Check a procurement scoring rubric before bids arrive

Example result.

Support and service levels (10 of 60 quality points)
5  Figures for all four dimensions, committed in the contract, with remedies if missed.
4  Figures for all four dimensions, committed in the contract. No remedies.
3  Figures for all four dimensions. No contractual commitment.
2  Figures for some dimensions, not all.
1  Describes support without any figure.
0  Does not address support.

AI critiques the wording here. It doesn’t choose criteria, set weights or score bids.

1. Write criteria as questions a bid can answer

A criterion is good when a bid can succeed or fail at it by what it says. “Excellent support” is a subject. “State your hours of cover, response time, resolution time and availability” is a criterion, because a bid either contains those figures or doesn’t.

Avoid demands that exclude suitable suppliers without improving the assessment, such as experience with your exact organisation size or a named reference.

Decide the price-to-quality split now, and write the arithmetic: if a criterion is worth 10 points and scored out of 5, its contribution is raw score ÷ 5 × 10. Leaving it until bids arrive is how a method ends up looking chosen to suit a result.

Why this works: most scoring disputes start with a criterion two honest evaluators read differently. That’s a wording problem, and wording is what a language model is useful for.

What you should see: a list of criteria phrased as answerable questions, with weights that add up and a written formula.

2. Ask for descriptor scales

Below are procurement criteria for [what you are buying], with weightings. For each criterion, write a descriptor for each score from 0 to [maximum score]: one sentence saying what a written bid must state, evidence or quantify to earn it.

Do not use "excellent", "strong", "good", "adequate" or "poor" as the word that separates levels. Score 0 means the bid does not address the criterion. Do not change my criteria or weightings, and do not add requirements such as references, years of experience or certifications. List any decision I need to make separately.

Criteria: [criteria and weights]

[what you are buying] is the requirement in a sentence; [maximum score] is your method’s top score; [criteria and weights] is your list from step 1.

Why this works: the banned adjectives are the whole instruction. Left alone, a model writes “5 = excellent response, 4 = good response”, which looks like a rubric and does none of its work. Many scales will step up by adding content, but not all: some criteria judge the quality of a method or use a formula, so don’t force every one into the same ladder.

What you should see: a scale per criterion in observable terms, and a short list of open decisions.

3. Stress-test the rubric with three separate questions

Here is the scoring sheet with descriptors. Act as a critic, not an editor. Do not rewrite anything. Answer these three questions separately, quoting the sheet:

1. Which criteria could two reasonable evaluators, reading the same bid, score differently? Quote the exact words that create the ambiguity.
2. Which criteria cannot be assessed from a written bid alone, because they need information a bidder has no way to supply in writing?
3. Which criteria overlap, so that a bid strong in one is automatically strong in another? Name the pairs and what they share.

Do not reassure me and do not comment on criteria that are fine.

• Question 1 finds wording problems. “Criterion 4 is somewhat subjective” isn’t a finding; “criterion 4 says ‘appropriate level of support’ and ‘appropriate’ is never defined” is.
• Question 2 finds criteria that are really about something else. It forces a choice: express the concern as something a bid can evidence, or move it to a separate stage with its own published criteria.
• Question 3 finds double-counting. If two criteria share most of their content, a strong bidder gains twice and the real weighting isn’t what the sheet says.

Why this works: three questions find three different faults. Asking for “feedback” gets a blended answer that misses all three. “Don’t rewrite” keeps the decisions yours.

What you should see: quoted wording for each problem, grouped under the three questions.

4. Dry-run the descriptors against two fictional bids

Write two short paragraphs of made-up bid text for your hardest criterion, one middling and one good, and test whether the descriptors pick the score.

Here is a descriptor scale and two extracts from fictional bids. For each extract, state which score the descriptors require and quote the descriptor wording that decides it. If the descriptors don't clearly decide a score, say which two levels it falls between and why. Do not use judgement beyond the descriptors.

Why this works: “falls between 3 and 4, because…” is the result you want now rather than in the evaluation room. Moving a level costs ten minutes before publication and is impossible afterwards.

What you should see: a score and quoted descriptor for each extract, or a clear note of where the boundary is unclear.

5. Keep the scoring human, and written

When bids arrive, people score them, and every score gets a note that names the bid section and quotes the relevant line. Don’t paste bid content into any tool your procurement rules don’t expressly allow.

For UK contracting authorities under the Procurement Act 2023, each supplier that submitted an assessed tender gets an assessment summary before the award notice. Cabinet Office guidance says authorities “cannot simply repeat the award criterion and state” that “the tender demonstrated x”, because reference must be made to the tender. Your procurement and legal teams own how that applies. Private buyers have no such duty, but the losing bidder will still ask.

Why this works: if evaluators reference the tender as they go, the explanation is largely written by the time anyone asks for it.

What you should see: a score-note format agreed before evaluation starts: criterion, tender section and quote, score, why the next score up wasn’t awarded, evaluator and date.

Check the result

• No two levels are separated only by an adjective.
• Nothing in a descriptor asks for more than the requirement does.
• Weights add up to the intended total, and each criterion has a written formula.
• Every criterion can be assessed from a written bid.

The usual failure: the AI improves the rubric by quietly adding requirements, such as a named reference at score 4 or “five years’ experience” at score 5. It reads as better wording; it’s a new requirement that could exclude suitable suppliers. Spot it by comparing every descriptor line with your original criterion.

The descriptors added [added requirement], which is not in my criterion. Remove it and restore the original requirement. Show the revised scale with changes listed, and list any open choices for the procurement owner instead of deciding them.

[added requirement] is the phrase the model introduced.

Fictional example. A distributor is buying warehouse management software. Quality is 60 points, price 40. The first draft of the quality criteria:

Technical capability          20
Implementation approach       15
Support and service levels    10
Experience in our sector       8
Cultural fit                   7

The stress test found three faults:

• Ambiguity. “Support and service levels” doesn’t say which levels. A bid offering 24/7 phone support and one offering four-hour resolution answer different questions. Fixed by naming four dimensions: hours of cover, response time, resolution time and availability.
• Not assessable. “Cultural fit” can’t be judged from a written bid. What the buyer meant was that the last supplier escalated everything through a sales manager. Rewritten as “the named people who will hold the account, their roles and the escalation route from first contact to resolution”.
• Overlap. Technical capability and implementation approach both scored integration. One was scoped to the product, the other to the plan.

The first support scale had a 4 that read: “All four dimensions with figures, contractual commitment, and a named customer reference.” That reference wasn’t in the criterion. The repair prompt removed it, giving the scale in the example result.

A fictional bid’s section 6.2 offers 08:00–18:00 cover Monday to Friday, two-hour response, next-working-day resolution and 99.5% availability, described as “target levels”. The score note:

Scored 3. Section 6.2 states 08:00–18:00 cover, two-hour response, next-working-day resolution and 99.5% availability: all four dimensions with figures, meeting the descriptor for 3. A higher score was not awarded because section 6.2 calls these “target levels” and the draft contract at Appendix C contains no matching commitment, which the descriptor for 4 requires.

Weak: “Scored 3 — adequate support offering.” It takes four seconds to write and a fortnight to defend. The weighted contribution is 3 ÷ 5 × 10 = 6 points.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Here is the price scoring formula I intend to use: [formula]. Construct three sets of fictional bid prices where this formula produces a surprising or unfair-looking result, such as an outlier, a zero price or two near-identical bids. Show the arithmetic for each. Do not recommend a different formula.

Using this rubric only, create four clearly fictional boundary extracts: unsupported assertion, equivalent wording, partial evidence across two levels, and possible double-counting across criteria. For each, explain what the evaluator needs to decide. Do not change the rubric to force a score.

Sources

• Cabinet Office: Guidance on assessing competitive tenders
• Cabinet Office: Guidance on assessment summaries