AI tender review means a machine reads the whole tender package instead of sampling it: it extracts every requirement with a citation, checks the documents against each other, and flags the deviations. The estimator keeps the decisions. The machine makes sure no requirement goes unread.
The setup
Not because reviewers are careless, but because the arithmetic does not work. A bid worth tens of millions rests on a tender package: about a hundred files of administrative conditions, technical descriptions, drawings, quantity schedules and appendices with appendices of their own. The meeting where the bid is signed off takes about 45 minutes. Reading every file properly, with the attention it takes to catch the one line that decides the outcome, would take four people a week. The gap is covered by trust.
The consequences sit in details that are typographically small and financially large. In one procurement, every single bidder missed the same line item, worth about 40,000 kronor, buried deep in the document hierarchy, and the whole procurement had to be run again. And roughly half of a project's cost is not the contractor's own work: it arrives as subcontractor quotes in a pile during the final days of the bid window, judged on gut feel. Many of them carry their own reservations, each of which would need to be checked against the tender documents. The largest single block of cost in the calculation gets the least tool-supported decision in the entire bid.
This is not negligence, it is time arithmetic. The reading load has grown past the reading time. A machine changes exactly that equation: it does not care which file in which appendix a requirement hides in, and it does not tire by the eleventh appendix.
What it does
Four things, all of them things a person does in principle but cannot finish in practice inside a bid window.
| Step | What the machine does |
|---|---|
| Total reading | Reads every file in the package instead of sampling the documents that seem to matter most. |
| Requirement extraction with citations | Pulls out each requirement and links it to exactly where, in which document and which revision, it appears. |
| Cross-document consistency | Compares the technical description, the bill of quantities and the drawing against each other and surfaces what does not match. |
| Deviation flags | Marks reservations, contradictions and requirements left unanswered, so a human can judge them. |
The difference from today's catch mechanism is large. Today the tool is manual keyword searches in PDFs: search for penalties, insurance, whatever hurt last time, and read around the hits. Keyword search catches the risks you already know to name. Machine review reads everything, including what you never thought to look for, and that is usually where the expensive line sits. It also reads consistently: which edition of a governing document actually applies can be referenced in ten different places, and the machine keeps track of all ten.
Blind spots
Human attention does not fail at random. It fails in patterns. Estimators learn the same document conventions, read the same files in the same order under the same pressure, at the end of a bid window with several bids running at once. Attention is fresh in the technical description and thin in the eleventh appendix. That is why it is not one bidder who misses a line but all of them who miss the same one. Independent readers are supposed to be the safety net, but they are only independent if their blind spots are.
A machine can be forced not to share blind spots. The same package can be read in several independent passes deliberately built to differ: what one pass glides over, another catches, and the findings are pooled. Where a review team converges on the same habits, the passes are pushed apart on purpose. It is the only reading strategy with no tired eleventh appendix.
Traceability
Only if you can check it. A useful machine review does not guess and does not summarize in passing. Every flag carries the requirement it rests on, the revision of the document it read, and a verifiable location you can look up. On uncertainty, the rule is to flag, not to conclude: the machine points at the line and lets the human decide. What has to be exactly right goes through checks, not through a language model's free phrasing. If you cannot look a finding up, you cannot defend it to the client, and then it is worth nothing.
Liability does not move. The estimator and the person who signs review, judge and sign exactly as before. What changes is what the signature rests on: a reading that actually happened, instead of a sample and trust. The signature gets safer, not looser.
What changes
The 45-minute meeting does not get shorter. It gets better. When the reading is already done, the minutes go to judgment instead of hope: which deviations matter, which subcontractor reservations need to be resolved, whether to bid at all. The team stops hunting and starts deciding. The expensive line gets a chance to be found before it becomes a cost.
That is what AI tender review does: not faster decisions on thinner material, but the same decisions on material that someone, or something, has actually read in full. Yesper is the AI civil engineer for construction and infrastructure, and reviewing a tender package is exactly the kind of work it is built for: large, repetitive, and too important to sample.
That's exactly what Yesper does with a tender: it reads every file, extracts the requirements with their source and flags the discrepancies, while the estimator keeps the decisions. Get in touch if you'd like to see it in a demo.
Book demo
