AI and tenders in 2026: what it does well, what it gets wrong
Using AI on UK tenders in 2026: reading the pack, extracting criteria, drafting answers, prompts that work, and why chatbots miss live notices.
Updated on September 5, 2026
Verified on 5 September 2026.
In 2026 the question is no longer whether to use a language model when bidding for public contracts, but where it gains you real time and where it loses you marks, or a contract. The dividing line is sharp: AI is excellent at reading, summarising and reshaping documents you give it, and very poor at going out to find the notices published this week.
What companies actually use AI for
Five uses come up again and again.
Reading the tender pack. A tender pack often runs to hundreds of pages across five or six files. A model that handles long PDFs will pull out, in minutes, the exact scope, the division into lots, the term and extension options, the KPIs and liquidated damages, the documents required and the awkward clauses in the conditions of contract.
Summarising for a bid or no bid decision. Reducing a tender notice to a fixed format, buyer, scope in one line, deadline, estimated value, conditions of participation, lets you triage a stack of notices far faster.
Extracting the award criteria. The instructions describe weightings and sub-criteria, sometimes in a table, sometimes buried in a paragraph on page 12. A structured extraction of those criteria, with weightings, becomes the plan for your quality submission.
Drafting the quality submission. The most common use, and the most dangerous. AI is useful to build a skeleton that mirrors the criteria, to rework paragraphs from previous bids and to check that no requirement in the specification is left unanswered. It is bad at inventing the value you add, which comes from the company.
Watching and classifying. Once notices have been collected from a current source, a model classifies them by relevance, assigns CPV codes and spots duplicates very well. The important words are "once collected".
What it does well, what it does badly
| Task | A general assistant | Why |
|---|---|---|
| Summarise a tender pack you upload | Good | The text is in front of it; the task is compression |
| Extract criteria, weightings and dates from the instructions | Good | Structured extraction, to be checked line by line |
| Map an activity to CPV codes | Good | The nomenclature is public and stable |
| Generate the vocabulary buyers use | Very good | No fresh data needed |
| Score and sort notices already collected | Good | Classification over supplied text |
| Rewrite and structure a quality submission | Acceptable, with review | The substance must come from the company |
| Find the notices published this week | Bad | Partial access to the portals, stale results |
| Give an estimated value that was not published | Bad | Pure invention |
| Cite a standard, a policy note or a section of the Act | Risky | Plausible but wrong references |
| Write a submission to be sent as it stands | Bad | Generic answer, scored as such |
The blind spot: finding current notices
This is the least understood limitation. General assistants, ChatGPT and Perplexity included, have only partial access to procurement platforms. Ask one to "find the open tenders for IT support in the West Midlands" and you will mostly get old competitions, often closed.
Three technical reasons. Buyer e-sourcing portals are poorly indexed: their content sits behind forms, sessions and JavaScript. The window of a tender is short, often three to five weeks, while a search index refreshes more slowly. And the reliable sources are APIs: Find a Tender and Contracts Finder both publish free feeds in the OCDS open data standard, and no general assistant queries them.
Hence the rule: the model comes in after discovery, never for discovery.
Three concrete risks
Invented references
A model produces plausible text, not verified text. In a quality submission that means standard numbers that do not exist, accreditations attributed to your company that it does not hold, imaginary project references, or a section of the Procurement Act 2023 cited with the wrong number. A false reference is not only a drafting error: bids are submitted with a certificate that the information given is accurate, and an inaccurate statement is a serious matter that can lead to exclusion. Every factual claim produced by a model must be checked by someone who knows it to be true.
The generic answer
Text generated without company-specific material shows itself in interchangeable phrasing and the absence of operational detail. Quality criteria reward exactly the opposite: named staff, a dated programme, a method applied to this site, measures matched to the requirement in the specification. A fluent but interchangeable answer loses marks precisely where the competition is decided, and evaluators who read forty submissions a week recognise it immediately.
Confidentiality
The pack sometimes contains site plans, operational data or the addresses of sensitive premises, and the bid you are preparing contains your commercial strategy and your prices. Three minimum precautions: check that the provider does not train on the data you send, prefer a business plan with a contractual commitment on that point, and strip out of the documents anything that does not need to be there. Many packs carry confidentiality clauses restricting disclosure to third parties, and an online service is a third party.
Prompts that work
These three aim at the tasks where the model is strong: the data is supplied, or it is stable.
Summarising a notice. Paste the text of the notice, then impose a fixed format.
Here is a tender notice. Return only, as a list:
buyer, scope in one sentence, type (works / goods / services),
procedure, lots, submission deadline, contract term,
award criteria with weightings, documents required.
If a piece of information is absent from the notice, write "not stated".
Add nothing that is not in the text.
The last instruction matters most: it stops the model filling gaps by invention.
CPV mapping. Useful when building a watch profile.
For the activity "{activity}", list the relevant CPV codes from the broadest
to the most specific, with the official label for each and a score from 1 to 5
for how likely a UK public buyer is to use it for this requirement.
Add the neighbouring codes to exclude to avoid noise. Answer in JSON.
Then check the codes in the official CPV browser: a model easily produces a code in the right format that does not exist.
Synonym expansion. Portal search is literal: searching "IT services" will not return a notice titled "managed service provision". Generating the buyer's vocabulary is what a model does best.
You are a UK public buyer. For the activity "{activity}", list 15 to 20
wordings a buyer would use in the title of a notice, procurement jargon
included, then the near terms to exclude. Answer in JSON.
Where buyers stand
There is no UK statute regulating artificial intelligence in procurement, and the Procurement Act 2023 says nothing about it. The reference point for central government is PPN 017, "Improving transparency of AI use in procurement", published on 17 February 2025 (verified 2026-09-05), which addresses how authorities handle AI in the commercial process. Otherwise the existing regimes apply: the duties of equal treatment and transparency under the Act, data protection law, and the confidentiality terms of the pack itself. In Ireland, buyers and suppliers are additionally subject to the EU AI Act, Regulation (EU) 2024/1689, which entered into force on 1 August 2024.
In practice, more tender documents now include a question asking bidders to declare whether and how AI was used in preparing the response, and a few go further and restrict it. Such clauses are hard to police, but they exist, which is reason enough to read the instructions to tenderers before you start drafting. No published case has turned on a bid written with the help of a language model. The sensible position is to treat AI as an internal drafting tool for which the company remains entirely responsible, exactly like a word processor or an external bid consultant.
Scoutee applies that separation: notices are collected from dated, structured sources, and a model only intervenes afterwards, to classify and summarise what was actually published.