Architect Meeting Minutes: Record, Summarise and Issue Same Day With AI

Ask everyone for consent before recording, capture the meeting with a transcription tool, then give the transcript to an AI assistant with a fixed template for decisions, actions and design-change flags. Check the draft against your own notes and issue it before you leave for the day. A 90-minute meeting needs roughly an hour of work afterwards.

The part that is easy to underestimate is not the AI step but what surrounds it: a clear yes from everyone to being recorded, a template that separates decisions from chatter, and a check against your own notes before anything is sent. Minutes are a record that clients, consultants and sometimes contractors rely on later, so a tidy summary that quietly turns "oak might be nice" into "oak approved" does more damage than late minutes would have done.

Follow me on Instagram@sagnikteaches

Ask for consent before you press record

Recording is the one step you cannot repair afterwards. Rules on recording vary from place to place and can depend on where each person on a call is sitting, so the safe habit is the same everywhere: get an explicit yes from every participant before recording starts, and put the request in the meeting invitation as well so nobody is surprised. This is not legal advice. If your appointment terms or professional body say more, follow them.

Connect on LinkedInSagnik Bhattacharya

A short script works better than improvising:

Subscribe on YouTube@sagnikteaches

Before we start, I'd like to record this meeting so I can write accurate minutes. The recording is used only for the minutes, it isn't shared with anyone, and I'll delete it 14 days after we agree the minutes. Is everyone happy with that? Please each say yes or no.

Let each person answer on the recording, then type "recording consent confirmed by all" at the top of the minutes. If someone says no, you have three honest options: do not record, take typed notes yourself and use the AI only to tidy them, or ask whether the objection is to the recording or to the storage and adjust. Never record and hope. Agree a pause rule too, so that "can we pause the recording" is a normal thing to say when a client starts discussing their finances or a dispute with a neighbour.

Pick the capture route by how the meeting happens

The AI step works from a transcript, so the first job is getting a clean one. Five routes cover most practices. Prices are USD list prices as of October 2026; confirm them on the vendor's page for your account.

A phone voice recorder plus any transcription service is a sixth route, and often the simplest for in-person meetings in the studio. The question to ask is where the audio goes when you upload it. Google's help page also says Meet's note-taker can be used for in-person meetings, but it handles one language per meeting, from a list of eight, so a client meeting that switches language halfway through will come back with patchy notes.

A quick sum shows why the meeting cap matters more than the price. A three-person practice on Otter Pro pays 3 x $8.33, about $25 a month billed annually; on Business it is 3 x $19.99, about $60. But a two-hour design review runs past Pro's 90-minute limit, so you either record in two parts or pay for Business. The Otter Business versus Pro comparison goes deeper on that choice, and the Zoom settings tutorial covers the Zoom route.

If the practice lives in Microsoft 365, Teams avoids adding a supplier. Microsoft's documentation says an admin can set a meeting policy that requires participants to give explicit consent to being recorded or transcribed, and that Copilot in a meeting can run "only during the meeting", which relies on speech-to-text data that is not saved afterwards. That mode suits a client who objects to a stored transcript, but you must copy the summary out before the call ends because there is no transcript to return to. The same page warns that your organisation's retention policies may still keep Copilot's prompts and responses, so "no transcript" does not mean "no record at all".

A same-day timetable for a 90-minute client meeting

This is a worked timetable for a design review on a house extension, with illustrative times. Total hands-on effort after the meeting comes to about 52 minutes, and the elapsed time is under an hour and a half.

The read-back at 15:30 is the cheapest quality control in the process. Two minutes of saying "so we agreed X, Y and Z, and I'll do A by Friday" puts the decisions into the transcript in plain, unambiguous sentences, and it gives the clients a chance to correct you in the room instead of by email a week later.

Minutes filled in: decisions, actions and design-change flags

Blank templates are everywhere. What helps is seeing one filled in. This is meeting 07 on the same house extension, Tuesday 3 November, with attendees listed by role. It puts decisions first because that is what people look for later. All content is illustrative.

Under these sit two short lists: "Discussed, not agreed" (the tile colour palette, the garden lighting) and "For information" (the approvals timetable, the next meeting date). Look at the wording of D07-02 and the status of F07-02. A condition and an open question are the two things a careless summary erases.

The prompt that keeps the AI honest

This prompt produces the structure above. It forbids invention, makes the assistant flag anything unclear, and asks for a transcript timestamp against each decision so your check is quick. Paste it into a business-plan assistant (ChatGPT Business or Claude Team, for example, which do not train on your content by default) or into Copilot if your transcript lives in Microsoft 365.

You are drafting minutes for an architecture practice from a meeting transcript.
Use ONLY the transcript and my notes below. Do not add anything that was not said.

Output, in this order:
1. Header: project, meeting number, date, attendees by role, apologies.
2. Decisions: only things explicitly agreed ("agreed", "yes, go with", "let's do").
   Keep conditions ("subject to ...") exactly as stated. Number D[meeting]-01...
3. Actions: who accepted the action and the due date if one was given.
   If nobody accepted it, write "owner to be confirmed". Number A[meeting]-01...
4. Design-change flags: any request that changes the approved layout, size,
   materials, budget, programme or consultant input, including casual ones
   ("can we just...", "while we're at it..."). Say who asked. Do not state a
   fee or programme effect; write "effect to be confirmed".
5. Discussed, not agreed: options mentioned without a decision.
6. For information.
Keep it to two pages. If a number, dimension or date is unclear, write [CHECK]
after it. Give the transcript timestamp for each decision.

TRANSCRIPT:
[paste]
MY NOTES:
[paste]

What comes back is usually close, and wrong in a handful of predictable places. These are the faults to hunt for:

• Conditions dropped. "Agreed subject to checking the clearance" becomes "Agreed". Compare every decision with its timestamp.
• Casual change requests missed or promoted. The assistant files a window move under chatter, or lists it as a decision. Section 4 exists to catch this.
• Actions given to the wrong person. With cross-talk, the assistant guesses who said "I'll do that". Your read-back is what settles it.
• Units and numbers. "Two-four" becomes 24 or 2.4. Every [CHECK] must be cleared before issue.
• Length. Small talk creeps in. Cut anything that is not a decision, action, flag or information.

Three of those faults can turn up in a single minute. This is the stretch of meeting 07 where the island and the window came up:

[32:10] Client A: So the island, could it be two-four?
[32:14] Architect: Two point four works if the clearance to the units is right.
        I'll check it on the plan.
[32:21] Client B: Fine, go with two-four then, as long as it fits.
[32:25] Client B: Oh, and could the kitchen window go on the side wall instead?
        While we're at it.
[32:29] Architect: Let me look at that and come back to you.

The 24 m would probably have been caught on a read-through, because it is absurd. The other two read like ordinary minutes, and a client who later finds "window moved" in an agreed record has every reason to think the change was accepted at no extra cost. The timestamps are what keep this check short: jump to 32:10, listen for twenty seconds, fix three lines.

Design-change flags: catching the casual request

Clients rarely announce a change. They say "could the kitchen window go on the side wall?" halfway through a sentence about worktops. A summariser treats that as chatter or, worse, as a decision. The flag's job is to protect the scope and fee conversation: it records that something was asked, by whom, and that the effect is open.

A second pass catches what the prompt misses. Search the transcript for trigger phrases such as "can we just", "while we're at it", "what if we", "I'd prefer", "could it be", "bigger", "move", "add" and "swap", or ask the assistant to list every sentence containing them with its timestamp. Then decide for each hit whether it is a flag, a decision or noise.

The minutes should never quote a fee effect or accept a change. They record that the request was made and that you will reply. The reply, with any fee or programme consequence, goes in a separate letter or email so the minutes do not become an accidental variation instruction.

A decisions log that outlives the meeting

Minutes answer "what happened on Tuesday". A decisions log answers "when did we agree the flat roof, and has it changed since?" It is one spreadsheet or table that every set of minutes adds to. Each row carries a reference, the date, the decision, a status and, if replaced, what replaced it.

The AI can keep the log current. After you have checked the minutes, give it the previous log and the new decisions with this instruction: add the new decisions, mark any earlier decision that a new one contradicts as superseded and name it, never delete a row, and list anything you were unsure about. Then check the superseded rows yourself, because an incorrect supersession is a quiet way to lose a decision.

The issue sheet: who received what, and when

An issue sheet is a short record kept with each set of minutes. It stops the argument that starts with "I never got those". Keep drawing issues in a separate register, as in the drawing issue register tutorial, and use this sheet for minutes.

The covering email can be three sentences:

Minutes of meeting 07 are attached. Please reply by Friday 6 November with any corrections; otherwise I'll treat them as agreed. Three items need a response from you: the clearance check in D07-02, the window move in F07-01 and the tile shortlist in A07-04.

The comment period is yours to choose. Meeting 07 allows three working days because the revised plan is due the same Friday; five working days is easier to remember as a default. Use whatever matches your appointment. The point is that the minutes state it, the sheet records it, and a person sends it.

Where transcripts mislead you

A transcript sounds authoritative, which is why its mistakes slip through. The common ones in a design meeting are predictable:

• Numbers and units. Dimensions are spoken in fragments ("two-four", "fifteen hundred"), and speech-to-text picks a reading without telling you. The [CHECK] rule and the read-back are your defences.
• Silent agreement. A nod is not in the transcript, and "we'll see" is not a decision. Anything uncertain belongs under "Discussed, not agreed".
• Product and detail names. A window system or timber species can be misheard into something plausible but wrong. Check names against the specification.
• Personal asides. A neighbour's name or a client's money worries should be left out of minutes that go to the consultants.

A realistic failure: a set of minutes recorded "Client approved oak flooring". The client had said "oak might be nice if the budget stretches". It surfaced three weeks later when the quote arrived, at the point where changing the specification cost money and goodwill. The cause was a prompt with no "discussed, not agreed" section and a check that skipped the transcript. The fix is the structure above, plus an audit pass after drafting, run as a second prompt in the same chat:

Check the draft minutes against the transcript. For each decision, quote the
sentence from the transcript that supports it and say whether it was clearly
agreed, agreed with a condition, or only discussed. List any decision you
cannot support with a quote.

On the oak example, the audit would come back with something like: "D06-04 oak flooring: the transcript at 41:12 reads 'oak might be nice if the budget stretches'. Discussed, not agreed. Cannot support as a decision." That single line is worth more than a full re-read of the draft. Anything the assistant cannot quote gets removed or confirmed with the client by email.

Meetings with a contractor are a different animal, with progress, programme and instructions, and the construction meeting minutes tutorial covers that side. When a site visit sits next to a meeting, the notes-and-photos method in the site visit report tutorial keeps the two records separate.

Who can see the recording, and how long you keep it

Decide this before the first recording, not after. In Teams, the recording lands in the organiser's OneDrive and everyone invited to the meeting gets permission to view it, while guests and external attendees see it only if it is shared with them. Meet saves its notes document in the organiser's Drive. Otter stores recordings on its own servers and, as noted above, its privacy policy lists model training on de-identified data. Whichever you use, share the minutes with clients and consultants, not the recording, and set a deletion date.

If you paste a transcript into an AI assistant, do it in an account where training on your content is off by default, such as a business plan, or switch off the model-training setting in a personal account's privacy options. A transcript contains everything said in the room, including things you would never put in minutes. Your practice's AI policy should say which tools are approved for client meetings; the AI policy checklist for architects has clauses you can adapt.

Checking it worked after five meetings

Run the process for five meetings, then look at four numbers: the time between the end of the meeting and issue (the target is the same day), the number of corrections clients ask for per set (two or fewer is a good sign), the number of decisions found missing later (zero), and the number of [CHECK] flags left in an issued set (zero). If corrections are mostly about numbers, tighten the read-back. If they are about missed decisions, review how the room discusses decisions, because the cue words ("agreed", "let's go with") may not be reaching the transcript. If they are about tone or length, adjust the template.

By meeting five you should also have a decisions log that you reach for without thinking. That log, more than any individual set of minutes, is what turns a same-day habit into something that protects the project when memories differ.

Questions architects ask about minuting meetings with AI

How do I minute a meeting where some people are on a call?

Run it in the meeting app, put the people in the room on one laptop or speakerphone placed in the middle, and ask everyone to say their name before their first comment. Test the audio before the first client meeting. Expect a weaker transcript from the room than from the call, and lean harder on the read-back at the end.

How accurate is the transcript with accents and technical terms?

Accuracy drops with cross-talk, background noise, accents and unfamiliar terms, and dimensions and product names are where mistakes cost most. Add a short glossary to the prompt (project name, consultants by role, product and material names) so spellings stay consistent, and check every number and name against your notes or the specification before issuing.

Does recording client meetings affect my insurance or appointment terms?

It may. Ask your insurer, and check your appointment terms and any client confidentiality clauses, before you start recording or putting transcripts into AI tools. Keep the answer you are given on file. If a client contract says nothing about recording, say what you intend to do in the meeting invitation so nobody is surprised.

Can I use the same workflow for meetings with consultants?

Yes, with two changes. The flags section becomes coordination items, such as information needed from another discipline, and actions carry drawing or document references. Keep one decisions log across both meeting types so a decision agreed with a consultant sits beside one agreed with the client, and mark which meeting each came from.

Further reads

• Copyright and Client Data Risks of AI in Architecture Practices — What happens to client data and drawings once they enter an AI tool.
• AI Consultant for Architects: What a First Engagement Costs — Seven consultant days priced deliverable by deliverable for a ten-person practice: audit, policy, one workflow, team session and a 30-day check.
• AI Automation for Architects: Enquiry to Fee Invoice, Step by Step — Six draft-only automation steps from enquiry form to overdue-invoice reminder, with costs, task counts and the architect approving every document.
• AI Tools for Architects 2026: A Toolkit From Brief to Handover — Which AI tool to use at concept, design, tender, site and handover, what each costs on its own pricing page, and three stacks for a five-person practice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Otter pricing and privacy policy pages, Microsoft Learn pages on Teams recording policies, recording storage, intelligent recap and Copilot in Teams meetings, Google Meet help on Take notes for me, Zoom support page on meeting summary. Prices are USD list prices checked in October 2026.