One person did the calculation. Two colleagues reviewed it. The deliverable was approved, signed and sent, and a small error went with it. Nobody was careless. The error survived because human review works the way it has always worked: one tired reader at a time.
The error
By looking right. Take a parking calculation: property types, scenarios, demand over time. One engineer produces it. A colleague reviews it. A second colleague reviews it again. The figure is plausible, the method is standard, the document is well written. It is approved and delivered to the client.
The result is a building with more parking than it will ever use, because a small deviation in one calculation passed one author and two reviewers. None of the three was a weak engineer. Three qualified people looked at the number, and the number was still wrong.
Run the same calculation through an AI engine that works differently, and the story ends earlier. It does not read the derivation and nod along. It fetches the inputs, re-derives the result on its own, compares it against the delivered figure, and flags the deviation on the first pass, while the building is still a drawing.
The mechanism
Start with what did not happen: negligence. The engineer ran the self-check, the reviewers did what reviewers do, and everyone followed procedure. If a quality system only holds when nobody has a bad Thursday, it is not a quality system. It is a bet.
Human peer review has two structural properties that no amount of professionalism removes. It is serial: one reader at a time takes the finished document and follows the author's reasoning through it. A reviewer who follows a plausible derivation tends to confirm it, and re-deriving every figure from scratch would roughly double the cost of the deliverable, so it is almost never done. And it is tired: review sits last in the schedule, closest to the deadline, competing with every other document on the reviewer's desk that week. Attention is a budget, and review is paid from what remains of it.
| Property | Human peer review | Independent machine review |
|---|---|---|
| Method | Reads the finished document, follows the author's reasoning | Re-derives every figure from the source data |
| Order | Serial: one reviewer at a time | Parallel: all calculations at once |
| Attention | A budget that runs down during the week | Constant, from the first page to the last |
| Full re-derivation | Doubles the cost, so it is rarely budgeted | Is the method |
Machine review inverts both properties. It re-derives instead of reading along, checks all of it instead of a sample, and gives the last figure the same attention as the first. It does not know the deadline is tomorrow.
An old idea
None of this is a new insight. W. Edwards Deming published his fourteen points for management in 1982, drawn from decades of quality work in manufacturing. The third point: "Cease dependence on inspection to achieve quality." Inspecting a finished product finds defects too late and misses too many. Quality has to be built into the process that produces the thing.
Construction engineering answered its quality problem with more inspection. Self-check, internal review, external review, the client's review: layer upon layer of qualified people reading the same finished document. Every layer helps, and every layer is another serial, tired reader. The parking calculation passed three of them.
Independent re-derivation is what Deming was actually asking for. The check stops being a person reading the output and becomes a second, independent production of the same result. Two derivations that agree tell you something a nod never can.
The signature
None of this moves the signature. The engineer still reviews, still decides, still signs, and still answers for the work, exactly as before. Who holds the pen is settled and stays settled; we wrote about that in The engineer gets the last word. What moves is the catch-rate: the share of errors found before the client finds them.
Every engineer carries a private list of mistakes that were caught at the last minute and still sting. The parking calculation is what that list looks like when the catch never comes. From now on, that class of error meets a second pair of eyes that never tires and never has a Thursday afternoon. The next demand figure that is quietly wrong stays on paper, caught before it is poured.
Sources
Yesper is the AI civil engineer for construction and infrastructure. AFRY, COWI, NRC Group and other Nordic firms use it to halve the time on a study, rerun calculations in minutes, and catch errors that would otherwise slip through. Get in touch if you'd like to see what it can do for you.
Book demoGet the next piece in your inbox. Unsubscribe anytime.
