Site meeting minutes with ChatGPT: method and checks
Site meeting minutes with ChatGPT: why the "8 days to object" rule does not exist by default, and the method.
By Educasium
Install the ChatGPT skill: Site meeting minutes →
A site meeting ends on an ambiguous note: the structural contractor reports being "at 60%," without specifying the basis for that assessment, and offers to intervene "on Friday" on a leak reported on the ground floor, without confirming the exact date or its own commitment. The architect asks ChatGPT to prepare the minutes from notes taken on the fly. The risk is not that ChatGPT writes poorly: it is that it unintentionally turns an approximate statement into an established fact, or a mentioned intervention into a dated commitment.
This article details what site meeting minutes can establish and what they can never establish on the participants' behalf, the reality behind a often-misunderstood objection rule, the method for preparing this document with ChatGPT without losing the nuance of the original notes, and what sets this ChatGPT version apart from the equivalent method already published for Claude.
Summary
- What minutes can establish, and what they can never establish alone
- The "eight days to object" rule does not exist by default
- Preparing the minutes with ChatGPT, step by step
- A table for not confusing statement, proposal and decision
- What sets this method apart from the Claude version
- When this pack is not enough
- Training to make meeting minutes more reliable
- Frequently asked questions
What minutes can establish, and what they can never establish alone
Site meeting minutes record what was said, observed and decided at a meeting on a specific date. They do not confirm the actual state of progress on site — only a visit or a technical inspection can do that — and they never turn a spoken proposal into a contractual commitment. ChatGPT can organize notes into a readable document; it cannot verify a progress percentage reported by a contractor, nor guess who actually agreed to intervene on a given day if that information is missing from the supplied notes.
A well-written sentence never proves a decision was made
The main risk of AI-generated minutes is not a visible factual error, but a writing fluency that makes a guess look like an established fact. "The contractor will attend on Friday" reads like a decision; "the contractor mentions a possible visit on Friday, unconfirmed" conveys the same information with its real uncertainty intact. That seemingly minor difference in wording changes what the document can prove if a later disagreement arises.
The "eight days to object" rule does not exist by default
A common confusion attributes to site meeting minutes an automatic eight-day objection window, after which their content is deemed accepted by any party that failed to object. This rule does not exist by default: where it exists, it comes from a clause the parties themselves inserted, or from an explicit reference to the NF P03-001 standard, which is itself only a voluntary standard.
What the Civil Code actually says about silence
The general principle set out in Article 1120 of the French Civil Code is that silence does not equal acceptance, except in circumstances set out by law, usage, business relations established between the parties, or an express contract clause. Minutes claiming that a point not objected to within eight days is automatically deemed accepted would apply a rule that does not exist absent that clause — a claim that should never slip into a generated document, even if an old minutes template contained it.
| Common claim | Reality |
|---|---|
| "The minutes are binding if no one objects within 8 days" | False absent a contract clause expressly providing for that window |
| "The NF P03-001 standard imposes this window on every site" | False: this standard only produces effect if the contract expressly cites it among its documents |
| "A point missing from later notes is considered resolved" | False: an absent mention proves neither closure nor agreement and must be checked before any closure |
Preparing the minutes with ChatGPT, step by step
Step 1: Gather the raw notes, the date, confirmed attendees and the last approved minutes. Without the previous minutes, ChatGPT cannot know that an already-open point should be carried forward rather than treated as new.
Step 2: Classify every sentence as an observation, a reported statement, a proposal or an explicit decision before drafting. Doing this classification before asking ChatGPT anything prevents a nuance of uncertainty from being lost in the rewording.
Step 3: Request the draft while explicitly preserving the wording of uncertainty. "The contractor reports 60% progress" must stay distinct from a verified measurement; ChatGPT must reproduce that distinction, never smooth it away for a more fluent style.
Step 4: Carry over action IDs from the previous minutes, without renumbering. An already-open action must keep its number from one set of minutes to the next, so it stays traceable through to its actual closure.
Step 5: Review the resulting Word file against the notes, line by line, before circulation. Check names, dates, reported percentages and photo references; confirm that no mention of an objection window not provided for by the contract has slipped into the document.
A table for not confusing statement, proposal and decision
Faced with an ambiguous note, good practice is to reword the raw note while preserving its uncertainty, rather than turning it into a fact.
| Received note | Correct rewording | Information still needed |
|---|---|---|
| Structure reported at 60% | Progress reported by the contractor: 60% | Basis for that assessment |
| Leak reported on the ground floor | Damp reported on the ground floor, cause not identified | Affected trade and corrective action |
| We will attend on Friday | Attendance mentioned for Friday, unconfirmed in writing | Exact date, committed contractor and scope of the intervention |
Fictional example for testing the method: action A07 concerns a leak still open in the previous minutes. The new notes do not mention it. The correct request to ChatGPT is to carry A07 forward as open, without closing it; a repair proposed "soon" never becomes a firm deadline in the resulting document.
What sets this method apart from the Claude version
Educasium also publishes an equivalent method for preparing site meeting minutes with Claude, built around a native skill installed once and for all in the tool. Both methods share the same central risk — turning a guess into an established fact — but differ in installation and in the pack's scope.
The Claude pack relies on a natively enabled skill that remains available from one conversation to the next once installed; the ChatGPT pack presented here relies on a project to which instruction files are attached, to be reused at every new meeting. In both cases, only the site-specific data needs updating at every meeting, never the pack itself — and in both cases, a mention of an eight-day objection window found in a reference file inherited from an old template must never be presented as a general legal rule applicable by default.
When this pack is not enough
Properly drafted minutes do not resolve every situation encountered at a site meeting.
A persistent disagreement between the client and a contractor over an announced progress figure — one asserting it was observed, the other disputing it — is not resolved with better-worded minutes, but with a joint inspection visit. A decision that significantly changes the site's schedule or budget must be formalized through an amendment or a dedicated written exchange, never left as a mere mention in meeting minutes. Minutes prepared with AI assistance, ChatGPT like any other tool, remain in every case a draft: they must be reviewed and approved by the person responsible for the site before circulation, particularly on contractual mentions.
Training to make meeting minutes more reliable
A well-designed pack is not enough on its own: it also needs, meeting after meeting, the skill to tell findings apart from decisions. For an architect practising as a self-employed professional (NAF code 7111Z in France), the FIFPL fund covers part of the cost of Qualiopi-certified training under 2026 criteria set at €300 per day and €900 per year, with e-learning capped at 50% of the daily rate — funding that must be arranged before training starts, not after.
Frequently asked questions
Is a point not objected to within eight days of the minutes automatically deemed accepted?
No, unless a contract clause expressly provides for it. The general principle set out in Article 1120 of the Civil Code is that silence does not equal acceptance, outside cases set out by law, usage, business relations or an express contract clause. Before applying such a rule, it is therefore necessary to explicitly check whether it appears in the contract's CCAP, rather than assuming it out of habit from an earlier site.
Can ChatGPT verify that progress reported by a contractor is real?
No: ChatGPT organizes and rewords the notes supplied, but never verifies a fact on site. A reported progress percentage must always stay attributed to the contractor who reported it, never presented as a measurement verified by a third party. The minutes must therefore carry the wording "according to the contractor" or equivalent for any progress figure, so the document's reader can immediately identify its source.
Can an action missing from the new notes be considered closed?
No: an absent mention in the latest notes proves neither closure nor resolution of an action. It must be carried forward with its original ID until an explicit closure is communicated. That carry-forward must appear in every new set of minutes until actual closure, even when the action is not discussed at the corresponding meeting.
Must the ChatGPT pack be reinstalled for every new site meeting?
No, as long as the ChatGPT project stays active with its instruction files: only the notes specific to each meeting need updating, never the pack itself. Keep the same project for the whole site, however, so action IDs carry over correctly from one set of minutes to the next. A new project should only be created for a different site, never to restart action numbering from zero on a site already under way.
Reliable site meeting minutes, whether prepared with ChatGPT or any other tool, rest on three constant elements: a strict distinction between findings, reported statements and explicit decisions, no assumed objection rule absent a written clause providing for it, and actions carried forward with their original ID until explicitly closed.
Before the meeting, our ChatGPT method for DCE consistency checking helps secure contract documents ahead of the site, and after a one-off visit, our ChatGPT method for defect tracking details how to trace a defect through to resolution.
Training 100% fundable via OPCO/FIFPL. Qualiopi-certified programme. To structure your site meeting minutes with AI as part of our AI for architectural management training, contact Educasium and specify your status (employee, self-employed, business owner) and your goal. This pack is also available on its dedicated ChatGPT skill page.