You can draft an engineering report with ChatGPT, and for a one-off document you will read closely yourself, it may be fine. For reports that carry findings a client will rely on, four problems get in the way: it can add new content that nobody reviews, it may leak your client's sensitive info, it cannot parse long videos for interesting screenshots, and it does not organize your reports. We build Tenera Reports, an AI tool that drafts engineering reports, so we have an interest here; the argument below stands on things you can check yourself, including a test that takes twenty minutes.
ChatGPT can add new content that nobody reviews
What engineers fear about AI drafting, correctly, is the hallucinated finding. A general-purpose tool is built to produce good text from whatever it has, so filling gaps plausibly is part of its design. Give it a rough set of site notes and it smooths them, connects them, and supplies the connective material a reader expects. Almost all of that reads well. Some of it describes things nobody observed, and the output carries no marker separating the two.
For most writing, gap-filling is the feature that makes the tool useful. In a report carrying findings a client may rely on, it is the exact failure mode: a specification detail, a condition, a cause that was never in anyone's account, stated with the same confidence as everything else. That is what makes review so hard. An engineer checks the specifics they know to be risky, and an invented sentence does not announce itself as one of them; you cannot check a sentence you do not know needs checking. The next problem has nothing to do with the writing at all.
ChatGPT may leak your client's sensitive info
Be concrete about what gets pasted: photographs of a client's building, notes naming a defect, sometimes a sentence about who built it badly. The moment that goes into an engineer's personal ChatGPT account, your client's sensitive information has left the firm's control. It sits in an account the firm does not control, under consumer terms nobody at the firm ever reviewed, with no record that it happened; the firm cannot see it, cannot audit it, cannot delete it, and the account walks out the door with the employee. That is a leak, whether or not anyone ever misuses it: material the firm was trusted with is now somewhere the firm cannot reach. Most engineering agreements carry confidentiality obligations written long before anyone imagined site photos in a chat window, and "we did not know it was happening" is not a sentence a principal wants to say out loud.
That is why blanket bans appear, and why they fail quietly: the usage just becomes invisible. The workable standard is the one you would apply to any vendor. Know where the data is stored, who can access it, and whether it can stay in the region your work requires, in writing, before client material goes in. The remaining two problems are mundane by comparison, and still expensive.
ChatGPT cannot parse long videos for the screenshots that matter
More and more of what comes back from a site visit is video: the walkthrough is filmed, the narration happens over it, and the best evidence for a finding is a screenshot from the footage at the moment the observation was made. A chat tool cannot parse an hour of walkthrough video and hand back the interesting screenshots, let alone place them as captioned figures in a report. You can pause the footage, screenshot it yourself, name the file, and paste it in with a description, but at that point you are doing the assembly and the tool is only writing prose around it. The gap is not the writing; it is everything between the footage and the finished figure.
ChatGPT does not organize your reports
Each report drafted in a chat tool is a conversation, and conversations accumulate: the site notes pasted into one thread, the follow-up in another, the revision in a third, and the deliverable itself living in a Word file somewhere else again. Nothing connects this month's visit to last month's report on the same building, nothing enforces the firm's template from one report to the next, and when a client asks a question in a year, the record of how the report was made is a scroll through chat history. None of this is a defect in ChatGPT; it was built for conversations, not for organizing reports. Reports need the other thing: one place per report where the draft and the evidence that produced it stay together.
The quieter problem: you cannot know what it read
Underneath all four sits a quieter issue: on any given run, you do not know whether the whole input informed the output. Here is a test you can run this afternoon with any long document. Number every page and every line. Feed it in and ask whether the model read the document completely; it will say yes. Then ask it to quote the first and the last line it read. When we first ran this on a numbered forty-page document, both came back from the middle, run after run, however hard we pushed.
Be precise about what that proves, because the sloppy version of the claim is easy to knock down. A model's account of what it read is itself generated text, so the test does not measure attention directly; what it demonstrates is that you cannot trust the model's own report of what it consumed, and nothing else in the output tells you either. There is also no fixed page count where skimming starts: behaviour varies by document, model, and provider, and shifts as products update. Unpredictability is the accurate description, and unpredictable is not a standard an engineer can work to.
What Tenera does about each
Tenera Reports is built for exactly these four problems.
- Nothing enters the draft that the engineer did not say. The report is drafted from the engineer's spoken narration, and the model cannot assert a finding the engineer never called: a defect in the draft is one the engineer named on site. By default it does minor image interpretation, at the level of describing what a photograph shows, and a template instruction can switch even that off. Constraint is the design, not a setting.
- Client data stays under the firm's control. Data is encrypted in transit and at rest, access is limited to you and the people you explicitly share a report with, and if your work requires data to stay in a specific region, talk to us and we will set that up. The account belongs to the firm, not to whoever happened to paste.
- Long videos are parsed for the screenshots that matter. Tenera extracts the relevant frames from site video and ties them to the transcript of what the engineer said while filming, so the figures in the report come from the moment the observation was made.
- Your reports stay organized, and they come back when you need them. Every report your team produces lives in the firm's Tenera account, in the firm's own template, with the evidence that produced it. When someone goes back to the same building next year, last year's report is where it was left, not buried in a chat history or scattered across personal drives. The finished document exports to Word or PDF whenever it needs to travel.
The questions any vendor should answer about data handling are answered in writing on our security page.
Where each option is the better choice
| General-purpose AI (ChatGPT and kin) | Word template and a disciplined author | Tenera | |
|---|---|---|---|
| Best used for | Drafting prose from scratch, rewriting for tone, one-off documents, summarising something you will read anyway | Low volume, one author, a stable deliverable | Repeat reports across a team, where consistency and traceability matter |
| Cost to start | Nothing; it is already on everyone's machine | Nothing beyond writing the template | Setup work: templates and questions written for the firm |
| Behaviour with your input | Unpredictable; may use all of it, may not, and will not tell you | Entirely predictable; it does nothing on its own | Constrained to the engineer's narration by instruction and template |
| Where it falls down | Adds content nobody reviews; no separation of observed from invented | Every report costs the same as the last; consistency depends on the person | Needs the firm to have decided its own standard first |
| When it is the wrong answer | Any document where a reader may rely on a specific factual claim | Any team where several people write the same report type | Five or six reports a year, one author; setup will not pay back |
The middle column deserves respect: a firm producing a handful of reports a year with one experienced author does not have the problem this category solves, and should not be sold as though it does.
What constraint does not fix
A constrained tool cannot invent a finding, which shrinks the failure surface, and it can still mis-transcribe a number, garble a product name, or structure something confusingly. The engineer reads the draft, checks the specifics, corrects what is wrong, and signs it, exactly as before. Banning consumer tools does not substitute for any of this: bans do not remove the blank-page pressure that makes people reach for them, they just make the usage invisible. Senior engineers tell us versions of the same story about junior staff and consumer tools; the useful response is a workflow where the tool cannot write anything the junior did not observe.
Run the test, then ask one question
Run the numbered-lines test on whatever your team uses now: twenty minutes, one document, three attempts. Then ask one question of any drafting tool you evaluate, in writing: can it introduce a finding the engineer never made? The answer, and how quickly it arrives, tells you most of what you need.
Run the comparison on your own document
If your team is drafting reports with a general-purpose tool today, run the twenty-minute test on a real document before deciding anything. Then, if the results bother you, book a demo of Tenera Reports and bring the same document: you will see what constrained drafting looks like on your own material, including the frames pulled from your site video. Or start a free trial and run the comparison yourself.