SOCIAL MEDIA ZEITGEIST

 

Three-quarters of architectural practices now use artificial intelligence on at least some projects. Only 17% of users think it has made their designs better. Seven per cent say it has enabled them to increase their fees.

That gap is the story of the RIBA’s AI Report 2026. Based on a survey of more than 1,100 built-environment professionals, it captures a profession moving rapidly into AI while remaining deeply uncertain about where it is going. Adoption has risen from 41% in 2024 to 59% in 2025 and 74% this year. Almost a third of practices now use AI on every or most projects, and 91% of current users expect their use to increase.

Yet only 19% have a documented AI policy. AI is entering architectural production faster than practices are developing the commercial models, governance and training systems needed to manage it. The question raised by the RIBA report is no longer whether architects will adopt AI. It is who will capture the value, who will carry the risk and what kind of profession will result.

Read the 74% carefully

The headline figure covers practices using AI on at least some projects. It does not mean that three-quarters have transformed their work. Only 31% use it on every or most projects and 15% say it is embedded in day-to-day practice. Sixty-two per cent remain at the exploratory stage, with another 17% formally trialling it.

Size is the strongest predictor of uptake. More than 90% of practices with at least 50 staff use AI, compared with 63% of those with between one and nine staff. Age is not: adoption remains consistently high among respondents with between one and 30 years of industry experience. The assumption that AI is primarily a young architect’s technology does not survive contact with the data.

The survey was self-selecting and some questions changed this year. Its results should therefore be read as a strong indication of direction, rather than a definitive census of the profession.

Pictures and paperwork – but not yet the physics

Early-stage visualisation is the most common application, used by 56% of AI-adopting practices. Compliance checking follows at 40% and practice management at 32%. Use then falls sharply: 10% for BIM creation or maintenance, 7% for building-performance simulation and 6% for environmental-impact modelling.

Pictures and paperwork, then, but not yet the model or the physics. UCL associate professor Tom Holberton describes this as AI’s “asymmetrical impact”. Models have advanced around the text and images available online, making them effective at research, reports and visualisation. They are much less capable with the structured drawings, models and data through which architecture is coordinated and delivered. Describing a detail is not the same as resolving one. Holberton raises a further problem: when plausible documents can be generated almost instantly, their existence no longer proves that the work behind them has taken place.

“As generative models expand, information becomes less useful as evidence that a team has done any thinking or work.”

Provenance and verification become more important as professional-looking information becomes cheaper to produce.

Productivity is arriving before better design

The clearest positive finding concerns efficiency. Seventy-three per cent of users report improved productivity and 57% a positive return on investment. Forty-four per cent say AI has released time for creative, higher-value work.

The design results are far less convincing. Thirty-eight per cent believe AI helps optimise design outcomes, but only 26% think it has improved buildings for users or communities. Just 17% say their designs are better because of AI, while 42% disagree.

AI can accelerate the production of options without improving the judgement used to choose between them. It can create a convincing image before the proposal has been tested.

Speed improves quality only when the capacity released is reinvested in design, technical scrutiny, consultation or client care. Human oversight remains essential—but it means little unless the reviewer has the knowledge, time and authority to reject an unreliable answer.

The finding that should concern anyone pricing work

Despite the reported productivity gains, only 7% of AI users say it has enabled their practice to increase project fees; 58% disagree. Only 13% of all respondents expect AI to support higher fees in future. Productivity gains do not automatically belong to the architect. In a market where practices frequently compete on price, efficiencies may be passed to clients through lower fees or expanded scope. They can also be absorbed by software, training, checking and governance. The profession could deliver more work more quickly without improving profitability.

Anyone who worked through the transition to BIM will recognise the pattern. BIM brought real coordination benefits, but also software, training and compliance costs – often before practices knew how to charge for them. Katie Fisher’s essay in the report warns that technological democratisation can create new hierarchies between firms with different investment capacity:

“If access to powerful AI tools is gated by subscription costs, technical literacy, and proprietary platforms, then ‘democratisation’ becomes a hollow phrase.”

There is little evidence yet of market expansion, but 57% of practices still plan to invest in tools and 50% in training. They therefore need to decide where the value should appear: in better design, a wider service, reduced risk or stronger margins. Without a commercial strategy, AI risks becoming another mechanism through which architects deliver more for less.

The governance gap

Only 19% of practices have a documented AI policy. Almost half have neither a policy nor a plan for one, despite 74% using the technology. Among large practices, 56% already have one; smaller firms are much further behind.

Once AI touches live information, governance is not abstract. Practices must decide which tools staff may use, what data may be entered, whether inputs are retained for training, how outputs will be checked, when use must be disclosed and who remains accountable.

Dan Rossiter of CIAT and BSI cites two instructive cases: the £20 million deepfake fraud against Arup and a Polish contractor that lost a public contract after its submission included non-existent legal rulings. The second should concern every bid team. Fluent, authoritative text is not evidence that a source exists.

“Given the lack of transparency, there is no choice but to ‘trust’ the output.”

Rossiter argues that the central issue is not trust but transparency. A spreadsheet formula can be examined; the path by which a generative model reaches an answer is generally opaque. BS ISO/IEC 42001 offers a structure for AI management, but every practice needs agreed boundaries, checks and accountability. Copyright law is still evolving, while data-protection responsibilities cannot be outsourced to a software provider.

Automating the bottom rung

There is little evidence yet of wholesale job losses: 9% of practices say AI has led to staff reductions. But 59% expect reductions across architecture in future, and 61% believe AI will make it harder for early-career professionals to acquire the skills they need. AI is strongest at high-volume, pattern-based work: research, schedules, reports, standard details and document comparison. These are also the tasks through which inexperienced staff learn how buildings, contracts and practices work.

Automate the bottom rung and the ladder remains, but becomes harder to climb. Protecting every repetitive task makes little sense; removing them without creating another route to competence is equally shortsighted.

Olivia Stobs-Stobart of Plan A offers a practical response: map a process before automating it and involve the people closest to the work. “Focus on metrics oriented around skill, and not time,” she advises. “Skill is a valuable commodity.”

Practices must replace incidental learning with deliberate exposure to source information, technical reasoning and senior decisions. Otherwise, today’s efficiency may become tomorrow’s shortage of judgement.
Harder to become an architect, easier to look like one.

Seventy-seven per cent agree that AI can never replace human creativity. At the same time, 76% believe it increases the risk of imitation and 55% think it will enable people without adequate qualifications to design buildings.

Generative systems can make proposals appear resolved long before they are technically, socially or environmentally credible. They also draw on existing work whose provenance is often unclear. The risk is not only copying, but convergence around the styles, places and assumptions most heavily represented in training data.

Nenpin Dimka goes further: “AI is reshaping the decision infrastructures that determine what is built, not merely the tools used to draw it.” If these systems privilege Western architecture or regulation, they can influence architectural possibilities before an architect begins drawing.

What practices should do now

RIBA has promised an AI Overlay to the Plan of Work and a responsible-use policy during 2026. Practices need not wait. The immediate priorities are straightforward:

1. Write a short policy defining approved tools, prohibited information, checking responsibilities and uses requiring senior approval.

2. Apply controls in proportion to risk, and train staff to verify sources, recognise plausible errors and understand when a task should not be delegated.

3. Protect the learning pathway by involving early-career staff in trials and exposing them to the reasoning above automated production.

4. Measure quality, risk reduction, service and skills—not simply hours saved—and decide where the value will go. If an efficiency does not improve the offer or business, it is likely to disappear into lower fees or additional scope.

The divide that will matter

AI is established enough to be useful, widespread enough to be unavoidable and immature enough to remain dangerous. It is improving productivity, but has not delivered an equivalent improvement in design quality, fees or collaboration.

The next phase will be less about spectacular outputs than institutional choices: what practices automate, what they protect, how they train, what they disclose and where they insist on human responsibility. The profession may divide between those who use AI and those who do not. On this evidence, however, the more consequential divide will be between practices that use it deliberately and capture some of its value, and those that simply absorb the cost and risk of keeping up.

The RIBA survey ran from March to April 2026 and received more than 1,100 responses. Respondents were self-selecting, making the findings indicative rather than statistically representative. The report also contains commissioned essays whose views do not necessarily represent RIBA’s position.