AI Can Help Write a Bid, But It Cannot Win One: The Evaluator's View

AI bid writing has moved from novelty to normal, and if you produce tenders for health and social care you are almost certainly using it already. The honest position is this: generative AI can help you draft a bid faster and structure your thinking, but it cannot win a UK public-sector care contract on its own. Evaluators can increasingly tell when a submission has been left to a model, and in a market where AI in health and social care tenders is now the default, the real question is no longer can AI write a bid, but what it still cannot do.

At a glance

• Adoption is mainstream: benchmarking reports that 79% of proposal teams use generative AI, up from 68% in 2024, and 62% now use it to generate answers to specific tender questions.
• AI prose gravitates to the middle of the scoring band, because top marks need specific, named, verified and locally true evidence a model does not hold.
• AI garbles UK care specifics: it confuses the Care Quality Commission with Ofsted, mis-states the Care Act 2014 section 42 duty, and invents a "4-part test" for mental capacity.
• Using AI is not banned and is not itself scored: Procurement Policy Note 02/24 says any question about AI assistance is for information only.
• The real danger is inaccuracy: under the Procurement Act 2023, inaccurate or misleading information can lead to exclusion.
• Evaluators can often tell, and "the AI made it up" is not a defence.

Can AI write a bid, and how far does it get you?

AI bid writing is no longer an edge a few teams have; it is the baseline most of them share. Benchmarking for 2026 reports that 79% of proposal teams now use generative AI, up from 68% in 2024, and that 62% use it to generate answers to specific tender questions. Separate APMP-backed research found 89% have at least experimented with it. Across the sector the average win rate is about 45% (UK teams about 47%), with teams spending around 33 hours on a single response.

The strategic point is uncomfortable: if most competitors use the same models, trained on the same text and prompted the same way, the output converges. Using AI for tenders can shave hours off a first draft, but it cannot make you stand out, because what it produces is, by design, the most average version of an answer. Differentiation matters more now, not less, and it comes from the material AI does not have: your verified outcomes, named systems, local knowledge and honest track record.

Why AI tender responses land in the middle of the scoring band

To see why unedited AI tender responses underperform, look at how care tenders are marked. Most quality questions use a rising scale. A middle score means the answer is acceptable and adequately evidenced: it addresses the question. Top scores go to answers that comprehensively meet and, where invited, exceed the requirement, with specific mechanisms, named systems and roles, clear timeframes and measurable outcomes.

That distinction is where generic prose fails. A weak answer asserts commitment: "we take safeguarding very seriously" or "we are committed to person-centred care". A top answer gives the mechanism, the owner, the system, the timeframe and the outcome: who does what, on which system, how often, checked by whom, and what result it produced last year. Left to itself, AI defaults to the assertion, because assertion is what the average of its training text looks like. As a matter of reading the rubric rather than a measured statistic, unedited AI answers tend to cap out around 3 out of 5: competent, compliant and unmemorable. In a competitive care tender, the merely competent answer loses.

What AI gets wrong in a UK care bid specifically

This is where AI in health and social care tenders becomes genuinely risky rather than merely bland. These are not rare slips; they are predictable failure modes.

It mixes up the regulator. In England, adult social care and healthcare are regulated by the Care Quality Commission (CQC); children's homes and children's social care by Ofsted. AI routinely applies the wrong one, citing an "Ofsted rating" for an adult service or the CQC for a children's service. To an evaluator who commissions in that space daily, that single error signals that nobody who knows the sector has read the answer.

It garbles the Care Act 2014 safeguarding duty. The section 42 duty is precise: where a local authority has reasonable cause to suspect that an adult with care and support needs is experiencing, or is at risk of, abuse or neglect, and cannot protect themselves because of those needs, the authority must make (or cause to be made) whatever enquiries it thinks necessary. As a provider you do not lead the statutory enquiry; you secure immediate safety, record and report, refer and share information, and support the enquiry the authority leads. AI frequently reverses this, which tells a commissioner you do not understand your own role.

It invents a "4-part test" for mental capacity. There is no such thing. The Mental Capacity Act 2005 sets out 5 statutory principles and a two-stage test of capacity. The four functional elements (understanding, retaining, using or weighing, and communicating a decision) sit inside that two-stage assessment; they are not a freestanding test. Stated as a "4-part test", it reads as confident and wrong, the worst combination in front of an assessor.

It fabricates your evidence. Ask for a strong answer and a model will invent case studies, KPIs, contract names, ratings, staff numbers and testimonials that read plausibly and belong to nobody.

It has no local authenticity. It does not know the commissioning council, the local demographics, partners and geography, or the named service you are bidding to run, so it fills the gap with generalities that could describe anywhere.

It ignores the specification. It does not mirror the buyer's wording, misses printed sub-questions, and overshoots word or page limits, any of which can be an automatic mark-down or a non-compliance.

It leans on buzzwords. Comprehensive, robust, person-centred, holistic: fine words, all worthless without a mechanism, role, system or timeframe behind them. AI produces them in bulk, and evaluators discount them on sight.

Can evaluators tell if a bid was written by AI?

Increasingly, yes, and rarely with a detection tool, because the tells are in the substance, not the syntax. The giveaways are the failure modes above: fluent prose that never names a system, person or date; answers that could fit any service in any area; the wrong regulator; the confident statutory error; the empty buzzword. When several submissions sound the same, all like the average of the internet, experienced assessors notice.

So the honest answer to can evaluators tell if a bid was written by AI is that they can often tell it was left to AI, which is the version that hurts you. A bid where a human supplied the evidence, the local detail and the statutory precision, and used AI only to draft and structure, does not carry those tells, because the substance underneath is real.

Do you have to declare AI use in a tender?

The UK position is clearer than many providers assume. Procurement Policy Note 02/24, issued by the Cabinet Office in 2024, gives buyers optional example questions asking whether AI assisted a submission. Crucially, the note states these questions must not be scored or taken into account in evaluation; they are for information only. Using AI is not banned, and disclosing it does not cost you marks.

So do you have to declare AI use in a tender? Sometimes a buyer will ask, and you should answer truthfully, but the disclosure itself is not the risk. Read the guidance further and the real message appears: it warns that content created with large language models may include inaccurate or misleading statements, and that buyers may need to carry out additional due diligence, seek clarification or ask for supporting evidence. The buyer is not policing whether you used AI; the buyer is alert to AI-generated inaccuracy, and is being told to check.

The legal risk: hallucination, exclusion and debarment

This is the part providers underrate, because there is now a hard legal analogue for AI-invented content. In June 2025 the High Court, in joined cases heard by the President of the King's Bench Division, dealt with lawyers who cited multiple non-existent, AI-fabricated legal cases. The court held that generative AI tools are not capable of conducting reliable research, and warned that putting fabricated material before a decision-maker can amount to contempt of court or even perverting the course of justice, with referrals to regulators and wasted-costs consequences. That was a legal-profession matter, not procurement, and no care provider has been disqualified on these facts, but the lesson is blunt: "the AI made it up" is not a defence, invented content does get caught, and it gets punished.

Bring that into procurement itself. Under the Procurement Act 2023, in force from 24 February 2025, a supplier that provides inaccurate or misleading information can face discretionary exclusion for improper behaviour, and excluded suppliers can in serious cases be placed on a central debarment list barring them from future contracts (for mandatory grounds, for up to 5 years).

Be fair about how this works. Exclusion for improper behaviour is discretionary, not automatic. Before excluding you, an authority must generally be satisfied that the behaviour gave you an unfair advantage and that no lesser remedy, such as seeking clarification, would do. An innocent slip you correct is unlikely to trigger it; knowingly or recklessly submitting fabricated content that gains an advantage is squarely in scope. And here is the mechanism that makes an unchecked hallucination dangerous: you sign a declaration that your submission is accurate. The model signs nothing. You do.

What still needs a human, and how to use AI for tenders well

None of this makes AI worthless for bidding; it means you have to know where the line sits. The following cannot be delegated to a model, and a named human has to own each one:

• Verified, checkable evidence, and real measurable outcomes rather than plausible-sounding ones.
• Local knowledge: the council, the population, the partners, the geography and the named service.
• Specification discipline: mirroring the buyer's wording and answering every printed requirement to its weighting.
• Statutory and regulatory precision: the right regulator, the Care Act duty stated correctly, the Mental Capacity Act stated correctly.
• Judgement about where the marks actually are, so effort follows the scoring.
• Data protection: never pasting service-user data, pricing or commercially sensitive material into external AI tools.
• Cross-document consistency, so the method statement matches the pricing, the staffing and the mobilisation plan.
• A named human signing off that every statement in the bid is true.

Inside that line, AI earns its place. It is genuinely useful for building an answer plan, drafting a first pass you can improve, summarising a long specification, checking readability and word count, and listing the sub-questions you must cover. The condition is constant: a human supplies and verifies every piece of evidence, every local detail and every statutory point, and signs it off. Treat the model as a fast, confident junior drafter who has never worked in care, never read this specification, and will never carry the consequences. You would check that person's work line by line; check the model's the same way.

FAQ

Can AI write a winning care tender on its own?

No. It drafts a competent, compliant first pass quickly, but on its own it lands mid-band, because top marks need specific, named, verified and locally true evidence it does not hold, and it makes sector-specific errors that cost you credibility. Winning depends on a human supplying the evidence, the local knowledge and the statutory accuracy.

Can evaluators tell if a bid was written by AI?

Often, yes, usually from the substance rather than detection software. Generic prose that never names a system, person or date, the wrong regulator, confident statutory errors and empty buzzwords all read as an answer left to a model, whereas a bid built on real human detail does not.

Do you have to declare AI use in a tender?

Sometimes a buyer will ask, and you should answer honestly, but Procurement Policy Note 02/24 says those questions must not be scored, so the declaration itself carries no penalty. The risk is not that you used AI; it is whether what you submitted is accurate.

Is it against the rules to use AI for tender responses?

No. Using AI is not banned or penalised in evaluation. What is penalised is inaccurate or misleading content, which under the Procurement Act 2023 can, in serious cases, lead to discretionary exclusion for improper behaviour. The tool is fine; unchecked output is the problem.

What is the safest way to use AI for bid writing?

Use it for structure and speed, not for facts. Let it plan an answer, draft a first pass, summarise the specification and check word counts, then have a human verify every piece of evidence, correct every statutory point, add the local detail and confirm it is all true. Never paste service-user data, pricing or commercially sensitive material into an external tool.

The providers who win are not the ones who use AI or the ones who refuse it; they are the ones who know exactly where its usefulness ends and where human evidence, local truth and statutory precision begin. Get that line right and AI makes you faster; get it wrong and it puts your bid, and your organisation, at real and avoidable risk.