The same handful of questions comes up in almost every conversation with an engineering firm and its IT department. Who is liable, does it train on our data, can we trust its citations, will it replace our people. Here are the honest answers, each linked to the fuller post behind it.
Responsibility and jobs
Yesper is the AI civil engineer for construction and infrastructure, and these are the questions engineers and IT departments put to it most often. Each answer is short on purpose, and each links to the post that works through it in full. We add to the list as new questions come up in real conversations.
The same engineer who was liable yesterday. A named person still reviews the deliverable, decides it is right and signs it; bringing AI into its production changes none of that, and the only thing that changes is what the signature rests on: a calculation chain recomputed in full by machine rather than sampled by a tired colleague on a Thursday afternoon. The full argument is in The engineer gets the last word.
No, though it changes the job. The mechanical share of the work compresses sharply, and some routine tasks disappear entirely: research, recalculation, formatting, cross-checking. Judgment, context, liability and the signature do not compress, which is why the engineer with an AI colleague tends to outcompete the one without, as set out in Will AI replace civil engineers?.
Data, trust and errors
No. Your documents are not fuel for a shared model; the system learns from the verdict, not the data, so your project files stay the compass and never become training material. Each customer's connection runs inside their own Microsoft tenant, with their own access scopes, and nothing is shared between customers.
Only if it carries its evidence. Demand four things of every machine-generated review comment: the requirement it rests on (a TDOK number, an AMA Anläggning code, a Eurocode clause with its national annex), the revision of the document it read, a location you can check in under a minute, and a flag instead of a conclusion when the evidence is thin. Reject any comment missing one, exactly as you would reject an unsourced comment from a human reviewer.
When the evidence is missing, ambiguous or contradictory, a well-built system flags the point for a human instead of asserting an answer: uncertainty is stated, never smoothed over. Underneath that, a machine recomputes the whole calculation chain, every figure rather than a sample, so deviations that a tired serial review can miss are surfaced and handed to a person to judge. The engineer stays the last line of defence, now working from a shorter, sharper list.
Fit, scope and value
Yes. Documents live in SharePoint, so that is where Yesper reads them, through a sync service on Microsoft Graph that mirrors your libraries, checks each file by byte count so nothing is read half-complete, and re-syncs every hour to follow the current revision. It reads your Swedish project documents the same way it reads any other, and the connection is authorized by your own IT administrator inside your own tenant.
No, and it does not try to. Yesper works on written deliverables: studies, technical descriptions, specifications and review comments, not CAD. It reads drawings and acts on what they contain, but the drawing itself stays with the engineer and the CAD tools you already use.
Judge it with your own numbers, not a vendor's multiple. Start from the hour anatomy of a real deliverable, price what re-running everything costs today when the brief changes, then add the two terms firms usually forget: the value of the last error that shipped, and the value of one more design iteration at a constant fee. That worksheet, rather than a payback-period promise, is laid out in The ROI of engineering AI.
If you have a question that isn't here, we're glad to answer it, straight and with a source. Get in touch and we'll take it.
Book demoGet the next piece in your inbox. Unsubscribe anytime.
