Evaluating bids: organising, not deciding

Why this matters

Evaluating bids is slow and detailed. Committee members read long documents, compare them against criteria and must be able to defend their scores. AI can help with the reading and the organising. It must not do the judging. A tender award that rests on a tool's score is open to challenge, and it takes the decision away from the people who are accountable for it.

What AI may do

• Put what each bid says against each published criterion in a table, quoting the exact words.
• Flag where a bid does not answer a criterion, or where its answer is unclear or contradicts itself.
• List the clarification questions a committee might ask each bidder.
• Check that all required documents appear to be present.

What AI must not do

• Give scores, ranks or a recommendation.
• Decide that a bid is non-compliant or reject it.
• Add criteria that were not published, or apply the published ones differently to different bidders. The criteria in the tender document are the only ones that count.

Rules for a fair process

1. Use only an approved tool. Bids and prices are confidential. Do not use a public AI tool.
2. Treat all bidders the same. Use the same prompt, the same criteria and the same steps for every bid.
3. Read the original. The committee scores from the bids themselves. The AI table is an index to find the right passage, not a replacement for reading.
4. Check quotes. A tool can misquote or invent text. Look up every quoted passage in the bid.
5. Declare conflicts first. A member with a conflict of interest stays out, whatever tools are used.
6. Keep the record. Save the prompt, the tool's output and a note of what the committee changed. This is your audit trail.

Watch for subtle bias

A summary can make a well-written bid look stronger than a plainly written one with the same content. A tool may also treat names, languages or company types differently. Where your rules allow it, hide bidder names while the committee compares content. Keep human oversight real: committee members must understand and be able to explain every score.

A prompt to start with

The compare bids against the evaluation criteria prompt sets out these limits for the tool. Use it only with a tool your department has approved for confidential material.

Try it

Invent two short bids for the same simple purchase and three criteria. Use the prompt above with invented text only. Then write down one thing the table got right, one thing it got wrong and one question you would ask a bidder.

Check yourself

1. Why must AI not give scores or a recommendation?
2. Why should the committee read the original bids even when it has an AI table?
3. What should you save in the record?

Answers: (1) Scoring and award are decisions for accountable officers and a tool's score can be challenged and cannot be explained as a person's judgement. (2) The table can misquote, miss or invent. It only helps find the right passage. (3) The prompt, the tool's output and a note of what the committee changed.