When AI helps compare supplier proposals, settle which offers meet the required conditions before presenting a final ranking. Keep a missing answer visible. Do not let an attractive weighted total compensate for a requirement that the organization has defined as essential.
The practical decision for a buying team is how to structure the comparison record. Use separate places for requirement evidence, eligibility decisions and scored differences between eligible offers. AI can help assemble that record in an approved environment; the purchasing owner remains responsible for the evaluation method and decision.
This guide proposes a review process for business teams. Its suppliers, requirements and scores are fictional. It is not procurement-law advice, a product recommendation or an independently tested supplier-evaluation system. Formal tender procedures need the applicable professional and legal review.
Research distinguishes compliance from differentiation
In a Journal of Business Logistics article first published on 14 May 2026, Finnegan McKinley, Anne Dohmen and Vincent Castillo examine 123 supplier bid packages from 31 Ohio public-sector IT projects. They used o3, Grok-3-Mini and DeepSeek R1-0528 with a retrieval pipeline, repeating each bid evaluation 15 times per model.
Their analysis distinguishes compliance signals from competitive differentiation and reports greater scoring volatility around differentiation. Human score comparison was available for 102 bids per model, while volatility used all 123. The study concerns alignment with existing human evaluations, not independently verified supplier delivery outcomes.
The authors propose a two-stage approach, using AI for compliance screening and people for competitive assessment. Their limitations matter: the setting was public-sector IT procurement, and redacted cost data limited the analysis to technical evaluation. These results do not establish that a current assistant can safely reject suppliers or select the best commercial offer.
Our recommendation is more conservative: use the distinction to organize evidence and review, then validate the actual workflow before automating any decision. Consistent scores alone do not establish that a requirement was interpreted correctly.
Agree the method before reading an AI ranking
New Zealand Government Procurement’s evaluation-criteria guidance separates pass/fail preconditions, conditions governing proposal submission and qualitative criteria. It advises reserving preconditions for essential requirements and expressing them clearly. That is guidance for its procurement context, not a universal legal rule for private businesses elsewhere.
For a business comparison, the useful design question is which requirements genuinely cannot be traded against other benefits. Specify the evidence needed, the point at which the condition must be met and who can resolve ambiguity. A requirement to have a capability now differs from a permitted commitment to deliver it before service starts.
The same organization’s methodology guidance distinguishes rating scales from evaluation models. Its weighted-attribute method applies agreed weights to criterion scores, and it discusses evaluating price either within or separately from those criteria. Our application is to record those choices before asking an assistant to calculate a total.
Do not ask the model to quietly invent weights from proposal length or persuasive language. If stakeholders disagree about the priorities, record that as a decision for them. Our AI decision-review guide addresses the related problem of confusing a preferred outcome with evidence supporting it.
Keep three outcomes visible in a fictional comparison
Imagine a fictional company selecting a maintenance provider. It has agreed that the provider must support a specified site during an overnight service window from the start of the contract. The team will compare service approach and whole-contract price among offers confirmed to meet that requirement.
The assistant receives three proposal packages. Our proposed record would look like this:
| Offer | Evidence about the required service window | Current review state | Consequence for ranking |
|---|---|---|---|
| Alder | Explicit commitment covering the specified site and hours | Supported, awaiting the owner’s confirmation | Can enter the eligible comparison once confirmed |
| Birch | Explicit statement that overnight coverage is unavailable | Apparent conflict with the requirement | Owner resolves the eligibility decision before ranking |
| Cedar | “Extended support” mentioned without site or hours | Unresolved evidence | Keep pending under the agreed clarification process |
These are invented conditions, not assessments of real suppliers. The labels deliberately distinguish what a document says from what the authorized evaluator decides.
Suppose an AI draft gives Birch an overall score of 91 and Alder 83 because Birch describes an impressive service approach and lower price. Those fictional totals do not answer whether Birch satisfies the required coverage. Displaying it as the preferred supplier would conceal the unresolved eligibility decision.
Cedar presents a different problem. Silence about exact hours does not establish that overnight coverage exists, but it also does not establish that the supplier cannot provide it. The evaluation process determines whether clarification is allowed, what evidence is acceptable and how unresolved information affects the decision. The assistant should preserve that uncertainty rather than decide the rule itself.
Keep missing evidence out of automatic arithmetic
Build the comparison around criterion-level records. For each, retain the requirement version, supplier document and passage, proposed interpretation, reviewer decision and reason. Where an answer refers to an appendix, check that the appendix was actually supplied and inspected.
A blank score needs an explicit meaning. It might mean not assessed, insufficient evidence, not applicable or excluded from this stage. Those meanings are not interchangeable. A numeric zero can be a legitimate assessed result under a defined scale; using zero as a convenient substitute for every missing value changes the comparison.
Do not automatically drop an unresolved criterion and rescale the remaining weights to 100%. That would let two suppliers receive superficially comparable totals under different methods. Follow the approved missing-evidence policy and visibly withhold or qualify a total where necessary.
Keep extraction and judgment separate. A correctly copied sentence can still concern another site, an optional service or a future product. Our source-field review guide covers verifying document values and their context. Here, the additional question is whether that accepted evidence satisfies this evaluation rule.
For arithmetic that is approved to proceed, use a transparent calculation rather than asking for an unexplained overall impression. Retain the raw scores, weights and calculation version. The reviewer should be able to reproduce the total without asking the model to regenerate its reasoning.
Inspect how the product combines AI and human scores
Sourced’s proposal-evaluation documentation, updated 10 October 2026, describes criterion scores, weighted totals and human scores replacing AI scores for criteria the human evaluates. It also says its AI leaves a score empty when data is insufficient. These are vendor descriptions of intended product behavior; we have not tested the service or verified its evaluation quality.
The relevant review question for any tool is what happens when a score changes origin or becomes unavailable. Inspect whether the total is recalculated, which criteria are affected and whether an earlier recommendation remains visible as though it were current. A “human reviewed” badge is too broad if only one of several consequential criteria was checked.
Keep the original AI suggestion and the accepted human evaluation distinguishable where the system supports an appropriate history. Do not describe a new total as independent human judgment merely because one evaluator changed one field.
Challenge the comparison before using it for an award
Prepare permitted fictional or suitably controlled sample packages with expected review outcomes. Include an explicit requirement conflict, an absent appendix, an optional extra described near an included service and a condition met only after a future date. Check that each reaches the intended decision state under the agreed method.
Also change one accepted criterion score and inspect the calculation and recommendation. Introduce a missing score and verify the documented handling. These are proposed acceptance exercises, not tests performed for this article. In consequential sourcing, have qualified evaluators determine the expected outcomes and appropriate scope.
Repeated AI evaluations can reveal instability worth investigating, but agreement across runs cannot replace checking evidence or authority. Investigate whether disagreement comes from document retrieval, ambiguous requirements, the scoring scale or a genuinely judgment-heavy distinction before editing a prompt.
For the next comparison, start with one requirement capable of disqualifying an otherwise attractive offer. Trace it from the approved rule through the source passage to the evaluator’s decision and the displayed ranking. A useful AI comparison makes that path easier to inspect; it does not hide it inside a confident number.