AI BID WRITING

01

Start with what actually goes wrong on a bid

Most conversations about AI in bidding begin with drafting, which is the wrong end of the problem.

Across eleven organisations we spoke to, the difficulty that came up most — eight of them, unprompted — was not writing at all. It was that the evidence needed to answer the question already existed somewhere and nobody could find it in the time available.

The second most common was that the person writing the answer is frequently not a bid writer: an engineer, an estimator or a health and safety lead, handed a question and a deadline alongside their actual job.

Neither of those is a text-generation problem. Which is why “AI writes your bid” both oversells and misses.

02

Where AI genuinely helps

Reading the pack. A tender pack running to thousands of pages across dozens of documents is a volume problem before it is anything else, and machines are good at volume.

Finding what you already have. Twenty years of submissions become searchable by meaning rather than by whoever remembers writing them.

Checking coverage. Comparing a drafted response against the requirements it is supposed to answer, and listing what has not been addressed. This is the highest-value use and the least discussed.

Making the standard consistent. Where several divisions or several authors contribute, quality tends to vary by author. A consistent check applied to all of it narrows that spread.

03

Where it does not help, and where it actively hurts

It cannot price your work. Rates, quantities and commercial modelling are not this. Any tool implying otherwise should be asked to demonstrate it on a bid you have already priced.

A general assistant does not know your obligations. It can summarise a scope. It cannot tell you which clause you will be liable for, because it has never seen your contract history, your standards or the specification you were actually sent.

It cannot supply evidence you do not have. If the project example does not exist, generating a fluent paragraph about it is the worst possible outcome — a confident answer that cannot survive scrutiny, in a document that is contractually binding.

Generic output scores badly. Evaluators read many submissions. Text that could have been written about any contractor reliably scores as though it was.

04

The questions worth asking any supplier

1. What is it reading? General training data, or our documents, our submissions and the client’s actual specification? This is the whole difference.
2. Can it show its working? Every claim should be traceable to the evidence behind it, or a reviewer cannot defend it.
3. What happens to our documents? Where they are held, who can see them, whether they train a model, and what happens after the bid closes.
4. What does it refuse to do? A supplier who says their product does everything has not thought about the parts that matter.
5. Will the people who actually write the answers use it? The buyer is a bid director. The user is often somebody who did not choose to be a bid writer, and a tool that makes their evening worse quietly stops being used.

05

“Why not just use Copilot or ChatGPT?”

The most common question, asked by five of eleven organisations before anything else, and it deserves a straight answer.

For summarising a document or improving a sentence, a general assistant is fine and you probably already pay for one. The gap opens where the tender does: checking a response against a specific client’s requirements, drawing on your own submission history, and telling you what is missing rather than what is there.

The difference is not the model. It is what the model is reading.

The detailed comparison → · And ChatGPT →

06

What it looks like when it works

Related

Start with one conversation

Bring a live bid, or a tender you have already submitted. Thirty minutes, no preparation needed.