ChatGPT or Claude in an engineering firm: prevent made-up answers by giving them the project context
An AI assistant makes things up mostly when it doesn't know the project: without the design assumptions, the contract documents and the approved choices, it fills the gaps with plausible generalities. Give it a written sheet for each project (assumptions, decisions with their rationale, open items), and ask it to cite its sources and say when it doesn't know. The AI drafts and checks consistency, the engineer approves, and nothing goes out unreviewed.
A project engineer asks an AI assistant to rewrite a design assumptions memo for an office building. The answer is clean and well structured, and along the way it cites an imposed load and a standard reference that nobody adopted for this project. The assistant didn’t lie in the usual sense: it filled a gap. It knew nothing about the brief, the contract documents or the choices already approved in meetings. This guide explains why this happens, what to give the assistant for each project, and how to organize review and sharing in an engineering firm.
Why an assistant makes things up when it doesn’t know the project
A language model produces the most plausible continuation of a text. When information is missing, it doesn’t stop: it fills in. OpenAI describes this in a September 2025 post on “hallucinations”: a model sometimes confidently generates a wrong answer, and current training and evaluations reward a guessed answer more than an admission of uncertainty. The company adds that ChatGPT still hallucinates, less often with recent models, but the problem hasn’t gone away.
Anthropic says the same in its documentation: even the most advanced models, like Claude, can generate text that is incorrect or inconsistent with the context provided. The techniques it suggests reduce hallucinations without eliminating them entirely, and it recommends always validating critical information.
In an engineering firm, the gaps are many and always the same:
- Design assumptions: loads, exposure classes, seismic zone, adopted geotechnical data. Without them, the assistant picks “typical” values that aren’t the project’s.
- Applicable codes and standards: a general-purpose model can cite a standard reference that doesn’t exist. Guides written for engineering firms flag this; Tensoria, for instance, notes that any reference produced by AI must remain verifiable and be validated by an engineer.
- Choices already approved: the option selected, the solution ruled out and why. An assistant that doesn’t know them will, in good faith, propose what was rejected three meetings ago.
- Contract documents: the client’s brief, the technical specifications, meeting minutes, the building control inspector’s comments. These are what count, not the model’s general knowledge.
The study published in January 2026 by the French construction industry’s skills observatory (Observatoire des métiers du BTP) puts it simply: review and verification become central, because generative AI isn’t free of bias and hallucinations. One small-business owner sums it up: “AI saves us time but sometimes misleads us.”
What to give it for each project
The answer isn’t to paste everything into the conversation. It’s to keep, for each project, a small set of short, up-to-date texts that the assistant reads before working and that any colleague can reread.
The project sheet
One page, no more: client, architect, address, type of structure, project phase, packages handled by the firm, other parties (building control inspector, cost consultant, other engineering firms), key dates. It keeps the assistant from guessing the context from a single document.
The assumptions
A dated list, each assumption with its source: “imposed loads: per brief dated March 12, page 4,” “soil data: G2 preliminary design geotechnical report dated February 3.” The source matters as much as the value: it lets the assistant cite and the engineer check.
Decisions, with their rationale
This is the part most often missing, and the one that prevents the most errors. A decision fits in three lines:
| Item | Example |
|---|---|
| Decision | Solid cast-in-place slab at level 1, no precast slab panels |
| Rationale | Many openings on the services side, requested by the HVAC package |
| Source and date | Design team meeting minutes no. 7, approved by the project lead |
Without the rationale, the assistant (or a new project engineer) will reopen the question. With it, they know the topic is closed and why.
Open items
What hasn’t been settled, who needs to settle it, and what you’re waiting for: a comment from the building control inspector, an answer from the architect, an additional soil investigation. Writing these down keeps the assistant from treating them as settled. You can explicitly ask it to flag any passage that depends on an open item.
Working instructions
A few sentences at the top of every request, based on the techniques Anthropic recommends:
- “Use only the documents provided, not your general knowledge.”
- “If information is missing, write ‘information not available’ instead of suggesting a value.”
- “For each claim, cite the passage in the document that supports it; if you can’t find a citation, remove the claim.”
These instructions don’t make the assistant infallible, but they make its answers checkable: anything uncited stands out.
The ground rule: AI drafts, the engineer approves
The split of roles is simple and should be written down in black and white at the firm:
- The assistant can format a memo from the material provided, summarize meeting minutes, compare two versions of the technical specifications, spot inconsistencies between documents, and prepare a list of questions for the next meeting.
- The assistant doesn’t decide on an assumption, a design value or an applicable code. It replaces neither the design software nor the engineer’s check.
- Nothing goes out without review: no document produced with AI is sent to the client, the architect or the inspector until the person responsible for it has reviewed and approved it.
This is also the position of the CNIL (France’s data protection authority) and France Num in their guidance sheet for small businesses: the final decision stays human, you need to keep a critical eye, and the company is responsible for decisions it makes based on AI. The same sheet advises writing a usage policy for all staff: how to use the tool, how to check results, what precautions to take for confidentiality. The construction observatory’s study reports testimonials along the same lines: the tool helps with summaries, but the professional checks and decides.
Client data confidentiality
A project contains drawings, client data and sometimes sensitive information about a site. Before uploading a document to an assistant, check two things: what your plan says, and what your contract with the client allows.
| Plan | What the vendor says about training |
|---|---|
| ChatGPT Business, Enterprise, Edu, API | OpenAI says it doesn’t use data from these plans to train its models by default, unless the customer explicitly opts in |
| Claude Team, Enterprise, API (commercial products) | Anthropic says it doesn’t use conversations for training unless you send feedback (thumbs up or down) or explicitly consent; an admin can turn off feedback |
| Consumer plans (ChatGPT Free and Plus, Claude Free, Pro, Max) | Use for training depends on the account settings; check account by account |
“No training” doesn’t mean “no retention”: how long the vendor keeps conversations falls under other rules, found in your plan’s terms. The CNIL’s sheet remains the simplest rule of caution: in consumer tools, don’t enter confidential, personal or strategic data, and anonymize when you can. For an engineering firm, that means at least a business plan for project documents, and a written list of what must never be uploaded.
Sharing between project engineers
A project sheet is only worth something if the whole team works from the same one. Yet every conversation with an assistant belongs, by default, to one person. If everyone pastes their own excerpts, two engineers on the same project get answers based on different assumptions.
The vendors offer Projects for this. In Claude, a project brings together documents, instructions and conversations; on Team and Enterprise plans, it can be shared with colleagues, with view or edit access. In ChatGPT, a project can also be shared, on all plans, with “chat” or “edit” access. That’s a good base for a project: the sheet, the assumptions and the decisions are attached once.
These tools stop where tracking begins: attached files don’t keep a history of who changed an assumption and why, the action items from the site meeting live elsewhere, and the assistant doesn’t update the decision list itself. A few habits make up for it:
- one person owns the project sheet and dates every change;
- a decision only goes into the sheet with its source (minutes, email, approval);
- when a project changes hands, the new project engineer reads the sheet before opening a conversation.
Where Tavimi can help
If your team uses Claude (Claude Code, claude.ai or Claude Desktop) or another MCP client, Tavimi is one option for keeping these sheets as a team. Tavimi is the shared working memory for a team and its AI assistants, over MCP. Each project has its own space; you invite the people involved by email, and an invited person sees only that space.
Assumptions and decisions are versioned notes: nothing gets overwritten, and previous versions remain available. Each note category sets its own writing rules and required headings: for example, you can require a decision to state its rationale, and an incomplete note is rejected. Open items and action items become tasks with an owner and a due date. When the assistant is missing a decision, it asks the question on the task instead of making something up, and only a human answers. Tavimi is hosted in France and sends nothing to an AI provider on its own; review by the engineer remains the rule.
Tavimi replaces neither your project management software nor your design tools. To see how it fits an engineering firm, visit the engineering firms page.
FAQ
What people ask us
Can you ask ChatGPT or Claude which standard applies to a structure?
You can ask it to find the references cited in the documents you give it, but not to decide on its own. Models can produce standard references that don't exist; the engineer must check every reference against the original text.
Is client data used to train the model?
On business plans (ChatGPT Business and Enterprise, Claude Team and Enterprise, the APIs), the vendors say they don't train their models on your data by default. On consumer plans, it depends on the account settings. Also check what your contracts allow before uploading a document.
Do you have to paste everything into every conversation?
No. Claude Projects and ChatGPT Projects let you attach files and instructions that are reused from one conversation to the next. What matters is that the project sheet exists, is up to date and is the same for everyone.
How can you tell whether the assistant made something up?
Ask it to cite, for each claim, the passage in the document that supports it, and to remove anything it can't cite. Whatever is left without a citation should be checked first.
Sources
- Réduire les hallucinations (documentation Claude, Anthropic), accessed October 3, 2026
- Why language models hallucinate (OpenAI), accessed October 3, 2026
- Étude sur la perception et l'intégration des outils d'intelligence artificielle dans les entreprises du BTP (Observatoire des métiers du BTP, janvier 2026), accessed October 3, 2026
- Outils IA pour bureau d'études : lequel choisir en 2026 (Tensoria), accessed October 3, 2026
- Fiche 4 : prendre les précautions indispensables à un usage responsable et sécurisé (CNIL et France Num), accessed October 3, 2026
- Is my data used for model training? Commercial products (Anthropic Privacy Center), accessed October 3, 2026
- Enterprise privacy at OpenAI (OpenAI), accessed October 3, 2026
- What are Projects? (Claude Help Center), accessed October 3, 2026
- Projects in ChatGPT (OpenAI Help Center), accessed October 3, 2026