Practical guide

How to create a procurement recommendation with AI

A procurement recommendation is not a ranking table with a preferred supplier highlighted. It is a decision record that lets an authorized reader reconstruct the need, the options considered, the evidence accepted, the commercial assumptions, the unresolved dependencies, and the conditions attached to the recommendation. AI can help organize evidence, compare like with like, test calculations,.

By Two Prune

A procurement recommendation is not a ranking table with a preferred supplier highlighted. It is a decision record that lets an authorized reader reconstruct the need, the options considered, the evidence accepted, the commercial assumptions, the unresolved dependencies, and the conditions attached to the recommendation. AI can help organize evidence, compare like with like, test calculations, and locate gaps. The accountable procurement and business owners must set the criteria, judge evidence, resolve conflicts, and approve the decision. Use the RECORD structure below: requirement, evidence, comparison, operating consequences, recommendation, and decision controls. Keep the source packet beside the finished artifact so every material statement can be checked.

1. Requirement: state the decision before describing suppliers

Open with the decision required, authorized decision owner, business outcome, scope, mandatory conditions, evaluation period, budget boundary, and date. Separate the underlying need from a proposed product. Include credible non-purchase options such as retaining the current process, improving it without a new supplier, building internally, or running a bounded pilot.

Record the evaluation criteria and their relative importance as they existed before proposals were assessed. If a criterion or weight changed, preserve the prior version, reason, authority, and effect on the result. This protects the recommendation from being reverse-engineered around a preferred option.

The U.S. Federal Acquisition Regulation requires covered competitive proposals to be assessed solely on stated factors and says the strengths, deficiencies, significant weaknesses, and risks supporting the evaluation should be documented. Those rules apply to their federal context; the transferable discipline is to make the comparison basis visible rather than silently changing it. FAR Subpart 15.3

Sources: [1]

2. Evidence: build a claim-to-source register

Create one row for every fact that could change the decision.

| Field | What to record | | --- | --- | | Claim | Exact statement used in the recommendation | | Source | Proposal section, contract term, demonstration record, reference, test, or internal analysis | | Status | Verified, supplier assertion, inferred, conflicting, stale, missing, or not applicable | | Scope | Product version, geography, volume, term, configuration, and date | | Owner | Person responsible for verification or resolution | | Decision effect | What changes if the claim is wrong |

Do not upgrade a supplier statement into an observed capability. A demonstration script, customer reference, security report, or contractual commitment may provide stronger evidence, but each answers a different question. If AI extracts proposal claims, compare the extraction with the original passages and keep page-level citations. Prompt text and model output are working material, not evidence about the supplier.

3. Comparison: normalize options without hiding tradeoffs

Give every viable option an evidence card using the same fields: fit to mandatory conditions, scored strengths, material gaps, total-cost range, implementation demand, data and integration dependencies, delivery evidence, exit implications, and residual uncertainty. Explain every disqualification separately from weighted scoring.

Use scores only as navigation. Show the underlying observations and run a sensitivity test. Ask whether a reasonable change in cost, volume, schedule, weight, or unresolved evidence reverses the leading option. If it does, describe the recommendation as conditional or sensitive.

FAR 15.4 distinguishes overall price analysis from analysis of cost elements and says the level of detail should reflect the circumstances. That is not a general procurement mandate, but it supports keeping the method and assumptions visible. FAR Subpart 15.4

Sources: [2]

4. Operating consequences: describe the work after selection

A recommendation must show more than purchase price. State the implementation work, internal capacity, controls, training, migration, service management, renewal, data return, transition, and exit effort each option creates. Name who would own each dependency and when it must be resolved.

The UK Government's Sourcing Playbook asks in-scope public organizations to consider delivery models, should-cost estimates, testing, risk allocation, supplier financial standing, resolution planning, and end-of-contract arrangements. It is public-sector guidance, not a universal rulebook, but it illustrates why a recommendation needs lifecycle consequences rather than an award score alone. The Sourcing Playbook

Sources: [3]

5. Recommendation: make the logic and conditions explicit

Write the recommendation as a conditional decision:

> Select [option] for [scope] because [evidence-backed reasons]. Do not proceed until [conditions] are satisfied. If [failure or changed assumption] occurs, [fallback, repair, or stop action] applies.

Then state why the closest alternative was not selected, which uncertainty remains, and what evidence would change the conclusion. Avoid claims such as “best supplier” unless the scope, criteria, date, and evidence make that statement defensible.

6. Decision controls: obtain domain-specific approvals

List the decision owner and the people approving commercial, financial, legal, privacy, security, accessibility, technical, operational, and records matters where relevant. Approval by one function does not substitute for another. Record dissent, exceptions, expiry, review triggers, and where the signed decision packet is retained.

Before release, have an independent reviewer perform five checks:

  • trace each material claim to a source;
  • recalculate totals, units, ranges, and denominators;
  • test missing and conflicting evidence;
  • rerun the stated sensitivity cases; and
  • confirm that every condition has an owner, due date, and consequence.

Worked example: service-desk knowledge assistant

A synthetic company is considering a knowledge assistant for its internal service desk. The approved criteria are workflow fit 30%, security and data handling 25%, three-year cost 20%, implementation feasibility 15%, and exit portability 10%. Three options remain: Vendor Cedar, Vendor Flint, and improve the current search process without a new supplier.

AI extracts proposal claims and constructs first-pass evidence cards. A reviewer finds that Cedar's statement about data residency is limited to one hosting configuration, Flint's price excludes migration, and the internal option lacks capacity for document cleanup. The total-cost model is rebuilt on common volumes and includes migration, internal review, support, and exit work.

| Option | Evidence-backed advantage | Material gap | Three-year range | Decision status | | --- | --- | --- | ---: | --- | | Cedar | Scripted test covers the required ticket workflow | Residency evidence does not cover the proposed configuration | 420–510k | Conditional | | Flint | Contract draft includes the required hosting region | Migration and retrieval-quality evidence are incomplete | 390–535k | Clarification required | | Improve current search | Lowest external spend and simplest exit | No funded owner for content cleanup and quality review | 180–310k | Not currently feasible |

Cedar leads under the approved weights, but a 20% migration overrun or failure to confirm the proposed hosting configuration changes the result. The recommendation therefore says: select Cedar only if security verifies configuration-specific evidence, a capped migration plan is agreed, and a 30-case pilot meets predeclared retrieval and escalation checks. If either boundary fails, return to Flint clarification or the current-process option; do not silently proceed because Cedar had the highest initial score.

The figures, vendors, and decision are synthetic. The worked artifact demonstrates traceability and conditional reasoning, not supplier fitness or price reasonableness.

How this fits a Two Prune assessment

A Two Prune assessment can ask a participant to create or audit a procurement recommendation from synthetic proposals, cost data, policy excerpts, demonstration notes, and conflicting messages. The assessment can observe source navigation, AI use, calculation checks, uncertainty handling, revision, and the final decision record. Two Prune reports observed performance and evidence strength. It does not select suppliers, certify procurement compliance, or make employment decisions.

Sources and limits

The RECORD method and worked example are original. Adapt them to the applicable law, policy, procurement route, authority, and consequence. This page is educational and is not legal, commercial, financial, or procurement advice.

  • Acquisition.gov: FAR Subpart 15.3 — attributed federal requirements for stated evaluation factors and documented findings.
  • Acquisition.gov: FAR Subpart 15.4 — attributed distinctions among price, cost, and cost-realism analysis in covered acquisitions.
  • GOV.UK: The Sourcing Playbook — current public-sector guidance on delivery models, cost, risk, testing, supplier resilience, and exit.

Sources: [1] [2] [3]

Sources

  1. 1.FAR Subpart 15.3 · U.S. Acquisition.gov
  2. 2.FAR Subpart 15.4 · U.S. Acquisition.gov
  3. 3.The Sourcing Playbook · UK Government

Explore this topic