Using generative AI to write and evaluate tenders, safely

Writing a good tender specification is slow. Evaluating twenty bids of two hundred pages each is slower. Generative AI promises to help with both, and it can, as long as you set clear boundaries.

For buyers: where it helps

Drafting specifications. Give the model your requirements in plain words and ask for a structured draft with scope, deliverables and service levels. It is often faster to edit a decent draft than to write from a blank page.

Market research. Summarise what is available in a market and common technical standards. Verify any specific claim.

Checking for consistency. Ask it to find contradictions between the specification, the evaluation criteria and the draft contract. This catches errors that otherwise become clarification questions or complaints.

Answering bidder questions. Draft replies that you review, then publish to all bidders at the same time.

Preparing evaluation. Summarise each bid against each criterion, with page references, so evaluators begin with an organised view. The evaluators still read the source and assign the scores.

For bidders: where it helps

• Reading a long tender pack and listing mandatory requirements, deadlines and documents.
• Producing a compliance matrix that maps each requirement to the part of your offer that meets it.
• Improving clarity and structure in your written answers.

The model cannot supply experience you do not have, and it should never invent references or certificates.

The rules that matter

Equal treatment and transparency. Public procurement depends on every bidder getting the same information and being judged on the published criteria. If AI is used in evaluation, it must not introduce hidden criteria or bias. Keep records of how it was used.

Confidentiality. Bids contain trade secrets. Never paste them into a public chatbot. Use a service with a contract that guarantees data is not used for training and is stored in a region you accept, or run a model within your own infrastructure.

Accountability. A named human signs the evaluation report and can explain every score. "The AI said so" is not a defence in a review procedure.

Accuracy. Models can state wrong things confidently. Every number, date and citation in an AI-assisted document needs a human check.

A safe workflow

1. Classify the task. Public information, internal, or confidential? Only the first can go to any tool.
2. Use an approved tool. Provide staff with one that meets your data rules, so they are not tempted to use a personal account.
3. Keep the human in the loop. The model drafts, the person decides.
4. Log it. Record which tool, which version and what for.
5. Review periodically. Check samples of AI-assisted work for quality and bias.

Writing good prompts for procurement

• State the role and the audience: "You are helping a municipal buyer prepare a tender for...".
• Provide the source text and ask the model to use only that text.
• Ask for page or section references so you can verify.
• Request a specific structure, such as a table or a list of criteria.
• Ask it to flag assumptions and anything it is unsure about.

What not to do

• Do not let a model score bids on its own.
• Do not use it to rank bidders on subjective impressions.
• Do not accept a draft specification that happens to favour a specific brand. Check technical neutrality.

The bottom line

Generative AI shortens the slowest parts of tendering, but only if confidentiality, equal treatment and human accountability are built into the process from day one. Write the policy first, then roll out the tool.

This is general information, not legal advice. Check the procurement rules of your market and your data protection or legal team's guidance before use.