Product pages list features. They rarely answer the question people actually have, which is what using the thing looks like on an ordinary Tuesday.
So this is one tender, followed through ConcursosGov from the moment it appears on Portal BASE to the moment a team decides to bid for it. It passes through the AI reading of the tender documents, the tender score and the decision log. The procurement below is an illustrative example rather than a live one. Every step, field and screen it passes through is the real product.
The company in the example is a software firm in Lisbon. Its commercial side is four people, one of whom handles public procurement alongside her actual job, which is running delivery. That combination is the normal case. It is also the reason most tenders get found late.
09:12 · The notice is published
A contracting authority publishes a procurement for the modernisation of an online services portal. Base value €1,100,000, submission deadline in 24 days, CPV code in division 72, which covers IT services.
At this point the notice is public. Anyone can read it for free. Nothing about that needs solving. The problem is volume. Roughly ninety other notices were published the same week, most of them for road surfacing, cleaning contracts and medical equipment. Nobody at the company has time to work out which of the ninety are theirs.
09:40 · AI reads the tender documents
The company monitors three CPV codes. This notice matches one of them and enters their list. Alongside it sit the procedure programme and the specifications, together running to about eighty pages.
AI reads them as soon as the notice enters the list. The fields that decide a bid come out as structured data rather than as something to go hunting for. The reading is done by Mistral AI models, processed within the European Union. When a value is not explicit in the documents, the field stays empty rather than being guessed.
In this example the accessibility specialist is the problem. The company does not have one on staff. That single requirement sits somewhere past page 40 of the specifications. Companies reading documents in deadline order find it on day 20, when there is no time left to bring in a partner.
Here it surfaces on day one, which turns a dead end into a decision: subcontract, partner or skip.
09:41 · The tender score arrives, at 82
A minute later the tender is scored. Part of the score comes from what the AI has just read, such as the award criteria and their weightings.
The tender score is 82. On its own that number would be useless, which is why it comes apart into the five factors that produced it.
The breakdown is the point, not the number. An 82 with a strong capability match and a warning on contract size is a different proposition from an 82 built on a perfect size fit with a weak capability match. The first says bid, with a consortium in mind. The second says bid alone.
It also means the team can disagree with the tender score on specific grounds. Someone who knows the authority tends to award on price regardless of the stated weightings can overrule a high score and record why.
The tender that scored 25, and why
An article that only shows the product saying yes is not worth much. The same week, a second procurement matched one of the company's CPV codes and came back with a tender score of 25.
It was a supply contract for ruggedised IT equipment. Base value €912,930, so larger than plenty of contracts the company has won. The notice listed an IT services code next to the equipment code, which is common. That is how it matched one of the three codes the company monitors. On the notice alone it looked worth a look.
The score said otherwise. The breakdown explained why in three lines. Capability match was weak: the object is hardware supply, while this company writes software. The AI reading showed price weighted at 85%, which removes any advantage a strong methodology might offer. And the procedure was restricted with prior qualification, asking for years of hardware supply references the company does not have.
This matters more than the 82 did. Nobody needs software to get excited about a contract that obviously fits. What eats a week is the procurement that looks plausible on the notice, gets someone reading eighty pages and turns out to be a hardware tender judged almost entirely on price. A low tender score with visible reasoning ends that in about fifteen seconds.
It also shows what the score is not. It did not say this contract is bad, because for a hardware supplier with the right references it is a perfectly good contract. It said this contract is not yours, which is a different claim and the only one a relevance score is entitled to make.
11:15 · The team decides
Back to the first tender. The person who found it marks it as interested and adds a note: strong fit on methodology weighting, need an accessibility partner, check whether the usual subcontractor is available. Author and timestamp attach automatically.
Her colleague sees the note that afternoon without anyone having to forward anything. He confirms the subcontractor. Eleven days later the tender moves to submitted.
This is the part that sounds least impressive and turns out to matter most. The alternative version of this story involves an email thread, a spreadsheet and a reason for the decision that exists only in somebody's memory. Six months on, a director asks why the company skipped two contracts in the same sector. The recorded version answers in seconds. The remembered version does not answer at all.
The 25 is recorded too, marked not interested with a one-line note about the qualification requirements. That record costs nothing to create. It prevents a colleague opening the same notice in three weeks and starting the analysis again from scratch.
What left the platform
Before the weekly commercial meeting, the list is exported to .xlsx: tender score, estimated effort, deadlines, contracting authority and current decision, filtered to the procurements still open.
Nobody rebuilds that document by hand, which is the real test of whether it gets produced at all in a busy week. The meeting then spends its time on the four tenders that matter rather than on reconstructing which four they were.
What ConcursosGov did not do
It did not write the proposal. It did not submit anything, because submission happens on the electronic platforms licensed by IMPIC. That step is deliberately outside its scope. It did not decide whether to bid. On a tender with a mismatch this specific, it should not.
It did not replace the full reading either. Once the team decided to bid, someone still read the procedure programme and the specifications from start to finish. The AI reading decided where that effort went.
It also cannot see several things that decide real bids. Whether the delivery team has capacity that month. Whether the company has worked with this authority before and how that went. Whether a competitor with an incumbent relationship makes the whole exercise expensive theatre. A tender score built from published data cannot know any of that, which is exactly why the final judgement stays with the people who do.
What it did was compress the distance between a notice being published and a person being able to make an informed decision about it. That distance used to be somewhere between several days and never. Here it was about two hours.
Frequently asked questions
Seven questions come up whenever someone sees the product work on a real tender. The answers are collected here.
How long after a tender is published does it appear in ConcursosGov?
How does ConcursosGov use AI?
What does a tender score of 0 to 100 actually measure?
Does the AI reading replace reading the tender documents?
Can the AI reading get things wrong?
Why record decisions on tenders you are not bidding for?
Can ConcursosGov submit the bid?
