Practical guide
AI procurement recommendation verification checklist
Use this checklist after a procurement recommendation has been drafted and before it is sent to a decision maker. It tests whether the recommendation can be traced to the stated requirements and evidence, whether alternatives were compared on the same basis, and whether unresolved uncertainty is visible. It is a release control for one recommendation, not a substitute for procurement policy,.
Use this checklist after a procurement recommendation has been drafted and before it is sent to a decision maker. It tests whether the recommendation can be traced to the stated requirements and evidence, whether alternatives were compared on the same basis, and whether unresolved uncertainty is visible. It is a release control for one recommendation, not a substitute for procurement policy, legal review, security review, or an accountable decision.
What the checklist is based on
For US federal competitive proposals, FAR 15.305 requires evaluation against the factors and subfactors in the solicitation and documentation of strengths, deficiencies, significant weaknesses, and risks. The UK government's Sourcing Playbook says testing and piloting should be considered early and reflected in the delivery model, sourcing strategy, bid documents, and evaluation. NIST's AI Risk Management Framework adds that risk work should be continuous across the lifecycle and connect technical questions with organizational responsibilities. These are useful control principles; the rules that apply to a specific procurement come from its jurisdiction, policy, solicitation, and approvals. FAR 15.305 Sourcing Playbook NIST AI RMF Core
Sources: [1] [2] [3]
Prepare the review packet
Freeze a review packet before checking the recommendation. Give each item an evidence ID:
Mark each check pass, fail, not applicable, or not evidenced. A pass needs an evidence ID and, where relevant, a reproducible calculation or test. Record the owner, due date, and decision consequence for every failure. Do not average a critical failure into a composite score.
- the decision question, permitted options, owner, deadline, budget boundary, and required approvals;
- the issued requirements and evaluation factors, including their weights or priority;
- vendor submissions, clarifications, demonstrations, references, and test results;
- the comparison workbook and all transformations or calculations;
- risk, privacy, security, accessibility, legal, data, and operating reviews that are required for this decision; and
- the exact recommendation version being reviewed.
TRACE release checklist
### T — Tie the recommendation to the decision
Hold condition: the recommendation answers a different question from the approved procurement decision, or introduces an undeclared criterion that changes the ranking.
### R — Reconcile claims to original evidence
Hold condition: a material conclusion depends on a source that cannot be located, a quotation that changes meaning outside its context, or an unsupported inference presented as verified fact.
### A — Apply one comparison basis
Hold condition: correcting a formula, comparison boundary, or unsupported override could change the preferred option.
### C — Challenge operating consequences
Hold condition: a critical control or operating dependency has no accountable reviewer, or the proposed pilot cannot observe the failure that matters to the decision.
### E — Explain the disposition and residual uncertainty
Hold condition: the narrative is more certain than the evidence, or an unresolved issue could reverse the decision without a condition, test, or escalation path.
- [ ] The decision, accountable decision maker, deadline, budget boundary, permitted alternatives, and required approvals are explicit.
- [ ] Every criterion in the comparison maps to an issued requirement or an approved decision need.
- [ ] Mandatory requirements are separated from scored preferences and optional benefits.
- [ ] All vendors were assessed against the same current criteria, evidence window, and interpretation rules.
- [ ] The recommendation states what is out of scope, including later implementation choices.
- [ ] Each material vendor claim links to the original submission, clarification, demonstration record, reference, or test result.
- [ ] Vendor assertions are labelled as assertions unless independently tested or corroborated.
- [ ] AI-extracted facts have been checked against the original source, including units, dates, qualifications, and exclusions.
- [ ] Missing evidence is recorded as missing rather than inferred from a similar feature or a different product tier.
- [ ] Contradictions between documents, demonstrations, references, and clarifications are logged and resolved or left visibly open.
- [ ] Price, term, volume, currency, taxes, implementation, support, usage, exit, and contingency costs use the same comparison period and boundary.
- [ ] Scores follow the published scale and weighting; formulas, overrides, and rounding can be reproduced.
- [ ] Mandatory failures are not offset by attractive optional features unless the decision rules explicitly permit that treatment.
- [ ] Alternative scenarios use coherent assumptions rather than changing whichever inputs favor one option.
- [ ] A reviewer has recomputed a sample of scores and the decision-driving totals independently.
- [ ] The recommendation identifies dependencies on data, integrations, people, process changes, controls, service levels, and supplier capacity.
- [ ] Proposed testing covers representative normal, boundary, failure, and recovery cases.
- [ ] Privacy, security, legal, accessibility, records, and sector reviews required by policy have named owners and dispositions.
- [ ] Lock-in, portability, transition, termination, monitoring, and material-change triggers are addressed.
- [ ] Benefits are separated into evidenced, estimated, and unmeasured claims, with owners for later validation.
- [ ] The recommendation explains why the preferred option is stronger for the stated decision, not merely why it has the highest total.
- [ ] Strengths, deficiencies, significant weaknesses, risks, open questions, and dissent are visible.
- [ ] The conclusion remains the same under plausible sensitivity cases, or a decision trigger covers the reversal risk.
- [ ] Conditions precedent, negotiation points, pilot exit criteria, monitoring measures, and reapproval triggers are explicit.
- [ ] Preparer, independent reviewer, approvers, evidence snapshot, version, review date, and refresh date are recorded.
Verification record
| Area | Result | Evidence IDs | Failure or uncertainty | Owner | Due date | Decision consequence | | --- | --- | --- | --- | --- | --- | --- | | decision alignment | | | | | | | | claim traceability | | | | | | | | comparison integrity | | | | | | | | operating consequences | | | | | | | | disposition and release | | | | | | |
Choose one overall disposition: release for the stated decision, release with visible conditions, return for repair, or hold the dependent decision. The disposition applies only to the reviewed recommendation version and evidence snapshot.
Worked verification: synthetic service-desk procurement
A fictional organization compares three AI-assisted service-desk tools for a twelve-month controlled rollout. The draft recommends Vendor North because it scores 86/100, ahead of Vendor Field at 81 and the existing workflow at 63. The review packet contains the requirements, proposals, demonstrations, reference notes, a cost workbook, and security questionnaires.
TRACE first finds sound decision alignment: the recommendation uses the issued criteria and separates two mandatory controls from weighted preferences. The claim review, however, finds that North's stated regional data-residency capability comes from a product-marketing page. Its signed proposal says the relevant processing route is available only on a higher tier, while the price workbook uses the standard tier. The item is changed from pass to not evidenced.
The comparison check then finds unlike cost boundaries. North's total excludes implementation and premium support; Field's total includes both. Restating both options over twelve months adds £74,000 to North and £18,000 to Field. An independent recalculation produces 78 for North and 81 for Field. The reviewer does not simply switch the recommendation: the scoring difference is within the approved sensitivity band, and Field has not yet passed the required failure-recovery test.
The operating challenge adds four test cases: a normal ticket with a complete knowledge article, a conflicting pair of articles, a ticket containing personal data, and loss of the retrieval service. Field passes the first three but fails to return a clear degraded-mode message during the outage. The failure could expose staff to an apparently complete answer without supporting material.
The disposition is release with visible conditions for a two-vendor final test, not an award recommendation. North must evidence the priced tier's residency route. Field must repair and retest degraded mode. Both must submit a portable export. The decision record names owners and a five-day deadline. This result is useful precisely because it does not convert missing evidence into a confident ranking.
Two Prune boundary
Two Prune can assess whether a participant traces a supplied procurement recommendation to evidence, checks calculations and contradictions, identifies operational risks, and communicates a bounded disposition. It reports observed performance in that assessment. It does not run or approve an organization's procurement, provide legal or financial advice, certify suppliers, validate a product's controls, or make employment decisions.
Sources
US Federal Acquisition Regulation 15.305: Proposal evaluation
- UK Government: The Sourcing Playbook
- NIST: AI Risk Management Framework Core
Sources: [1] [2] [3]
Sources
- 1.FAR 15.305 · U.S. Acquisition.gov
- 2.Sourcing Playbook · UK Government
- 3.NIST AI RMF Core · National Institute of Standards and Technology