How to use AI for tender and proposal writing

AI is most useful in tender writing as a drafting and assembly tool, not an author. The approach that works is to build a structured library of your real evidence (case studies, policies, accreditations, CVs), use an AI model to draft answers to each tender question from that library, then have a human rewrite the win themes, check every claim and sign off the final document. Used this way, AI removes most of the repetitive writing from a bid while the judgement that wins contracts stays with your team. Used carelessly, by pasting questions into a chatbot and submitting whatever comes back, it produces generic answers that evaluators mark down and claims you cannot stand behind.

"AI should write the eighty per cent of your tender that restates what you already know. The twenty per cent that wins it still has to come from you."

Dean Cookson, founder, Operosus

Why are UK SMBs looking at AI for tender writing now?

Three things have converged.

First, the prize is large and under-claimed. The National Audit Office reports that public sector bodies spend around £125 billion each year on common goods and services alone, before you count construction, care, IT and professional services contracts. Yet Tussell's procurement tracker found that only 20% of public sector procurement spend went directly to SMEs in 2024. The money is there. Smaller firms are not winning their share of it, and the most common reason we hear is not capability, it is capacity: nobody has three spare days to write a tender response.

Second, the rules have changed in SMBs' favour. The Procurement Act 2023 came into force on 24 February 2025, with an explicit duty on contracting authorities to consider the barriers facing small businesses, a single central platform for notices, and 30-day payment terms through the supply chain. It is the most SME-friendly procurement regime the UK has had.

Third, the tools have matured and your competitors are already moving. The ONS found that 23% of UK businesses were using some form of AI by late September 2025, up from 9% two years earlier, one of several adoption numbers we keep sourced in our UK small business AI statistics table. Bid writing is one of the places AI earns its keep fastest, because tenders are full of exactly the kind of structured, repetitive, evidence-based writing that language models handle well.

What can AI actually do at each stage of a tender?

Tendering is a pipeline, not a single writing task. AI helps at almost every stage, but the help looks different each time, and so does the human role.

The pattern across the whole table is consistent: AI compresses the mechanical work, humans keep the judgement and the accountability.

Which prompts actually work for tender writing?

Generic prompts produce generic bids. The prompts below work because each one forces the model to use your material and the buyer's criteria, rather than its own training data. (The general craft of writing prompts that hold up under repetition is covered in our guide to business prompt writing.)

1. The question deconstruction prompt. Run this before writing anything:

"Here is a tender question and its scoring criteria: [paste both]. Break this down into every separate thing the evaluator is asking for, as a numbered list. Flag anything that is implied but not stated. Do not draft an answer yet."

2. The evidence-matching prompt. This is where most of the value sits:

"Here is the deconstructed question from before, and here are extracts from our case studies and policies: [paste]. For each requirement, tell me which piece of our evidence answers it and where we have a gap. Do not invent evidence to fill gaps, list them as gaps."

3. The constrained drafting prompt:

"Draft an answer to this question in no more than [X] words. Use only the evidence I have supplied. Structure it as: our approach, how we have done this before, how the buyer benefits. Write in plain British English, first person plural, no superlatives, no claims that are not in the evidence."

4. The red-team prompt. Use a fresh chat so the model has no investment in its own draft:

"You are an evaluator scoring this answer against these criteria: [paste both]. Score it, justify the score, and list the three changes that would most improve it."

5. The consistency prompt, run across the assembled document:

"Here is our full response. List every factual claim, number and named individual, where each appears, and any place where two sections contradict each other."

Two rules sit above all five. Never paste another organisation's confidential information into a consumer AI tool whose terms allow training on your inputs; use a business-grade service with appropriate data terms. And treat every factual statement in an AI draft as unverified until a person has checked it against the source.

Where do humans have to stay in the loop?

This is the part that decides whether AI raises or lowers your scores. The non-negotiables:

• The bid/no-bid decision. A model can summarise the opportunity. It cannot weigh your pipeline, your capacity next quarter or your relationship with the buyer.
• Win themes. The two or three reasons you specifically should win this specific contract have to come from people who know the client and the competition. AI can polish a win theme. It cannot find one.
• Every claim and every number. Language models will confidently assert things that are not true. In a tender, an unsupported or false claim is not just embarrassing, it can be grounds for exclusion.
• Pricing. Keep AI away from your commercial model entirely, beyond formatting and sense-checking.
• Social value and local content. Evaluators increasingly score what you will actually do in their community. Boilerplate is obvious and scores accordingly.
• Final sign-off. A director signs the form of tender. That signature means a human has read what is being promised.

The eighty-twenty split from the top of this guide is the whole method in one sentence. If you remember nothing else, remember the split.

What does this look like in practice? Bidwell as a worked example

We built Bidwell because we kept watching capable UK SMBs decline tenders they could have won, purely on time. The story of how it went from client problem to product is told in full in our case study. The design decisions in it are a useful template whether you use Bidwell, another tool, or a general-purpose model with good discipline.

• The evidence library comes first. Bidwell's core is not the writing model, it is a structured store of the company's real material: case studies, accreditations, policies, team CVs, past answers. Drafting only ever happens against that library. This is the single biggest defence against hallucinated claims, because the model is assembling your facts rather than generating plausible ones.
• Drafting is question by question, against the criteria. Each tender question is answered individually, with the word limit and scoring weight attached, rather than generating one long document and hoping it maps onto the questions.
• Gaps are surfaced, not papered over. If the library has no evidence for a requirement, that shows up as a gap for a human to resolve, by adding real evidence or deciding how to handle the weakness honestly.
• A human reviews before anything leaves the building. The output of the system is a draft for review, never a submission.

The same pattern, structured real-world data in, AI drafting in the middle, human review before anything reaches a customer, is how we build most client systems at Operosus, from veterinary booking flows to facility-hire matching. Tender writing is simply the use case where the payoff per document is largest.

What mistakes should you avoid?

• Submitting the first draft. Evaluators read dozens of responses and AI boilerplate has a recognisable flavour. Score-winning answers are edited by someone who knows the client.
• One mega-prompt for the whole bid. Quality collapses when you ask for 8,000 words in one go. Work question by question.
• Letting the model pad to the word limit. A 500-word limit is a maximum, not a target. Dense, specific answers beat full ones.
• Recycling a previous bid wholesale. AI makes it temptingly easy to reskin last month's response. Buyers notice when their requirements are answered at someone else.
• Ignoring data protection. Check the terms of any tool before client names, staff data or commercially sensitive figures go into it.
• Skipping the compliance matrix. AI can help build it, but a non-compliant bid loses before a word is read, so a person owns it.

Where to start

You do not need new software to begin. This week:

1. Build a minimal evidence library. One folder: your three best case studies, your accreditations, your key policies, short bios for your delivery team. Plain documents are fine.
2. Pick one live or recent tender and run the five prompts above against one question, using a business-grade AI tool.
3. Compare the edited result with what you submitted last time. Judge the gap and the hours saved honestly.
4. Write down your red lines. Which steps a human must own, which information never enters an AI tool. One page, shared with everyone who bids.
5. Then decide whether to systematise. If you bid more than a few times a year, a purpose-built workflow with a managed evidence library beats ad-hoc prompting, because the library compounds with every bid.

If you want to see the systematised version working on your own tenders, that is what Bidwell does, and we are happy to show you it running on a real opportunity you are considering. Our tender writing use case shows the full system with real production numbers. Either way, start with the library and the human red lines. The drafting is the easy part.