Short answer
Using AI in a UK public sector bid is not prohibited, and the Cabinet Office compares it to using a bid writer. Authorities may ask you to disclose it, and the government’s own example disclosure questions are explicitly not to be scored. In the US we found no federal rule on a bidder’s use of AI, which means the solicitation governs. The leverage is not in writing faster — it is in choosing which tenders to enter and never losing a requirement buried deep in a long specification.
One thing to say before the detail: this is a UK-led picture. Using AI to write tender responses is a question the UK government has actually answered in writing, and the US section below is shorter because no equivalent general rule turned up — not because it was researched less.
What the UK actually permits
The UK public procurement regime was rebuilt recently: gov.uk records that “The Procurement Act 2023 came into force on 24 February 2025”. Separately, and more usefully for a bidder, the Cabinet Office had already published a policy note on exactly this question.
PPN 02/24, Improving Transparency of AI use in Procurement, puts it in a single sentence: “suppliers’ use of AI is not prohibited during the commercial process but steps should be taken to understand the risks associated with the use of AI tools in this context, as would be the case if a bid writer has been used by the bidder”.
It even frames the upside from the buyer’s side: “There are potential benefits to suppliers using AI to develop their bids, enabling them to bid for a greater number of public contracts.” For a small supplier that has avoided public work because the paperwork was disproportionate, that is the relevant sentence in the entire document.
One scope note worth holding on to: the PPN applies to “all Central Government Departments, their Executive Agencies and Non-Departmental Public Bodies”, and other public sector contracting authorities “may wish to apply the approach set out in this PPN”. A council or an NHS trust is not bound by it, so their tender documents may say something different. Read them.
The disclosure question, and how it is scored
The fear that stops bid teams using AI is not legal. It is that ticking “yes, we used AI” marks the bid down. On central government tenders following the PPN, that fear is addressed directly.
The example disclosure questions in Annex B come with an instruction to the buyer: they “should not be scored or taken into account in tender evaluation, and should be used for information only”. The PPN adds a warning in its own words: “It is important that contracting authorities do not discriminate against particular suppliers in the use of these questions or in the interpretation of supplier responses.”
The honest caveat: the same annex says authorities “can however continue to ask and evaluate any further relevant questions about use of AI as part of their award process”, and that whether those are scored must be set out in the procurement documents. So the unscored question is the standard one. A scored AI question is possible and will say so on its face — which is a reason to read the scoring methodology before assuming either way.
Practically: answer the disclosure question plainly. A supplier who discloses AI use and can describe its own review process reads as competent. A supplier who conceals it and is caught by a clarification question has created a different conversation entirely.
The US picture, including what we could not find
Federal opportunities are published centrally and the thresholds that shape them are in the Federal Acquisition Regulation. FAR 2.101 defines the simplified acquisition threshold as “$350,000, except for” a list of exceptions, and the micro-purchase threshold as “$15,000” with its own exceptions for construction and service contract labor standards.
Those two numbers matter more to a small supplier than any drafting tool, because they mark where procedure lightens. An opportunity under the simplified acquisition threshold is a different commercial proposition from one above it, and knowing which side of the line you are on should precede any decision to invest days in a response.
On AI specifically: we looked for a federal rule governing a bidder’s use of AI in preparing a proposal and did not find one. We are reporting that as a gap in our search rather than as a conclusion about federal law — agencies set their own terms, and an individual solicitation can impose requirements that no general regulation contains. The instruction that survives either way is the same one every experienced bidder already follows: the solicitation is the rulebook.
Where AI earns its place
The instinct is to point AI at the blank box. In LYVIA’s experience the return is larger, and the risk lower, three steps earlier.
| Stage | What AI does well | Why it matters |
|---|---|---|
| Qualification | Read the notice against your own capability and past contracts, and score the fit | The cheapest bid is the one you decide not to write, and capacity spent on an unwinnable opportunity is capacity not spent on a winnable one. |
| Requirement extraction | Pull every mandatory requirement, deadline, certificate and word limit out of a long document into a checklist | Compliance failures are decided before evaluation. A missed mandatory item disqualifies a bid that would have scored well. |
| Answer reuse | Find the approved answer you already wrote for a similar question, and adapt it to this specification | Your best material already exists and is already accurate. Rewriting it from scratch introduces error. |
| Consistency check | Compare the whole submission for contradictions in dates, names, figures and commitments | Multi-author bids contradict themselves. Evaluators notice, and it reads as carelessness about delivery. |
Why the drafting is the wrong place to start
Generated prose is fluent, generic, and evaluated against a scoring methodology that rewards specificity. A model that has not seen your delivery record cannot name the contract, the volume, the named client or the measured outcome — and those are the sentences that score.
There is a second, quieter problem. A fluent draft feels finished, which suppresses the questioning that a rough draft invites. The section reads well, nobody challenges the claim inside it, and the claim was invented. That failure mode is not unique to tendering, but tendering is where it is most expensive. The Cabinet Office puts it to buyers in its own example question: AI tools “may also introduce an increased risk of misleading statements via ‘hallucination’”. That is what an evaluator has been primed to look for.
The answer library is the real asset
The durable investment is not a prompt. It is a maintained set of your own approved answers — accreditations, insurance, safeguarding, social value, methodology, case studies with real numbers — each with the date it was last verified and the person who owns it.
With that in place, retrieval does the work that generation was never suited to: the model finds what you have already said and had checked, and adapts it to the question in front of you. Without it, every bid restarts from nothing and every answer is a fresh opportunity to be wrong. This is the same pattern as retrieval over company documents, applied to the one document set where accuracy is contractually enforceable.
The four ways this goes wrong
- An invented reference or credential. The most serious failure available, because it moves from a weak bid to a misrepresentation. Every factual claim must trace to something a human has verified.
- Confidential material in a public tool. Client names, pricing, drawings and staff details pasted into a consumer chat interface. Check what your tool retains before anything sensitive goes near it — the same discipline as in AI and cybersecurity.
- Answering the question the model expected. Public sector questions are oddly worded on purpose and map to a published scoring methodology. A generically excellent answer to a slightly different question scores badly.
- More bids, not better ones. Capacity gained at the drafting stage is easy to spend on marginal opportunities. Spend it on qualification instead.
A workflow that survives scrutiny
- Score the opportunity before anyone writes. Fit against capability, realistic competition, and whether you can evidence every mandatory requirement. Decline in writing, with the reason, so the decision is reviewable.
- Extract requirements mechanically into a compliance matrix. Every mandatory item, its evidence, its owner, its deadline. This is the highest-value automation in the whole process.
- Draft from your library, not from the model’s memory. Retrieve, adapt, then have the owner of that subject read it.
- Verify every number and name against a source, by a person, before submission. Not a review of the writing — a check of the facts.
- Answer the disclosure question honestly, and keep a short internal note of which tools you used and how output was reviewed. If a clarification comes, you answer it in a sentence.
Where this sits in a wider automation program is covered in our AI strategy roadmap, and the document-handling foundation is in AI document management. For everything else a smaller company might automate, start with our guide to AI automation for small business.
Frequently asked questions
Are we allowed to use AI to write a public sector bid?
In the UK, yes, and the Cabinet Office says so directly. Procurement Policy Note 02/24 states that "suppliers' use of AI is not prohibited during the commercial process but steps should be taken to understand the risks associated with the use of AI tools in this context, as would be the case if a bid writer has been used by the bidder". The comparison to a human bid writer is the whole point: it is a normal commercial practice with normal risks attached.
Will disclosing AI use count against us?
Under PPN 02/24 it should not. The guidance says of its example disclosure questions that they "should not be scored or taken into account in tender evaluation, and should be used for information only", and warns authorities that "it is important that contracting authorities do not discriminate against particular suppliers in the use of these questions or in the interpretation of supplier responses". Buyers may still ask separate, scored questions about AI where it is relevant to what they are buying.
Is there a US federal rule on using AI in a proposal?
We could not find one, and we are not going to imply otherwise. We checked the Federal Acquisition Regulation definitions and the federal acquisition site and found no rule governing a bidder's use of AI in preparing a proposal. That is an absence of evidence rather than proof of absence: individual agencies and individual solicitations can and do impose their own terms, so read the solicitation.
What is the fastest way AI loses a bid?
A fabricated reference or a specification the model invented. Public sector evaluation is a documentary exercise: an assertion you cannot evidence is worse than an omission, and a misstatement about past performance is a different order of problem again. PPN 02/24 raises it directly: its example disclosure question warns that AI tools "may also introduce an increased risk of misleading statements via 'hallucination'".
What thresholds decide whether a US opportunity is worth bidding?
The Federal Acquisition Regulation sets the simplified acquisition threshold at $350,000 and the micro-purchase threshold at $15,000, each with exceptions. Below those lines the process is materially lighter, which changes the economics of bidding for a small supplier more than any efficiency gain from drafting faster.
Where does AI genuinely help in tendering?
Before the writing. Deciding which opportunities to pursue, extracting requirements from a long specification so nothing is missed, and reusing what you have already written and had approved. The win rate moves more from bidding on fewer, better-matched opportunities than from producing more text per bid: a scoring methodology rewards fit, and fit is settled before anyone writes.
If tendering is a real channel for you and the bottleneck is capacity rather than capability, the first thing worth building is the answer library, not the drafting. Book a call.
