The AI agent for public procurement: draft the documents, analyse the tenders

Three documents matched, four inconsistencies — and only one blocks.
Blocking: the tender rules announce three weighted criteria (price 50, technical 40, programme 10); the specification names a fourth, environmental performance, which appears nowhere in the weighting. A bidder answering the specification works on a criterion that will not be scored.
Non-blocking, to correct:
· One quantity in the price schedule differs from the specification take-off on one item — 1,240 m² against 1,190.
· The rules require a site visit; the timetable sets no date.
· Two clauses of the conditions refer to each other in a loop.
What I have written, correction by correction: for the fourth criterion, both possible wordings — weight it in the rules, or remove it from the specification — each complete and ready to substitute; the quantity aligned on the take-off, with the item's workings; the timetable line for the visit; and the conditions' loop untangled into a single cross-reference.
What is left to the buyer: settling the pack. It is the document that will ground the bid analysis — its coherence is an act of the buyer, and that is exactly what makes it defensible before an unsuccessful bidder. One review, one choice on the criterion, and the pack goes out. tender-pack-highways-2027_4-inconsistencies.pdf4 inconsistencies · 1 blocking
⛓ Source · rules, specification, conditions and price schedule

It becomes one the day an unsuccessful bidder notices — and it takes ten minutes to settle today.
What the rules say: "a site visit is compulsory; the visit certificate is attached to the bid".
What the timetable does not say: when, with whom, and by what date it can be requested.
What that produces: bidders ring the user department, which arranges visits case by case. On your comparable 2024 tender, five visits took place over three weeks, two of them after the date the department believed was the cut-off.
What I have prepared: three slots taken from the user department's actual availability, all falling at least ten days before the bid deadline, and the paragraph to insert into the rules — place, person to contact, deadline for requesting a visit.
What is left to you: settling them. They commit an officer's time and bear on equal treatment between bidders — which is what means that, once published, they hold for everybody.

An observation beyond this pack: of your eleven tenders last year, four required a compulsory visit and none published a slot. That is a rules template with a line for it and no field — the corrected template is ready too. rules_visits-with-no-slot.pdf11 tenders · 4 visits with no date
✎ Support · 4 tenders in the same state — the dates stay for the department to set

Anomaly detection runs on the procedure itself, not only on the bids.
What it picks up across your 14 live tenders:
· Missing documents — of 41 applications received, 9 are incomplete: 4 without a tax compliance certificate, 3 without an insurance certificate, 2 without signing authority. Each is flagged with the document named and the date beyond which it will be too late to ask for it.
· Inconsistencies between the three documents that must agree — 7 discrepancies: a performance period of 12 weeks in the specification and 10 in the tender rules; a quantity in the price schedule absent from the description; a criterion weighted 40 % in the rules and 30 % in the notice.
· Abnormally low bids — 2 bids lie more than 35 % below the average of the others and more than 30 % below your estimate. I flag them with the detail of the line items that create the gap: here the “site set-up” line at €1,200 against €14,000 on average.
· Equal treatment — an answer given to one bidder was not published to the others, and the tender closes in 9 days.
What anomaly detection produces: every flag carries the document, the page and the two diverging values, ready to be added to the evaluation report. An abnormally low bid is not rejected by a flag: it opens the written request for justification the rules require, and I draft it. Rejection, award and legal characterisation stay with the buyer and the panel — which is precisely what makes the procedure defensible.
The figure I publish about myself: of my first 120 inconsistency flags, 38 were beside the point — 32 %. All of them compared a clause of the specification with a clause of the tender rules that dealt with two different things under the same word “period”: performance period on one side, bid validity period on the other. I now separate the two and compare clause with clause by subject, never by heading: over the next 300 flags, beside-the-point ones fell to 5 %, and all 7 discrepancies above were confirmed by your procurement officer.
⛓ Sourced · 14 tenders, 9 incomplete applications out of 41, 7 confirmed inconsistencies, 32 % beside-the-point down to 5 %

19.4% of the original value, and the fourth variation being prepared would add 6.2.
The detail:
· Variation 1 — 4.1% — additional work under instruction.
· Variation 2 — 8.7% — scope extension agreed in committee.
· Variation 3 — 6.6% — unforeseen ground conditions.
· Variation 4, in preparation — 6.2% estimated.
What I do not say: at what percentage a variation becomes irregular. Those rules depend on the nature of the contract and the procedure used, and I do not recite them.
What I do say, because it is arithmetic: the cumulative figure would go from 19.4% to 25.6%. That is the figure your procurement team will want before preparing the variation, not at the point of signing it.
The note is ready, with all four variations, their grounds and their dates. highways-variations_cumulative.pdf4 variations · 19.4% → 25.6%
⛓ Source · original contract, 3 notified variations, draft variation 4

Two, and one of them is quieter: its variations are small and numerous.
· Cleaning contract — 7 variations, between 1.2% and 3.4% each, cumulative 14.8%. None drew attention; their sum does.
· IT maintenance contract — 2 variations, cumulative 11.1%.
Why the first is the more interesting: each variation was examined on its own, and each looked modest. Nobody has the cumulative figure at the moment of preparing the next one — it takes opening seven files to get it.
What I can do, on your approval: show the cumulative figure and the effect of the draft in hand, when each draft variation is opened. Not an alert, not a block: a number, where it is useful.
And a quarterly summary to procurement: contracts whose cumulative figure passes a threshold you set, with their trajectory. contracts_cumulative-variations.pdf3 contracts · cumulative figures and trajectories
✎ Proposal · cumulative figure shown at opening — no variation blocked

Four things, and two of them show in delivery rather than in the bids.
· Three bids received in 2022, one excluded as incomplete — no technical statement.
· The successful contractor incurred 4 penalties over the life of the contract, all for late response to a reported fault.
· The specification set no numbered response time — the penalties rest on a general formula, and two were disputed.
· Two firms downloaded the pack and did not bid, in 2022 as in 2018.
What I checked rather than concluding anything: what those two firms did elsewhere. One of them has bid on two of your other tenders since 2022, and won one. A download without a bid has many causes, and indifference plainly is not one of them here — the cause still has to be asked for, and the message that asks it is written.
What it gives concretely for the new specification: a numbered response time, failing which the penalties stay disputable. The clause is drafted, with the time set on the shortest of the four penalised incidents. grounds_2022-review.pdf3 bids · 4 penalties · 2 disputed
⛓ Source · 2022 tender, contract, 4 penalties, pack downloads 2018 and 2022

Yes, and it is the cheapest and most profitable question of a tender.
What I have prepared: a short message to firms that downloaded the pack without bidding, drafted, addressed by name, and dated to go out after the close, unconnected to the award. Three questions: was the time too short, did the scope not fit, did a document stop you?
When it goes, and why not before: after the close, never during. Contacting a bidder mid-procedure bears on equal treatment — and the message only has value afterwards. Sending it carries the department's name: one review, and it goes.
What your data already shows, without asking anything: across your eleven tenders in 2025, 34 pack downloads for 19 bids submitted. Nearly one download in two leads nowhere.

A proposal, if you approve it: the message goes automatically eight days after the close, to downloads without a bid only, once. The replies feed an annual note to procurement — by tender, never by firm. downloads-without-bid_11-tenders.pdf34 downloads · 19 bids
✎ Proposal · message after the close — never during the tender

On the “roads 2027” tender, one submission began before the deadline and finished after it, and your register records only one time.
What the platform's trace says: deadline 12:00. Upload started 11:52, finished 12:04, 180 Mo transferred in 12 minutes. The receipt issued to the bidder carries the finish time.
This kind of case had not yet been taken into account in your submissions register: it records one time of receipt, and nothing says which one when a submission straddles the deadline. So I have not settled it on my own — and I do not qualify this submission, neither as received in time nor as late.
Three ways to record it. Record the finish time: it is the one on the platform's receipt, at the cost of a time that depends on file size and the bidder's bandwidth rather than on their action. Record the start time: it matches the moment the bidder launched the upload, at the cost of a time that appears on no receipt and that nobody but me has read. Record both times and the size transferred, with the platform's timestamped trace, and put the question to the procurement department before the bids are opened: the buyer decides on facts, and the decision is taken before it is discovered.
I recommend the third. The general rule belongs to your procurement department, and I do not recite it from memory — what I can do is set out the timestamped facts before opening rather than after, when an unsuccessful bidder asks for them.
You can also give me another rule of your own — for instance: record the finish time, and have me flag every submission whose start precedes the deadline, so the department decides case by case.
Shall I remember it for the next ones? Across your 14 live tenders and the 41 submissions received, 3 finished after the deadline having begun before it, and 2 of the 3 carried more than 100 Mo. If you approve the rule, I will record the next ones that way, and I will give you the count with every tender.
⛓ Source · platform trace, upload 11:52 to 12:04, 180 Mo · 3 straddling submissions out of 41

Written exactly as you just said it, and the trace is filed with the papers.
The rule, in one sentence: a submission straddling the deadline causes both times and the size transferred to be recorded; the platform's timestamped trace is filed with the tender papers; the procurement department is seized before the bids are opened; I qualify neither admissibility nor timing.
Why your condition matters more than it looks: a trace living in a technical register disappears with the tool that produced it. Filed with the papers, it has the same life as the tender — and it is two years later, on a challenge, that it gets looked for.
What it changes on this tender: both times and the 180 Mo are in the register, the trace is filed with the “roads 2027” papers, and the procurement department has the question this morning, before the opening set for 10:00 tomorrow. The 2 other straddling submissions among the live tenders are handled the same way.
What lets you undo it: the rule is written into your submission procedure under its own name, with its date, the procurement head who approved it and the submission that gave rise to it. You remove one line, and the register goes back to a single time.
✎ Rule approved · submission procedure, entry “submission straddling the deadline”, approved on 12/05 by the procurement department

Seven times, and it is the number of bids received that suffered.
What I looked at: your tenders over the last three years, the submission traces kept, and the number of bids finally analysed on each.
What comes out of it: 7 submissions finished after the deadline having begun before it. 4 were set aside, and all 4 carried between 140 and 310 Mo. The 3 tenders concerned went from 3 bids to 2, and one from 2 to 1. Your average is 2.9 bids per tender, which is already low for arbitrating a price.
The cause is not with the bidders: none of your tender regulations states how long submitting a file of that size takes. A bidder launching their upload at 11:52 believes they are in time, and nothing in the papers they read says otherwise.
What I propose, and it is one sentence to add to the tender regulation: state the indicative size of the complete file and recommend starting the upload one hour before the deadline.
What it is worth, on your figures: 4 bids set aside in three years across 3 tenders, one of which was judged on a single bid. One more bid on a works tender of 340,000 € is counted in price points, and your weighted analysis shows it contract by contract. The calculation comes from your own tenders, not from a national statistic.
The decision is yours: the tender regulation is drafted by your department, and I change none of its clauses.
✎ Proposal · 7 straddling submissions over three years, 4 set aside, 3 tenders down from 3 bids to 2 — the regulation belongs to the procurement department