AI-assisted business writing
Where AI-assisted bid writing goes wrong
AI can take much of the drudgery out of bid writing, but it brings risks that a general AI policy doesn’t cover. We look at where model-assisted bids go wrong and how to avoid those problems while keeping the time the tools save you.
Bid writing involves a great deal of repetition. Teams answer similar questions from one tender to the next and rework the same material each time, and many have started using AI models to speed that work up. Given a library of past responses, a model can produce a plausible first draft of a new answer in minutes, which can save days of copying at the start of every bid.
The risks are specific to bids, though, and a general AI policy won’t cover them. Some can’t be seen until an evaluator is reading the submitted document.
The answer that fits every question
Evaluators score each response against the published criteria, which reward evidence that you’ve understood the question and can deliver what’s being asked. A model working from a question and a library of past answers tends to produce something fluent and complete but general. The responses that score well usually name the site, the contract, the constraint, and the thing the buyer is worried about. A model can’t know what the buyer is worried about because that knowledge comes from your conversations with the buyer, and those aren’t in the library.
Use the model for the boilerplate, and keep it away from the material that explains why you’re the right supplier for this contract. Then have the bid manager rewrite the opening of every scored answer with the specifics. Our guide to how tender evaluators read a scoring rubric explains what they look for.
The commitment you didn’t mean to make
What a bid says you’ll do can end up in the contract, and models are generous with commitments. They add phrases such as “24/7 support”, “a dedicated account manager”, and “guaranteed uptime” because they appear in the library, without anyone deciding to offer them for this contract at this price. In a long response, a promise in the third paragraph of the ninth answer is easy to miss at review and expensive to discover once the work has started.
A person needs to read every commitment in a model-assisted bid and decide whether to make it. The most practical way to do that is a commitment sweep before submission: search the document for the words that usually signal a promise, such as “will”, “guarantee”, “ensure”, “dedicated”, and “within”, and check each one against what you’ve priced.
The reference that was never real
Models are prone to inventing details in past-performance sections and case studies. A case study has a familiar shape that’s easy to reproduce, and the model doesn’t check the facts it fills in, so responses can end up citing project values that don’t match the contract or dates that are a year out. Buyers can and will check these details. In EU public procurement, a supplier found guilty of serious misrepresentation in the information used to assess it against the selection criteria can be excluded from the competition (Directive 2014/24/EU, 2014, art. 57(4)(h)).
Write case studies from the project records instead of the library, and trace every figure in them to a document that someone could produce if asked. Our guide to finding hallucinated sources applies the same discipline to references in other documents.
The same bid, submitted twice
When two suppliers use the same model on similar libraries, their responses can start to read alike. Boilerplate that reads like everyone else’s probably won’t cost you much on its own, but if the sections meant to set you apart sound generic too, the evaluator has little reason to prefer you. Where many bidders use AI, the bids that stand out tend to be written in the supplier’s own voice, by people who know the contract.
The confidentiality leak
Bid libraries often hold client names and pricing from earlier contracts. Putting that material into a tool your organization has no agreement with can breach the confidentiality terms of the contracts it came from. Bid teams generally know this, so the question to ask is which tools people actually use late at night before a deadline. An AI policy that names the approved tools is the minimum. For bid work, someone also needs to enforce it.
The document nobody read
The largest risk is also the simplest. A model can produce a two-hundred-page response, and on a tight deadline it’s tempting to submit it after reading only the pricing schedule and the executive summary. Every answer in the document is scored, though, and each is a place where the model may have hedged, overpromised, made something up, or drifted away from the question. A bid hasn’t really been reviewed until someone has read each answer against the question it responds to.
Keeping the time you’ve saved
The tools take real drudgery out of the process, and dividing the work lets you keep that benefit. The model assembles and reformats the boilerplate, while people write the answers that are specific to this contract, decide which commitments to make, check the case studies, and read the whole document before it goes. Teams that work this way can keep much of the time they’ve saved without putting their scores at risk.
If your deadline is close and the full bid hasn’t been read yet, we can do that reading for you. We check each answer against its question and flag unpriced commitments, invented details, and answers that have drifted, so that you can fix them before you submit. It’s a large part of what refining an AI-assisted draft involves when the draft is a bid.
References
- Directive 2014/24/EU of the European Parliament and of the Council of 26 February 2014 on public procurement and repealing Directive 2004/18/EC. (2014). Official Journal of the European Union, L 94, 65–242. https://eur-lex.europa.eu/eli/dir/2014/24/oj
Bid due soon?
Send us the draft and the deadline, and we’ll agree the scope with you before quoting.
Get a quote