No. AI (artificial intelligence, which in practice means the language and image models everyone has tried by now) cannot hold a license, carry liability, or seal a drawing, and nothing about the legal structure of practice is changing on a timescale a firm owner plans around. What AI will do, and is already doing, is absorb a set of tasks: minutes, first drafts, code lookup, and some production work.
So the honest answer to “will AI replace architects” is: not the architects, but a real share of the hours, and the hours it takes are concentrated at the junior end of the studio. That is a staffing problem, and staffing problems are solvable.
What do the numbers say about architects’ jobs?
Start with the boring data, which is more reassuring than the headlines and less reassuring than the profession would like.
The BLS (Bureau of Labor Statistics) Occupational Outlook Handbook, puts architect employment at about 124,600 jobs in 2025 and projects 4% growth from 2025 to 2035. That is modest growth, not decline, and the projection was published after language models became ordinary office tools.
NCARB (the National Council of Architectural Registration Boards) counts more than 123,000 licensed architects in NCARB by the Numbers 2026, and the path to joining them still runs through 3,740 hours of AXP (the Architectural Experience Program) and the six divisions of the ARE (Architect Registration Examination). More than 39,000 candidates are on that path. There is no version of that exam a model can sit.
Firm structure matters too. The AIA (American Institute of Architects) Firm Survey Report 2024 finds that about three in four firms have fewer than 10 employees, while firms with 50 or more people are a few percent of firms but produce more than half of total billings. Large firms have production departments to automate. A six-person firm is already mostly judgment, coordination, and client contact, which is the part AI does worst.
Which parts of the job can AI actually do today?
Think in tasks, not jobs. The traditional fee split (AIA Handbook of Professional Practice, stable for decades) puts SD (schematic design) at 15%, DD (design development) at 20%, CD (construction documents) at 40%, bidding at 5%, and CA (construction administration, the phase after the permit when the building gets built) at 20%. CD is the obvious target for automation, and it is where the tools are furthest from working. Text-to-BIM is a demo, not a workflow.
What AI does well today lives in language. 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. In practice the work is mostly text: meeting minutes, RFI (request for information) drafts, proposal narratives, a first pass at zoning or code research. The pillar page goes through each.
Notice where most of that sits. Deltek Clarity 2026 puts firm-wide utilization just under 60%, and the proposals, admin, and marketing in the other 40% are where the first hours come out. That is real money. It is also not replacing architects. It is replacing the Sunday-night proposal.
Where AI touches billable work, it touches junior work. Take a 30-person firm with a 40,000-square-foot TI (tenant improvement) on a six-week permit clock. The PA (project architect) spends the week on the landlord’s work letter, the mechanical engineer’s shaft, and a plan checker who wants the egress diagram on the cover sheet. The junior spends it on the code summary, the plan-check response letter, and the finish schedule. A model can draft the first two today. It cannot do any of the PA’s week.
Why can’t AI seal a drawing?
Because the seal is a legal act by a person, and every mechanism around it assumes a person.
Licensure is granted by a state board to an individual. The AIA Code of Ethics and Professional Conduct requires licensees to seal only work they have reviewed and are competent to judge. Professional liability insurance attaches to a firm and its people, and the standard of care is what a reasonably prudent architect would have done. A tool cannot be sued into learning.
Ownership points the same way. 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, and the D.C. Circuit’s 2025 decision in Thaler v. Perlmutter affirmed that requirement. A building nobody designed has no author, which makes it hard to own and hard to license to an owner. The copyright post has the details, and where AI falls flat has the practical failures behind the legal ones.
Boards and legislatures could change all of this, but rules of this kind move over decades, and a firm plans over years.
Who should actually worry?
Four groups, honestly.
- Staff whose day is mostly tasks. Code lookup, minutes, first drafts, and eventually redline pickup once tools can read drawings. If a role is 80% of that list, the role shrinks even if the person does not.
- Firms that sell hours instead of judgment. A task billed hourly that now takes a fifth of the time is a fifth of the fee. The fee structure post is worth rereading with that in mind.
- Drafting-only service providers. Production without design responsibility is the most exposed business there is.
- The training pipeline. This is the one that deserves the worry. The profession trains people by having them do exactly the work AI is best at. AXP needs 3,740 hours of real experience, and if the junior work goes to a model, those hours have to come from somewhere else, on purpose.
Who should not worry: anyone whose value is the client relationship, the coordination call at 4:30 on a Friday, the site visit, or the judgment call on a flashing detail. That list has not changed since drafting boards.
What should a firm do in the next 12 months?
- Re-price judgment. Look at every hourly contract for tasks AI has shortened, and move toward fixed fees for scope where you carry the risk and keep the gain.
- Move the saved hours somewhere billable. Many firms report CA consuming 25% to 30% of effort against a 20% budget (practitioner consensus, not a survey figure). Hours saved on minutes can go to the site visit that prevents the RFI.
- Train juniors deliberately. Have the junior check the model’s code summary against the code, then have a senior check the junior. The junior learns by verifying, which is roughly how they used to learn by drafting.
- Clean the content. A model drafts from what your firm gives it. A project template that saves 20 hours and a curated detail library are the difference between a useful draft and a confident mess.
- Write the policy, then train. The AI policy post and the training post cover both.
- Measure. Hours before, hours after, by task and by phase. Without that you are guessing, and so is everyone at the conference.
The profession absorbed CAD, BIM (building information modeling), and email. Each one moved the work up the ladder and made the bottom rung harder to reach. AI is doing the same thing faster. The firms that lose are the ones that let the junior work disappear without replacing the training it used to provide.
AECPulse Project Manager shows which phases and which people the saved hours are actually landing on, because a firm that cannot see its hours cannot tell whether AI is helping or just hiding the work.