We compared 11 procurement AI platforms on 26 features

We scored eleven procurement platforms against the same 26-capability checklist: ten competitors, researched from their own documentation and independent sources, plus our own. Three results were not what we expected. Not one of the ten has a documented second pass that re-checks its own AI's findings. Only one checks tender documents against national procurement law. And eight of the ten publish no price at all.

The most useful thing we found has nothing to do with who wins. It is that "AI for procurement" describes at least four different products, and they get compared to each other as though they were one.

1. How we built this, and what to distrust in it

We fixed the checklist first — 26 concrete capabilities, from requirement extraction to lot handling to bidder screening — then scored every platform on the identical list, so no vendor got graded on the dimensions its own marketing happens to favour. Every profile was then handed to a second research pass whose only instruction was to find what the first one got wrong. That pass earned its keep: it caught a wrong acquisition year for Responsive (RFP360 was 2021, not 2022), a customer count for Loopio recycled from a 2021 press release, and a Mercell price table read off the wrong row of the source PDF.

Two warnings before you read the grid. First, our own row is self-reported. The ten competitors were researched and fact-checked externally; we filled in our own column from the inside, and it did not go through the same adversarial pass. Weigh it accordingly. Second, and more important: this grid measures whether a capability exists, not whether it is any good. Not one vendor in this market — us included — publishes measured precision or recall for requirement extraction or gap detection, and no independent benchmark exists. A tool that flags 200 issues of which 60 are wrong scores the same here as one that flags 40 correct ones.

2. Four products wearing one label

Response management platforms — Responsive, Loopio, AutogenAI, Arphie — are the biggest and best-funded segment, and the one most people mean by "AI for RFPs". They are vendor-side knowledge systems: a governed library of approved answers, plus AI that maps incoming questions and drafts replies. The unit of work is the question, not the requirement. They are very good at volume and consistency, and structurally uninterested in whether a finished bid complies with a tender's mandatory terms.

Bidder lifecycle platforms — Brainial is the clearest example — cover discovery, qualification, requirement extraction and drafting in one place, built around how an EU tender actually arrives: a zip of PDFs from TenderNed or e-Vergabe, not an Excel questionnaire.

Opportunity intelligence — TED and Deltek GovWin IQ — sells knowing what exists. TED is the free, legally mandated EU register, which is why every commercial tool ingests it instead of competing with it. GovWin sells what scraping cannot produce: 150+ human analysts reading agency budgets and board minutes to surface US opportunities before an RFP exists. Both stop at the opportunity.

eProcurement rails — Mercell, JAGGAER — are the systems of record, where a notice and a bid live end-to-end in one audit-trailed workflow. Both are indispensable if you must publish and receive bids compliantly. Neither reads documents.

3. The grid

4. The empty half of the board

Read across the grid and it splits cleanly. Content libraries, drafting, collaboration, multilingual handling and export are covered almost everywhere. Requirement extraction is broadly covered too — pulling requirements out of a solicitation is the easy part. What happens to that list afterwards is where products stop.

Among the ten researched platforms, five capabilities scored a single "yes" or none at all:

Verification is the emptiest column on the board. The closest thing to it in the commercial tier is Responsive's TRACE score, which is a confidence number produced by the same system that wrote the answer — useful, but not a second opinion. Severity classification is nearly as thin, and case-law precedent appears exactly once outside our own product.

The gap is sharpest at JAGGAER, because JAGGAER has documented the problem itself. Its own blog on running Claude in procurement warns that the model "occasionally invents plausible supplier details – including fabricating ISO certifications for real suppliers," and concludes that "any AI-generated supplier list needs cross-checking before it moves anywhere internal." That is an honest and unusually candid thing for a vendor to publish. It also describes a check that its product does not perform.

5. "Compliance" means two different things

This is the single most expensive misreading available in this market, and no vendor disambiguates it. In the response-management segment, a "compliance matrix" means: does my draft address every requirement the customer listed. It does not mean: is this document lawful. Those four platforms score a flat no on checking anything against a statute or a directive — including one that has raised roughly USD 65M and one that took a USD 200M growth investment.

Citations have the same problem. Loopio's and Responsive's answer citations resolve to entries in your own content library. That is provenance for reuse — proof of where your text came from — not proof of what the buyer's tender demanded. Only Brainial, Arphie, JAGGAER, Riigihanked.eu and ourselves anchor findings to text in the source documents.

6. The platforms that receive bids cannot read them

Mercell and JAGGAER are the only researched products that compare multiple bids side by side. Both compare structured questionnaire fields typed into their own forms, which is why both score only partial on gap detection. Nothing in either suite reads a free-form bid document against a tender pack requirement by requirement. The neatest illustration of the split: JAGGAER, the largest source-to-pay suite here, licenses Responsive to write its own RFP responses.

One column separates products built for European public procurement from products built for US corporate sourcing, and it is an unglamorous one: lots. Responsive, Loopio, AutogenAI, Arphie, Brainial and GovWin IQ all score no or unknown on lot handling. Lots are the default structure of an EU tender. A tool that cannot model them was not designed for this market.

7. Nobody publishes a price

Eight of the ten publish nothing. The only firm figures in this entire research set come from the UK's G-Cloud framework, where procurement rules compel disclosure — and they are worth seeing, because third-party estimates for the same products disagree by an order of magnitude.

8. Where we land

Our column is strongest exactly where the market is emptiest, which is not a coincidence — it is the thing we set out to build. Findings are requirement-level, severity-classified, each carrying a verbatim quote from the source document, re-checked by a separate verification agent, and graded against per-jurisdiction rulebooks for Latvia, Estonia, Lithuania and Poland. We publish our price: EUR 6 per credit, no subscription, first analysis free, with a typical 100-200 page procurement costing 5-10 credits.

The comparison that unsettled us most was not with any of the funded incumbents. It was Riigihanked.eu, an Estonian product carrying the full VAKO disputes archive back to 2007 auto-linked to the statute paragraphs each ruling cites, with published pricing from EUR 49 a month. It is operated by a company with one employee. On the capabilities that actually distinguish this segment, it scores higher than platforms that have raised nine figures.

Which is the real finding. The assurance layer of procurement AI — checking a document against the law, quoting the passage that proves a finding, ranking what is fatal against what is cosmetic, and having something verify the machine before a human signs — is not where the money in this market has gone. It is being built by small teams in specific jurisdictions, because it requires knowing one legal regime properly rather than selling to everyone. If you are evaluating tools, the question is not which is best. It is which of the four products you are actually trying to buy, and then whether anyone in that segment checks their own work.