AI for Architects in 2026: RFIs, Code Research and Your Seal

AI can now draft fee proposals, give submittals a first read, answer code questions with citations and turn site dictation into field reports. It can't hold responsible control, decide an RFI or seal anything. Use it freely on the business of your practice, as a checker during construction, and only under your direction on technical submissions.

Most lists of AI tools for architects stop at renderings, and rendering is a small slice of the week. The rest is contract documents, consultants, plan review comments and a contractor who wants an answer by Friday. This guide sorts that work by how close each task sits to your seal, because that distance tells you how much to hand over.

The paper behind the drawings, and how many firms use AI

Architecture carries more contract paper than its reputation suggests. The Bureau of Labor Statistics describes the job as designing buildings and also preparing contract documents, managing construction contracts, checking designs against building codes, zoning and accessibility rules, and visiting worksites. O*NET's list of architect tasks adds representing clients in bids and writing proposals.

Adoption is running ahead of results. Deltek's 2026 Clarity study of 896 architecture and engineering firms reports AI use at 70%, up from 53% a year earlier, while only 38% of respondents report a measurable business benefit. In the UK, RIBA's 2026 report says 74% of practices use AI on most projects. AIA's 2025 research found about 6% of US architects used it regularly, so firm-level adoption and daily habit can be very different things.

AI for an architect's paperwork, task by task

AI helps an architecture firm most where a tool can read a document you already have and hand back a draft or a flag. The table runs from the work furthest from your seal to the work closest to it.

The business of practice: hand over the first draft

Fee proposals are the safest place to start, because a wrong first draft costs you an edit, not a building. Monograph says its AI Proposal Generator reads your lead context and generates a proposal ready for review. That helps with the scope narrative and phase descriptions, and with a public call for proposals that asks for the same boilerplate again. Set the number yourself, and check the exclusions against your last similar job.

The fee model itself deserves a second look. An AIA Architect article from May 2026 reports that faster AI-assisted work may put hourly billing under pressure, while one attorney it quotes warns that flat fees demand cost and schedule certainty. If a tool cuts the hours in a phase you bill by the hour, your proposal is pricing that gain away.

Construction administration: a first reader, never the respondent

Submittal review is the clearest fit, as long as the tool stays inside your role. AIA's own explainer says that under B101-2017, § 3.6.4.2 you review submittals for the limited purpose of checking conformance with the design concept. Under A201 § 4.2.7, absent an approved schedule, you act with reasonable promptness while leaving time for adequate review.

Nomic reads the package, cross-references specs and drawings and flags deviations, with individual plans from $20 a month, and says the licensed professional still reviews and stamps every package. InspectMind's submittal checker offers a free check of up to 10 drawing pages and says approval authority stays with the architect or engineer. One caution of our own: don't paste a tool's "compliant" labels into your returned comments. That wording can claim a broader review than your agreement describes.

RFIs need more care, because the answer can change the work. A tool can find the sheet and spec section in question. Contractors have drafting help too: Procore's RFI Agent produces pre-populated RFIs, so a polished question is no proof of a real conflict. Read the cited sheets before you respond.

If the job runs on the contractor's Procore account, ask what your login can see, and learn the limits. Procore's Assist help page says this AI assistant can't read image files or markups in the Drawings tool, and that new activations are paused while Procore moves to its next AI. An answer from Assist can miss a redline your team drew.

For field reports, Cogram drafts from dictation in your templates, works offline and offers 14 free days; pricing comes with a demo. Read every report before it goes out, because "installed per drawings" becomes the record the moment you issue it. And on change orders, AI can compare a contractor's request with the spec and your earlier instructions, but AIA's documents have the owner, contractor and architect sign.

Technical submissions: direct the work while it's made

Responsible control is the rule that shapes everything here, and it looks at how a document was prepared, not only how it was checked. Rhode Island's architecture rules, in wording several other states share, define it as control over and detailed professional knowledge of a submission during its preparation. They add that reviewing or correcting work others prepared doesn't count, and they require records of your control or review for at least five years after sealing. Your own board's text governs, so read it.

Those rules don't mention AI, and NCARB doesn't want regulators limiting tools. But its position, updated in April 2026, says any regulation must keep the licensed practitioner in responsible control, and that you remain answerable to the standard of care. It also lines up with a joint regulators' statement that only a licensed practitioner may seal technical submissions. A workflow where software produces a sheet set and you look it over at the end sits badly with that language. Directing and correcting as the set develops, with notes showing you did, sits much better. The AIA AI Task Force says the architect remains the professional of record.

For code research, UpCodes says Copilot links every answer to the adopted code section, remembers jurisdiction and code year, and can generate QA/QC checklists. It comes with the Professional plan: $59 per user monthly on a yearly plan, $68 if paid month to month. The 93% accuracy on its page is the company's own figure, so open the section it cites, and confirm local amendments and interpretations with the authority having jurisdiction.

For specs, Deltek's Dela Insights flags a referenced requirement missing from the project or a product type mentioned in only one part of a section. Deltek says it supports professional judgment rather than replacing it, and that coordination with drawings and Revit isn't part of this release, so it won't catch a spec that disagrees with your details. In the model, Autodesk Assistant answers questions like how many columns sit on a floor and creates a door schedule from a prompt. Autodesk calls it a tech preview with a focused set of workflows.

AI renderings: the copyright question to ask first

Typing a prompt doesn't make you the author of an AI image, in the view of the Copyright Office's January 2025 report. Protection needs enough human expression in the result, such as your own work visible in the output or your creative changes to it. The D.C. Circuit held in Thaler v. Perlmutter that the Copyright Act requires a human author, while leaving open how much human input is enough. So a rendering generated from your own model has a stronger human contribution to point to than an image made from a sentence, though no court has drawn that line for buildings.

Tools built for architects can start from your geometry. Chaos Veras works with Revit, SketchUp, Rhino, Archicad and others, at about $16.90 a month for Starter and $29 for Pro on annual billing. Check whether the tool keeps or trains on your uploads before you feed it an unannounced project, and tell clients when an image is a mood study rather than a design the drawings support.

What to settle before project data goes into any AI tool

Four checks come before anyone buys a seat.

• Your owner agreements. AIA's AI Firm Toolkit lists provisions to look for: training restrictions, disclosure requirements, ownership terms, indemnities and audit rights, plus flowing those terms down to subconsultants.
• Your data classes. The same toolkit uses four levels (public, internal, confidential or client-covered, and security-sensitive) and treats any tool whose terms you haven't verified as public. A secure facility's plans don't go in a free chatbot.
• Your insurer. The toolkit advises having your professional liability insurer review a firm AI policy before you adopt it. At renewal, ask your broker directly whether any new AI exclusion applies to you.
• Recording consent. Federal law allows one party to consent, but California's Penal Code requires all parties' consent for confidential talks, and an OAC call can span states. Say so aloud when you hit record.

Where Merlin fits for architects

Merlin handles the correspondence around a project: the structural engineer's question, the client's third look at the tile, the plan reviewer's follow-up. This Mac app follows who you work with and what you owe each of them. Onboarding insists on Gmail plus Google Calendar, so it won't suit a firm on Microsoft 365 mail alone.

Merlin doesn't manage RFI or submittal logs, read your model, research code, write specs, price fees, send invoices or answer phone calls. Those jobs belong to the tools above.

• Who's waiting. If the civil engineer emailed you directly last week and you never answered, the People panel shows it, up to 30 days back. Merlin judges from mail alone, so a question asked on a site walk won't appear.
• Consultant threads. Ask in chat what the MEP engineer last said about the ceiling, and Merlin reads your mail to answer. Each person also gets a timeline and your own notes.
• Drafts you approve. A reply to a consultant that you request in chat arrives as a card with the text ready, and stays unsent until you confirm.
• OAC calls. The Meetings tab records on your Mac on the free plan. Speaker-labeled transcripts come with paid plans, from $29 a month, and are stored on your account, so apply your own data classes. Our note taker comparison covers other options.
• Repeat mail. Routines fire on a timetable or on matching Gmail, such as summarizing plan review messages each morning. They work unattended and can act without a card, so keep sending out of them until you've tested one.

How to test AI on one project this week

A closed project is the fairest test, because you already know what a careful reviewer caught.

1. Download AIA's short sample policy from the toolkit, about two pages, and mark it up for your firm. Send it to your insurer.
2. Pull ten old submittals with known deviations, from a project whose owner agreement allows it. Run them through a first-pass checker whose terms you've read, and count what it caught, missed and invented.
3. Dictate your notes from one past site visit into a field report tool, then compare the draft with the report you issued.
4. Ask a code tool three questions you've already resolved with a plan reviewer, and read each cited section.
5. Keep a log of tool, task, data class and reviewer. The toolkit's detailed policy includes an AI Use Log.