AI procurement portal: what it must read, and what it cannot see

What no procurement portal can see

The Act itself permits non-publication. Section 94 lets a contracting authority withhold information where disclosure would be against the interests of national security, or where the information is sensitive commercial information and there is an overriding public interest in withholding it. [3]

There is a compensation, and it is the useful part: section 94(3) requires the authority to publish the fact that it has withheld, and which limb it relied on. So the absence is visible even when the content is not — which means "we have redacted this under s.94" is itself a signal a discovery tool can surface, and almost none do. [3]

This is the sentence that separates an honest coverage claim from a marketing one. No tool can cover what the statute permits not to exist. Ours cannot either. Everything below is written on that basis.

The notices that arrive before the tender does

This is the material every competitor leaves on the table, and it is all statute. Three notice types exist specifically to tell you about a contract before there is a tender to bid on.

Now the part that makes this commercially decisive rather than merely interesting. Section 54(4) sets the minimum tendering periods, and one of its rows exists only because a planned procurement notice was published: where "a qualifying planned procurement notice has been issued", the tendering period can be as short as 10 days. [9]

Read those two together. A supplier who watches only tender notices can face ten days to respond on a procurement that was legally visible up to twelve months earlier. The Act did not shorten your time; it moved your time earlier, and put it in a notice type most tools do not watch.

Ten days, or eighteen months

The same contract gives you either a fortnight or a year and a half of warning, and which one you get depends on notices published long before the tender — not on how fast your alerts are.

An AI tender notification system has eight things to watch, not one

Most alerting products watch one event: a tender notice appears containing your keyword. A UK contract generates notices across roughly three years, and at least six of them carry information you would act on.

An alerting product that watches only the fourth of those is watching the moment your influence is already over and your deadline has already started. The first three are where the work is; the last three are where the next cycle's work comes from.

AI-powered search built for bidding is a matching problem, not a search problem

Search returns documents containing your words. Bidding needs the opposite operation: given a capability, return the notices you could actually win, with a reason attached to each.

The reason matters more than the match. A bid team that cannot see why a notice was surfaced cannot tune the system, cannot trust it, and eventually stops opening the email — which is how most alert subscriptions die. So the useful unit is not a relevance score. It is a demand: a specific thing this buyer has said it wants, quoted from the notice, with the sentence it was found in.

There is a second reason to insist on the evidence sentence, and it is about honesty rather than usability. A system that must produce a quotation for every claim cannot invent claims. That is a constraint in the data layer, not a promise in the marketing — which is the only kind of promise worth anything.

Why this notice matched you

Below is our own matching engine, run over real Find a Tender notices, showing what it found and — on the last two tabs — what it refuses to find.

Three properties are visible in those five tabs, and each one costs us matches on purpose.

And the counting law that follows from all of it: these signals are read from notice descriptions, and a description is a summary. Most tender packs demand things the notice never mentions. Every count derived this way is a floor, and this page will not present one as a total.

Where a machine learning procurement portal actually helps — and where it does not

The honest split: machine learning is good at the fuzzy, forgiving parts of discovery and bad at the parts where being wrong is expensive.

There is a version-control point hiding in there too, and it is the difference between a system you can trust over time and one that quietly drifts. Our matcher carries a classifier version. Change the rules and the version bumps; a bump makes every stored row stale, which obliges a re-run rather than leaving old and new judgements sitting side by side in the same table pretending to be the same thing. A model you retrain without versioning its output gives you a corpus where you cannot tell which rows came from which system.

Where machine learning is genuinely load-bearing in this product is forward prediction — reading a buyer's publication cadence to estimate when a contract will come round again. That is a projection, it is labelled as one everywhere it appears, and a past expected date renders as "overdue, may re-tender at any time" rather than as a forecast. A projection presented as a fact is the failure mode this whole category is prone to.

Machine-readable, or not at all

A procurement portal that scrapes rendered pages is a fragile product. One that reads a documented, licensed, standardised feed is a boring one, and boring is the correct ambition here.

Find a Tender publishes its notice data under the Open Government Licence and offers it in Open Contracting Data Standard JSON, mapped to OCDS 1.1.5, through a documented API. [15] The endpoint is cursor-paginated with a maximum of 100 releases per request and accepts an updated-from and updated-to window. [22] Bulk daily XML archives are also published on data.gov.uk. [23]

One observed detail, because it is the kind of thing that gets discovered in production rather than read in documentation: no numeric rate limit is published. The contract is a 429 response with a Retry-After header, and it fires. We hit it on 15 August 2026 while collecting the notice samples in the tracer above, at roughly twenty-two consecutive 100-release requests. [22]

Nobody publishes how many UK procurement notices there are

This section exists because the number you would expect to find here does not exist, and saying so is more useful than sourcing it badly.

There is no official UK statistic for the number of procurement notices published per year. We looked on gov.uk, on the National Audit Office's site and on parliament.uk, and there is no publication of it. Every "X thousand notices a year" figure in this category that we could trace ran back to a vendor blog, including some in our own estate.

It can be derived — from the Find a Tender OCDS API or the data.gov.uk daily archives — but a derived figure only means something with its method and its as-at date attached, and it goes stale immediately. So rather than publish one here and defend it forever, this page states the absence and points at the two places you could compute it yourself. [22] [23]

The same applies to our own corpus size. It moves daily, the only correct answer is a query against the database, and this route holds no database connection by design — so no corpus figure appears on this page at all. A number hard-coded into a marketing page is a number that will be wrong within a month and will still be there in a year.

Eleven criteria for AI procurement portal software

Criteria as rows, three honest answers as columns: what a free government service gives you, what a typical commercial aggregator gives you, and what we do. Where we lose, the row says we lose.

"Typical paid aggregator" is a summary of the seven named vendors in the next section, each read on their own site on 15 August 2026, and it is a fair-minded generalisation rather than a measurement — the per-vendor detail is where the checkable claims are. Our own column describes what is built today, not a roadmap.

The alternatives, named

Seven products people actually compare us with, each read on the vendor's own site on 15 August 2026. Coverage claims are quoted as claims — "X says Y about X" — because none of them is independently verifiable, including ours.

The pattern across all seven is the point. Six of the seven make a coverage claim that cannot be checked; one names three sources and is thereby demonstrably incomplete and demonstrably honest. Given that section 94 means no tool can be complete, being checkable is the only available virtue — and it is the axis the entire category has decided not to compete on.

Why the UK procurement market is this hard to see

Not because the data is secret. Because the routes to market multiplied faster than anyone counted them, and government says so itself.

The National Audit Office, reporting on efficiency in government procurement of common goods and services in July 2024, put it plainly: "Government does not know how many frameworks exist, but third-party data service providers have identified more than 21,000 from published data." [19]

And the providers do not agree with one another. The same report records Tussell identifying 8,990, BiP Solutions 11,194 and Delta Market Analytics 21,561 — a spread of better than two to one between the highest and the lowest count of the same thing. [19]

The NAO draws the consequence for suppliers explicitly: "This fragmentation prevents the government from acting as a single buyer… resulting in duplication of effort and increasing bidding costs for suppliers." [19] The same report notes that 19,606 organisations used the Crown Commercial Service to buy common goods and services in 2022-23. [19]

This is also, in one sentence, the strongest available argument for why anybody's coverage claim should be read sceptically. If three specialist data providers counting frameworks from published data differ by a factor of two and a half, a vendor asserting that it monitors "every portal" is asserting something nobody in the market can currently verify — themselves included.

Government has said the same thing about itself. The Cabinet Office's 2022 transparency paper listed as a problem: "No single picture of procurement — There are multiple portals where information on tenders and contract awards are published, meaning that to get a full picture of what is happening you need to check each different portal." [20] That was written before the Act and describes why the central digital platform was built; it is not a description of the position today, and we cite it only as the stated motivation.

Discovery ends where qualification begins

The most valuable thing a discovery system produces is a confident no. It is also the thing no product markets, because it sounds like selling less.

Once a notice is in front of you, discovery has done its job and a different set of questions starts: can we meet the conditions of participation, is the incumbent likely to retain it, is the evaluation weighted where we are strong, and is this worth the fortnight it will cost. Those are qualification questions, and this page hands them off rather than half-answering them.

Keeping those apart is not tidiness. A tool that scores "fit" by blending coverage, capability, incumbency and win probability into one number has made all four unauditable at once, and the first time it is wrong you cannot tell which part was wrong.

Sector portals: NHS, and everything that looks like one

"NHS procurement portal" is one of the commonest ways people search for this, and it rests on a misconception worth clearing up: there is no single NHS register.

NHS trusts are contracting authorities under the Procurement Act 2023 like any other, so their above-threshold and new below-threshold notices go to Find a Tender. [2] What people usually mean by "the NHS portal" is one of three different things: a trust's own e-sourcing instance, NHS Supply Chain's framework arrangements, or a regional collaborative's system. Each is a route to market, none is a publication register, and watching one of them is not watching the NHS.

The consequence for discovery is the general one in a specific dress. The statutory notices tell you the procurement exists; the framework operator's system is where the call-off actually happens; and the second is exactly what Find a Tender excludes when the arrangement was established under earlier legislation.

Who this is for, and who it is not for

The second half is the one worth reading, and the one you should ask every vendor in this category to write.

Every item in the right-hand column is a reason not to buy from us, written by us. It is here because a coverage claim you cannot falsify is worth nothing, and the fastest way to make ours falsifiable is to say where it fails first.

How rfp.quest does tender discovery

Two registers, named. Matching that shows its reasons. Forward notices treated as first-class rather than as an afterthought.

We ingest Find a Tender and Contracts Finder under the Open Government Licence, keep the OCID stable across every notice about the same procurement so a contract's history holds together, extract requirement signals with the sentence each was found in, and surface pipeline, planned and market-engagement notices alongside tenders rather than behind them. It is free to start and there is no card. Sign up free, or search live tenders first without an account.

We publish no time-saving multiple, no win-rate uplift and no coverage percentage for our own product. Every benchmark we could find in this category is a vendor measuring itself, and we are a vendor in this category. Where an independent measurement exists we will cite it and say whose it is.