AI analysis of a EUR 68 million rail tender. It took 23 minutes.

On 11 June, Latvijas dzelzceļš published an open procurement for the modernisation of the rail line between Jāņavārti and Ogre — design and construction works, two lots, with estimated values of up to 36 million euros for the first lot and 32 million for the second. The document package runs to 13 files: the nolikums, a draft contract, two EIS procurement-requirement documents (one per lot), five separate clarifications, one amendment, two submission forms, and the site-visit notice.

We downloaded all 13 and pointed our agents at them. 23 minutes and 35 seconds of machine time later, we had a ledger of 53 extracted requirements, a 19-item preparation checklist and 17 selection and award criteria projected from it, and 8 findings from the quality review. It cost about €9.

Everything below is downloadable. You can check our work against the original documents, which are public.

1. What we actually ran

Several Tendergate agents specialised in procurement analysis. In the first pass they extracted a requirement ledger — obligations, deadlines, forms, selection and award criteria, submission instructions — each recorded with the sentence it came from and a clause reference. That took 13 minutes 20 seconds. In the second pass they ran a quality review, reading the package as a procurement specialist would, looking for internal contradictions, missing pieces, and fair-competition problems. That took 10 minutes 15 seconds.

2. The AI analysis worked out which law applies

Several laws govern procurement in Latvia, and two of them matter here. Publisko iepirkumu likums covers ordinary public contracts. Sabiedrisko pakalpojumu sniedzēju iepirkumu likums covers utilities — energy, water, transport. A railway falls under the second one. Best not to confuse the two — and the AI analysis did not: it identified SPSIL and declined to raise findings under PIL.

3. The two contradictions the AI analysis flagged

Both of the significant findings share a shape: they are invisible if you read any single document, and only appear when you hold several of them against each other.

What follows is what our AI found by reading the documents. Every point is checkable in the original text, and a procurement specialist may read it differently.

The amendment that did not fully land. Amendment No. 1 rewrote clause 3.2.3 of the nolikums to let a bidder submit for one or both lots, naming a priority lot. But the published "Final" nolikums in the same package still carries the original clause — bidders may submit for one lot only — and the copy of the application form embedded inside it is still the old one, without the priority-lot field, even though the form attached separately to the package has been updated. The amendment legally supersedes the original. Both texts are sitting in the same folder anyway. A bidder who reads the nolikums and not the amendment could pass up the second lot — 32 million euros of it — or fill in the form bound into the document they were reading.

The scoring window that moves. The additional-experience criterion is worth 10 points. The EIS requirements document says the qualifying projects must have been completed "within the previous 12 years (from 2013 to the submission date)". The nolikums says the same criterion runs from 2014. A bidder whose project manager's reference project finished in 2013 scores 10 points under one document and zero under the other, in a scoring criterion, on a contract this size.

The agents then checked the arithmetic. Twelve years measured back from a 2026 submission is 2014, not 2013 — so the EIS document's own description does not agree with the year it gives. They recommended 2014 as the correct figure and said which document to fix.

The four smaller findings are of the same family: the amendment introduces a clause numbered 8.2.2.3 when the nolikums already has a different 8.2.2.3, so references to it are now ambiguous; a cross-reference points to clause 6.2.2.5 where the cybersecurity requirements actually sit at 6.2.25; the award formula gives 90 points to price and 10 to quality, with the single quality criterion being all-or-nothing — which the agents called legally permissible but weak at differentiating bids on a contract this size; and the technical specification is not published with the package but released only after an e-signed request, which the agents flagged as a transparency question while noting that the buyer justifies it as protection of commercial secrets.

4. The AI also recorded what is good

Two of the eight findings are strengths, and we think that matters more than it sounds.

The agents singled out the selection and exclusion regime as thorough and verifiable — exclusion grounds tied to SPSIL Article 48, sanctions screening, and detailed cybersecurity requirements under the Nacionālās kiberdrošības likums. They also noted that the price formula is clearly defined and that the financial guarantees line up: the 10% performance guarantee matches between the nolikums and the bid annex, and the draft contract's liability cap resolves to the penalty ceiling set in the annex. They noted that the 20-million-euro turnover requirement is proportionate, sitting well below twice the estimated contract value. Stacked with a project manager who must have led a 20-million-euro project and a 300,000-euro bid security, it still narrows the field noticeably.

Its overall verdict was "minor improvements needed", with no critical deficiencies. This is a competently assembled procurement with a consolidation problem, and that is exactly what the analysis concludes.

5. Every entry carries a document reference

The requirements file has a column most people skip past and shouldn't. Every entry leads back to a sentence or a table cell, and carries a clause reference — "Nolikums, 2.5., 8.4.punkts, 5.pielikums" — pointing at where in which document it lives. Roughly two thirds of the quotes are findable by searching a distinctive fragment of the original. Where a requirement comes from a table, the quote is assembled from the cells and no such sentence exists in the source; the values are the buyer's, the sentence is ours.

Any entry can be checked against the original text in seconds. When the extraction is wrong — it can happen, and we strongly recommend reviewing AI work, certainly still in 2026 — you can see that it is wrong and why without re-reading the whole nolikums. There is an example in this very file: entry 18 records the additional-experience criterion as running from 2013 while citing the nolikums, which says 2014. The year came from the EIS document and the citation from the nolikums — the same conflict the quality review flagged separately. An extraction you cannot check back to a clause is a list of assertions, which is not much use to anyone who has to defend a decision later.

6. How long does this kind of work take?

Twenty-three minutes is the part people react to, so it is worth being precise about what it does and does not mean.

It does not mean the work is finished. It means a first pass across 13 documents — 53 requirements with citations, a checklist, the selection and award criteria, and 8 reviewed findings — took 23 minutes and 35 seconds and cost under ten euros. A procurement specialist doing the same first pass would need to read all 13 files, build the requirement list by hand, reconcile five clarifications and an amendment against the base document, and then notice that a scoring window says 2013 in one file and 2014 in another. We put that at around three working days, call it 24 hours of focused effort.

So: read the 13 documents yourself. They are public and linked below. If you find the 2013/2014 discrepancy faster than 23 minutes, we would genuinely like to know — and if you find something our agents missed, we would like to know that even more.

7. Where human involvement is needed

The skepticism about AI in procurement is reasonable, and we share more of it than our position might suggest.

Everything above is what our AI surfaced by reading the documents. AI can be wrong, and interpretation matters — on several of these points a procurement specialist might read the same clause differently and be right to. Treat each finding as an observation to verify against the linked source, not a settled fact. We are not suggesting Latvijas dzelzceļš made mistakes; we are showing what an automated first pass produces on a real, live package, defects and strengths together.

That is also why the citation column exists, and why the checklist is projected from the ledger rather than generated separately. The machine reads faster than any of us and will never get bored on document nine of thirteen. It does not decide anything. A person still reads the finding, opens the clause, and judges whether it is real.

This is what Tendergate does — the same agents we ran here are the ones our customers point at their own tenders, from either side of the table. Buyers run the quality review before publishing, which is where the 2013/2014 problem is cheap to fix. Bidders run the requirement extraction to get the requirement list and checklist without three days of reading. The tender above is open until 15 September — a deadline extended by that same Amendment No. 1.