As of September 2026, AI (artificial intelligence, which in a design office mostly means the large language models behind tools like ChatGPT and Microsoft Copilot) is dependable at the reading and writing that surrounds a project: meeting minutes, RFI (request for information) drafts, proposal text, first-pass zoning and code research, spec markups, and summaries of long documents. With supervision it is useful for early massing studies, existing-conditions capture, and flagging a project that is drifting over budget. It cannot produce a permit set, read your drawings, know your local amendments, or take responsibility for anything.
That is the honest state of AI for architects: a fast, tireless junior who has read everything and admits nothing.
This page is the map for the whole pillar. It gives you a working mental model, the numbers on who is actually using these tools, the places where they do real damage, and what a firm of 3 to 75 people needs in place before any of it pays off. The detailed posts are linked at the end.
What counts as AI for architects, and what does not?
Four different things get called AI, and they fail in different ways.
- Generative AI: language and image models that predict plausible words or pixels from patterns in their training data. Both the hype and the usefulness live here.
- Generative design: algorithms that search options against constraints you set, such as the massing and layout tools inside Autodesk products. Deterministic, a decade older, and a different technology. Generative design vs. generative AI untangles the two.
- Rule-based checking: clash detection, model checkers, and spreadsheet logic. Not AI at all, though vendors are now attaching language models to them as a front end.
- Pattern detection on your own data: the statistics behind budget early warnings and resource forecasts. Unglamorous, useful, and only as good as your timesheets.
The distinction matters because it tells you what a tool can promise. A rule engine is right or wrong in ways you can audit. A language model is fluent whether or not it is right, and it is right most often about general questions and least often about your jurisdiction, your contract, and your detail.
Which tasks does AI already do well in a design office?
The pattern is consistent. AI works where the input is text you already have and the output is a draft a human will read before it leaves the office.
- Meeting minutes: transcribe the OAC (owner, architect, contractor) meeting, pull out decisions and action items, then correct owners and due dates. A 45-minute chore becomes 10 minutes.
- RFI responses: draft from the question, the spec section, and your earlier answers, then check the draft against the drawings.
- Proposals: first drafts of the approach narrative from your past proposals. The fee, the team, and the schedule stay human.
- Research first pass: zoning, code, product data. Good at finding the right question and the right chapter, unreliable on the final number.
- Specifications: editing a master spec to a project and catching orphaned references. Not writing Part 2 from nothing.
- Existing conditions: point clouds and photos to rough models.
- Feasibility: yield studies on a parcel in an afternoon instead of a week.
- Budget early warnings: pattern detection on logged hours against phase budgets.
- Renderings: mood imagery in SD (schematic design). Not the marketing shot, and not anything with a dimension in it.
A concrete example. A 12-person firm runs a Tuesday OAC meeting on a 24-unit multifamily project. The PM (project manager) used to lose Wednesday morning to the minutes. Now the recording goes to a transcription tool, a model drafts the decisions and action items, and the PM spends 15 minutes fixing names, dates, and one decision the model recorded backwards. That backwards decision is the whole story: the tool saved 2 hours and nearly created a change order.
Deltek Clarity 2026 (the 47th edition of its A&E, architecture and engineering, industry study) found generative AI use rising from 64% to 78% of firms in a year, but only 38% reporting measurable business impact. That matches what we see. The wins are in words, not in drawings.
Where does AI still fall flat?
- It cannot read a drawing set. Language models read text. They treat a plan as a picture and a dimension string as decoration. Ask one whether your corridor is wide enough and it will say 44 inches, because that is what the IBC (International Building Code) asks for, not because it measured yours.
- It invents citations. Section numbers, ordinance clauses, ASTM designations, product model numbers. NIST (the National Institute of Standards and Technology) lists confabulation among the risks in its Generative AI Profile of July 2024. The word is polite.
- It does not know your jurisdiction. Training data lags local amendments by months or years, and the plan checker’s interpretation letters are not on the internet.
- It is confident about everything. A wrong answer and a right answer arrive in the same tone, so your review process has to assume every output is wrong until checked.
- It leaks. What you paste into a consumer tool may be retained or used for training. Client confidentiality clauses still apply.
- It cannot hold a design. Ask an image model for the same building from another angle and you get a cousin of it.
- It has no author. The U.S. Copyright Office’s January 2025 report on copyrightability states that purely AI-generated material is not copyrightable and that human authorship is required. The copyright post covers what that means for your drawings.
Where AI falls flat is the full reality check.
How many firms are actually using it?
Fewer than the conference talks suggest, and less formally.
The AIA (American Institute of Architects) Firm Survey Report 2024 counts something like 19,000 to 20,000 architecture firms nationwide, three in four of them under 10 employees. A firm that size has no technology staff. The principal is the help desk.
Deltek Clarity 2026 finds the top technology challenge is knowing where and how to apply it (62% of firms), ahead of integration (56%) and data security (52%). Read those findings together and the picture is clear: at most firms, adoption means two or three people using a chat tool on their own, with no policy, no shared prompts, and no record of what was checked.
The profession’s gate has not moved either. NCARB (the National Council of Architectural Registration Boards) counts more than 123,000 licensed architects in the United States in NCARB by the Numbers 2026, and the path still runs through 3,740 hours of AXP (the Architectural Experience Program) and the six divisions of the ARE (Architect Registration Examination). Nothing in that pipeline has changed because of AI.
What does a firm need in place before AI pays off?
Four things, and none of them is a subscription.
- Clean content. A model drafts from what you give it. If your detail library holds three parapet details for the same condition, the model will pick one, and not the one you meant. The same goes for master specs, Revit families, and old proposals. An organized detail library, a BIM (building information modeling) content standard, and a firm knowledge base are prerequisites, not extras.
- A policy. Which tools, which data, who checks. NIST’s AI Risk Management Framework 1.0 (January 2023) organizes this as Govern, Map, Measure, and Manage. For a 15-person firm that translates to one page, and the AI policy post has the template.
- Honest time data. Budget early warnings run on timesheets. If people enter time on Friday afternoon from memory, the model learns fiction. Low-friction time tracking comes first.
- A verification habit. Someone owns checking each output, and the checking time is booked to the project rather than disappearing. Deltek Clarity 2026 puts firm-wide utilization just under 60%. Most early AI wins sit in the other 40%, the proposals, minutes, and admin, which is fine as long as verification does not quietly eat what the drafting saved.
How should a small firm start?
Pick three tasks and measure them.
Start with meeting minutes, RFI drafts, and proposal first drafts. All three have a human reader built in, a low cost of error, and a before-and-after you can time. Leave code review, specifications, and anything client-facing until the policy exists and the team has burned itself on something harmless.
Run a four-week trial with one champion per studio. Log hours: a task that took 3 hours and now takes 1 is worth writing down, and so is the task that took 20 minutes and then needed 40 minutes of checking. Both happen. Training your team lays out a six-week onboarding plan, and what clients think covers disclosure.
What about liability, copyright, and the stamp?
The stamp does not move. The AIA Code of Ethics and Professional Conduct requires licensees to seal only work they have reviewed and are competent to judge, and no licensing board treats software as a licensee. The standard of care is unchanged, which means a model’s mistake in your code summary is your mistake in your code summary.
Three practical rules follow. Document where AI was used and who reviewed it, the way you would document a consultant’s input. Keep client and project data out of any tool the policy has not approved. And expect ownership questions in contracts, because the Copyright Office position on human authorship means an image or a paragraph that nobody at the firm materially shaped belongs to nobody. Will AI replace architects takes on the jobs question, and the copyright and policy posts above handle the rest.
Where to go next
- Will AI Replace Architects? An Honest Look: the jobs numbers, which tasks go, and why the stamp stays human.
- How to Write an AI Policy for a Small Architecture Firm: one page and four data classes.
- Where AI Falls Flat in Architecture (A Reality Check): the full list of what it does badly, so your policy can name it.
The firms getting value from AI are not the ones with the most tools. They have clean libraries, a one-page policy, and a habit of checking. Everything else in this pillar is detail.
AECPulse Project Manager puts the AI assistants that have earned their keep, meeting minutes, RFI drafts, zoning research, proposal drafts, and budget early warnings, inside the same system that already holds your timesheets and phase budgets.