberton's. Asked for a schedule, a chatbot describes one. A schedule derived from the drawings is inferred, and may, as he warns of inference in general, "inject specificity into a schedule where there should be none". An agent that compiles the schedule from objects in the model leaves a trail of actions someone can audit. On a project run under ISO 19650-2 of 2018, the schedule would face the same checks whichever route produced it. ISO 19650-2 is the international standard from the International Organization for Standardization, or ISO, for managing information on projects using BIM. Clause 5.6.3 requires a quality assurance check, and clause 5.6.4 requires the task team to review the information and approve it before it is shared. Holberton ends with a question for every stage of the Plan of Work, which is when inference shifts from creative expansion to a lack of control. Read together, the four essays on practice place AI where the practice understands the task, has written down how it is done, and can audit what the tool did.
The remaining five essays widen the frame beyond the practice. Nenpin Dimka, architect, academic and Co-Chair of the RIBA AI, Generative Design and Data Expert Advisory Group, questions the usual framing of AI as the successor to computer-aided design and BIM. That framing, he writes in 2026, is incomplete. AI is becoming part of the decision infrastructures that determine what is built, where and for whom. By decision infrastructures he means systems such as procurement platforms that rank bids, planning tools that pre-assess applications, and environmental, social and governance, or ESG, scores. Architects meet their outputs as constraints without seeing the logic behind them. Drawing on hybrid systems in African cities, he warns that training data dominated by Western precedent can erase local spatial identity. He casts the architect as an editor who reads these systems and reconfigures them.
Katie Fisher, founder of CARD Projects and a lecturer at Ravensbourne University, asks in her 2026 essay who AI is designed for. She finds the debate led by a narrow group, predominantly male and predominantly tech-led. She warns that bias does not disappear in code but becomes harder to see, and that automating entry-level work hollows out the learning ladder. She also presses on the water that data centres consume. Her preferred target for AI is repetitive administration. The door schedule sits on both sides of her argument, as administration and as training.
Philip Allsopp is a habitat systems scientist, a RIBA Trustee and the other Co-Chair of the Expert Advisory Group. He treats human habitat as one of the most important public health, economic and social determinants of our time. He cites large health effects of housing and neighbourhood design, but his 2026 essay gives no sources for those figures, so I do not repeat them. His argument is that AI's most transformative contribution is the quieter one of linking data that are rarely connected, and he sets it out as a sequence. It starts with data on built form, health, economics and exposure, each held in its own silo. Those records are then linked through shared identifiers, in what he calls a common data environment. The linking produces a place-based dataset that can be queried and simulated. That dataset can then inform decisions on finance, planning and design. He names one digital twin platform, Nextspace, as an example. Two cautions follow. His common data environment is broader than the one in the ISO 19650 series. There, a common data environment manages information containers through the work in progress, shared, published and archived states, as the episode on the common data environment explains. And he notes that AI in the built environment is still controlled by large technology players whose incentives may differ from those of its users.
Three members of Arup's Global Foresight team, Florence Van, Malina Dabrowska and George Harrington, write in 2026 about foresight. They mean the structured practice of thinking about possible futures, and they offer it as a way to work with AI's uncertainty. They draw a parallel with social media, welcomed for participation and transparency before its longer-term harms became clear. Foresight, they stress, does not predict. It explores plausible futures and asks better questions. Their example is the Port of Dover Strategic Masterplan 2050, where three scenarios were used to stress test assumptions. A consistent theme of Arup's AI for Future Cities series, they report, is that AI should remain a tool, with human oversight and responsibility maintained. They also ask what professional liability and insurance mean when adaptive systems keep shaping outcomes after a project is delivered. They urge architects to define the purpose of AI as well as its use, and to develop bespoke tools collaboratively across the profession. Their examples are largely Arup's own tools and projects, so read them as illustrations by their developers rather than independent evaluations. The essay also restates the survey's 74 per cent as individuals using AI tools, a small shift from a question about practices.
The last essay turns to trust. Dan Rossiter, Sector Lead at the British Standards Institution, or BSI, and Vice-President Technical at CIAT, cites misuse in the sector in his 2026 essay. An engineering firm, which his footnote identifies as Arup, lost 20 million pounds to a deepfake impersonating a senior officer. A Polish construction company had its contract award overruled because its submission cited legal rulings that did not exist. His diagnosis is anthropomorphism, the habit of treating software as someone rather than something. His correction is blunt. A computer infers rather than reasons, and confabulates rather than hallucinates, and he makes the same correction for seeing and thinking. People trust a spreadsheet, he argues, because its formulae can be audited. That is transparency rather than trust, and most AI tools lack it.
Rossiter's remedy is standards, and he names three. ISO and IEC 42001 of 2023, from ISO and the International Electrotechnical Commission, or IEC, sets out an AI management system. With it, a practice could establish checks and balances for bid writing assisted by AI. ISO 26000 of 2010 gives guidance on social responsibility, and the ISO and IEC technical report 24028 of 2020 addresses trustworthiness in AI. He still recommends adopting AI, within checks that standards help a practice set. A practice with an AI management system would decide, and record, how a door schedule drafted by AI is reviewed. He ends his essay with a statement he attributes to IBM in 1979, which I come back to at the close.
One last caution concerns the other surveys you may set beside this one. Two have their own episodes. One is the RICS report of 2026 on AI in commercial property and construction, by Tom Storey, Anil Sawhney and Katherine Pitman. The other is the NBS Digital Construction Report 2025, from the specification platform NBS. Their numbers are not comparable with RIBA's, because the questions, the samples and the units all differ. RIBA asks respondents about their practice, so its unit is the practice as each respondent sees it. Compare the pattern of findings instead. The RICS survey finds the same shape as RIBA's, with pilots common and embedded use rare, so use is spreading faster than it is embedded. The NBS authors read their own results as a move beyond experiment, while the architects in the RIBA survey are far more guarded. What the RIBA report adds is what only architects can report, their concerns about imitation, creativity, qualification and early careers.
That brings the evidence together, so return to the question. The survey gives the first half of the answer. For now, AI is used at the front of the project and in running the office, and it is only beginning to reach technical design. The essays give the second half. AI belongs where the task is understood, its method is written down and the tool's work can be audited, and where a named person answers for the result. A door schedule meets those conditions only if the practice sets them. AI can describe a schedule, derive one or compile one. On a project run under ISO 19650-2, the schedule would still pass its quality assurance check and its review before sharing. Beyond that, several of the report's contributors insist that a qualified person reviews what AI produces, and a policy decides how. So the choice between time and training is a false one. A practice can let a tool draft the schedule while a junior colleague checks it and learns from it, and a named person signs it off. Rossiter's statement from IBM puts the accountability point in two sentences. "A computer can never be held accountable. Therefore a computer must never make a management decision."
In your own practice and research, read every figure in the report as an indication from self-selected respondents, which is how the report itself describes it. Wide use and shallow embedding are both part of what it indicates. Keep claims about process apart from claims about product. A tool that saves hours on a schedule has not yet shown that the schedule is right, and the survey's own respondents draw that line. And treat accountability and learning as things a practice designs. A written policy, a named person who answers for each output and a protected place for junior staff to learn can all exist before any tool arrives.
The full report rewards reading in its authors' own words. The episode on the RICS report on AI in construction and property and the episode on the NBS Digital Construction Report 2025 read two other surveys with the same care. The episode on parties, teams and processes in the delivery phase covers the checks that ISO 19650-2 requires before information is shared.
Because I am Dr Sepehr Abrishami and not RIBA, every reading in this episode of Reimagining AEC that goes beyond the report's own words is mine to answer for. Ask of the next routine task in your own practice or research what this episode asked of a door schedule.