AI for architecture practices: where it actually helps

If you run an architecture practice and are weighing whether AI is worth the bother, it already earns its place on the admin and early-stage work, and not on final design. It saves real time on documents and concept iteration. The design judgement, and the accountability that comes with it, stay with a qualified architect.

AI is now mainstream in the profession, and adoption is moving fast. RIBA's Artificial Intelligence Report 2025 found 59% of practices using it, up from 41% the year before. Large practices are past 80%, while smaller studios sit around 48%. The useful question is not whether to adopt it, but where AI belongs in the work: mapped to the stage of the project, the reliability of the source information, and the level of professional responsibility attached to the output.

Where AI helps a practice now

The honest picture is a profession moving quickly onto the tool, but concentrating it in a few sensible places. RIBA found AI use most prevalent in early design visualisations and specification writing, where the time saving is immediate. The gains cluster in a handful of tasks.

Concept and early-stage iteration. AI image and generative tools let a studio explore massing, mood and options quickly at the front of a project. Used as a way to generate starting points for a human to judge and refine, this compresses hours of early exploration. It is a sketch pad, not a drawing board.

Documents and correspondence. Drafting reports, design-and-access statements, specifications, client updates and routine written work is slow and repetitive, and AI takes a real bite out of it. The output is a first draft that an architect edits, never a final document that goes out unread.

Analysis around the design. Environmental and sustainability work benefits from AI's ability to run many scenarios fast. Some 65% of architects expect AI to have an increasingly positive effect on productivity in construction, and early-stage performance analysis is where that promise is most concrete today.

Practice admin. Bid preparation, information chasing and project paperwork all take load off the team. This is the least glamorous use and often the most valuable, because it gives fee-earning people their time back.

Use AI by project stage, not just by task

AI is not equally safe at every point in a project, so the sharpest way to think about it is by stage. The RIBA Plan of Work sets out eight stages from strategic definition to use, and the risk is not flat across them. It is lower where outputs are exploratory or communicative, and it rises as they start to influence statutory compliance, tender information, construction decisions and professional liability.

The pattern is a risk gradient. For most practices the safest ground is stages 0 to 2: briefing, precedent research, feasibility framing, concept options and client explanation. From stage 3 onward the work carries more weight, because planning, coordination, specifications and technical design begin to bind. By construction and handover, AI is best kept to search, summaries, logs and checks, and away from anything that reads as professional sign-off. The later the stage, the stronger the audit trail and the review need to be.

AI is also an information-management issue

Architecture already runs on controlled information. A practice using BIM, CAD, a specification system or a shared project space has rules about versions, approvals and who is allowed to rely on what. AI has to sit inside those rules, not route around them.

That discipline has a standard behind it. Under ISO 19650, every piece of project information in a common data environment sits in one of four states: work in progress, shared, published or archived. Those states exist so that people know what is safe to build from and what is still a draft. An AI answer built on a superseded model, an old planning condition or an unapproved consultant note can be more dangerous than no answer at all, precisely because it looks finished.

So before AI touches project information, four questions keep it honest. Is this work in progress, shared or published information? Is the source current? Can the output be traced back to a drawing, specification or policy so a person can check it? And who signs off the result? A plausible answer nobody can trace is not a time saving. It is a liability someone has to find later.

What AI cannot do: judgement and accountability

The line sits at judgement. Final design decisions, the resolution of a difficult site, the subjective calls on aesthetics and client preference, and anything that carries professional liability stay with a qualified architect. The tool proposes; the architect decides and signs.

This is not caution for its own sake. AI produces plausible output that can be confidently wrong, and in architecture a confident wrong answer can reach a drawing, a specification or a planning submission. The profession seems to know this: RIBA found a clear view that AI will develop and augment the architect's role rather than replace it. Only 18% of practices expect AI to lead to job losses, and just 4% think human creativity will no longer be needed for building design.

Accountability is also formal, not just cultural. The ARB Architects Code that came into force in September 2025 builds its standards around competence and the responsible exercise of professional judgement, and it expects that judgement to be documented. A tool cannot hold competence or account for a decision. The named person still does, which is the same reason context is the real advantage: the model is generic, but your knowledge of the client, the site and the brief is not.

The risk rises when AI touches statutory material

Planning statements, design-and-access statements, building-control narratives and technical compliance notes are not ordinary copy. They can shape how a project is assessed, priced, approved, challenged or built, and other people rely on them: planners, contractors, inspectors, insurers and clients.

Planning itself is going the same way. The government has rolled out AI tools for local authorities, including one that turns planning documents into structured data and a prototype that helps officers weigh applications faster. The official line is worth borrowing: these tools support the decision, they do not make it, and, as the government puts it, the planning officer remains the decision maker. The same standard belongs on your side of the table. AI can draft a statement or a checklist, but it cannot know whether a site constraint has been read correctly or whether a planning condition is still live.

Building safety raises the stakes again. The Building Safety Act regime puts weight on competence, design coordination and demonstrable compliance, and it created a Building Regulations Principal Designer role, a distinct duty from the older role of the same name under the construction health-and-safety regulations. Where AI helps with technical or regulatory work, keep a clear record of what was generated, which sources were used, who reviewed it and who approved it. Treat AI here as a drafting and checking assistant, never a compliance authority, and never let it become the source of technical truth. If a mistake would reach a client or an inspector, it needs a human who can catch it first.

Check your insurance before AI reaches client work

Using AI does not move the liability. If a tool produces a wrong output that reaches a client, the responsibility generally sits with the practice, not the software vendor, and professional indemnity insurers are starting to ask how firms oversee and validate AI. The particular danger in architecture is repetition: one wrong assumption can travel through drawings, schedules, specifications, reports and correspondence, so a single bad input becomes many.

A simple way to gauge exposure is to sort the use by how far the output travels. Drafting a social post or an internal note is low risk. Summarising a meeting or a brief is low to medium. Drafting a planning statement is medium. Checking a specification against drawings is medium to high. Producing technical compliance wording is high. Making design decisions, certifying compliance or signing off defects is not something to delegate to a tool at all.

The practical response is to check your professional indemnity position before AI touches technical design, specifications, compliance material, planning submissions or tender packs, and to write a short policy that says which tools are approved, what information may be entered, which uses are off limits, and what review is required before anything leaves the practice. Our simple AI policy for small firms is a good template.

Generic AI is only the first layer

The first wave of AI in a small practice is usually generic: write this email, summarise this policy, make this image, turn these notes into a client update. That is useful, and it is where most studios start. It is not the whole story.

The larger shift is towards AI that works with architectural information itself: models, drawings, specifications, planning documents, queries and past project decisions. Practice-knowledge retrieval that surfaces a relevant detail or lesson from a previous job, specification checks that flag gaps and mismatches, and model querying in plain language are where the sector is heading. These tools are more powerful because they touch real project data, and for the same reason they need stronger controls: version discipline, data security, audit trails and competent review. For many practices the serious value will come less from replacing design work and more from finding, checking and connecting the information that already exists.

A practical starter setup

You do not need a strategy document to begin, but you do need a few clear rules. Adoption has run ahead of governance here: RIBA found only 53% of practices expect to have an AI policy within two years, which means most do not have one now, even as usage passes 59%. A small, deliberate start with the rules written down already puts you ahead of most.

• Pick one low-risk task. Report or specification drafting, or early concept iteration, is the natural first step. Prove the time saving there.
• Use a business-grade tool. Choose an account with a data agreement that keeps your drawings and client information out of model training, and check each provider's terms rather than assuming.
• Work only from the right information. Feed AI current, appropriate-status project data, not a superseded model or an unapproved note.
• Keep outputs traceable. Prefer results you can check back to a drawing, specification or policy, and record who approved anything client-facing.
• Review after a month. Keep what saved time, drop what created rework.

First steps for a small studio

Pick one recurring task your studio does most weeks that carries no design liability. A design-and-access statement or a standard client report is ideal. Run it through a business-grade AI tool for a month, with an architect editing every draft before it goes out, and compare the time saved against the corrections needed. That bounded test teaches you more than any vendor demo, and it keeps the design judgement, the IP and the client relationship firmly in your hands.

If it would help to work out where AI belongs in your practice and where a human must stay in charge, or how to choose help without the hype, book a call. We can also help you tell a good adviser from a good salesperson before you spend a penny.

Common questions

How many architects actually use AI?

RIBA's 2025 survey found 59% of UK architecture practices using AI, up from 41% a year earlier. Large practices lead, past 80% adoption, while smaller studios sit around 48%. Use is concentrated in early-stage and routine work such as visualisations and specification writing, rather than finished design.

Where in the RIBA stages is AI safe to use?

It is safest in the earlier stages: strategic definition, briefing, precedent research and concept design, where the work is exploratory. Risk rises from spatial coordination and technical design onward, because outputs start to affect planning, specifications, tender information and construction. By technical design, treat AI as a checking and searching aid, not a source of technical authority.

Can AI check a design against UK Building Regulations?

Not on its own. AI can support checking, raise useful questions and flag possible issues, but compliance with the Building Regulations remains a professional responsibility, and the Building Safety Act regime expects demonstrable competence and a clear record of who reviewed and approved the work. Keep a qualified person in charge and keep the paper trail.

Does using AI change who is liable if it gets something wrong?

No. If an AI output reaches a client and is wrong, the responsibility generally sits with the practice, not the tool vendor, and the risk is that one wrong assumption repeats across drawings, specifications and reports. Check your professional indemnity position before AI touches technical or client-facing work, and keep a human review that can catch a mistake before it travels.

What should AI not be used for in architecture?

Final design decisions, professional judgement, and anything carrying liability or safety implications. AI can be confidently wrong, so keep a qualified architect in charge of everything client-facing, and treat AI output as a draft to check. RIBA's own findings show the profession expects AI to augment the architect's role, not replace it.

Does AI put your practice's intellectual property at risk?

It can, and most architects think so: 67% are concerned AI raises the risk of their work being imitated. Public tools may reuse what you feed them. Use business-grade tools that keep your data out of training, keep confidential drawings and briefs out of consumer tools, and check the licensing terms before using AI-generated imagery commercially.

---

This is general information, not legal advice. Check planning, Building Regulations and other statutory questions with a competent professional.