Practical guide

AI partnership assessment verification checklist

A partnership assessment is ready for a decision only when another reviewer can trace its important claims, reproduce its calculations, see unresolved contradictions, and understand the recommendation boundary. This checklist verifies a completed assessment. It does not replace the broader work of comparing partnership with other routes, investigating a prospective partner, or obtaining legal.

By Two Prune

A partnership assessment is ready for a decision only when another reviewer can trace its important claims, reproduce its calculations, see unresolved contradictions, and understand the recommendation boundary. This checklist verifies a completed assessment. It does not replace the broader work of comparing partnership with other routes, investigating a prospective partner, or obtaining legal and specialist advice. Use CHECK after the assessment has been drafted: confirm the decision frame, hunt every material claim back to evidence, examine calculations and contradictions, check terms and operating risks, and keep a reproducible decision record. Copy the checklist into the review file and preserve failures rather than silently editing them away.

C: Confirm the decision frame and review boundary

Start with the question the assessment must answer. Record the proposed collaboration, purpose, scope, geography, duration, decision owner, affected customers or users, and realistic alternatives. A reviewer should be able to tell whether the evidence supports this decision rather than a broader endorsement of the organisation.

Check that the assessment distinguishes a limited experiment, ordinary supplier relationship, alliance, and formal joint venture. Each can create different commercial, governance, competition, tax, data, and legal questions. The review should identify which questions require qualified advice and should not imply that this educational checklist completes that work.

Mark each item pass, fail, not applicable, or unknown:

The UK Competition and Markets Authority's joint-venture advice asks collaborators to define the true purpose, scope, duration, customer benefit, and limits of collaboration. Those questions make a useful scope test, but the advice does not approve a particular arrangement.

  • [ ] The decision, accountable owner, due date, and intended users are named.
  • [ ] Business as usual and at least one non-partnership route are compared.
  • [ ] Scope, geography, duration, excluded activities, and exit conditions are explicit.
  • [ ] The assessment separates verified facts, estimates, assumptions, and judgments.
  • [ ] Specialist reviews are assigned where the structure or jurisdiction requires them.

H: Hunt material claims back to evidence

Build a claim ledger for every statement that could change the decision. Give each claim an identifier, quote or paraphrase the claim, link the source, name the source owner, record the publication or retrieval date, and explain what the source does and does not establish.

Prefer primary evidence for legal status, ownership, financial information, certifications, customer commitments, security controls, delivery capacity, and intellectual-property rights. Interviews and partner-supplied materials may be necessary, but label them as representations until independently checked. If two sources conflict, retain both and create a disposition rather than choosing the more favourable one without explanation.

  • [ ] Every material claim has an attributable source or is marked unsupported.
  • [ ] Source dates are suitable for the decision date.
  • [ ] Partner representations are distinguished from independent records.
  • [ ] Access restrictions and missing documents are visible.
  • [ ] AI-assisted summaries link back to the underlying material and were checked.
  • [ ] Contradictory evidence has an owner, next action, and deadline.

E: Examine calculations, comparisons, and scenario assumptions

Reperform the calculations that influence the recommendation. This may include demand estimates, capacity, staffing, unit economics, implementation cost, revenue shares, currency conversion, delivery dates, or downside exposure. Record formulas, units, periods, source cells, rounding, and scenario assumptions.

Test at least one downside case and one break-even or stopping case. A polished base case is not enough when one uncertain assumption drives the conclusion.

  • [ ] Units, currencies, dates, and denominators are consistent.
  • [ ] Formula inputs trace to the evidence ledger.
  • [ ] Duplicates, omissions, and one-off items are handled explicitly.
  • [ ] Base, downside, and stopping cases use stated assumptions.
  • [ ] A second reviewer can reproduce decision-critical totals.
  • [ ] The recommendation changes appropriately when a critical assumption fails.

C: Check contributions, governance, information boundaries, and exit risk

Reconcile what each party contributes with the evidence that the contribution is available. Cover people, funding, systems, intellectual property, data, market access, certifications, facilities, and decision time. Then check who can decide, approve exceptions, access information, resolve disputes, change scope, and close the relationship.

The assessment should not treat a contract clause as proof that a control operates. Evidence may include named owners, current procedures, configured permissions, acceptance tests, incident records, and completed reviews. Competition-sensitive information must stay within a justified boundary. The CMA specifically warns against sharing sensitive business information beyond the collaboration's legitimate scope.

  • [ ] Contributions, dependencies, owners, evidence, and shortfalls are reconciled.
  • [ ] Decision rights, reserved matters, escalation, and deadlock routes are recorded.
  • [ ] Data, confidentiality, intellectual-property, and information-sharing limits are explicit.
  • [ ] Conflicts, incentives, related parties, and concentration risks are addressed.
  • [ ] Review, renewal, suspension, and exit triggers have executable actions.
  • [ ] Residual risks are accepted by a named authority rather than buried in prose.

K: Keep a reproducible decision record

Package the decision so a later reviewer can understand what was known at approval. Preserve the assessment version, evidence ledger, calculation workbook, contradiction log, reviewer comments, specialist opinions, decision, conditions, expiry date, and change triggers. Do not overwrite a failed check after repair; append the repair and retest evidence.

The final record should state one of four outcomes: stop, obtain missing evidence, run a bounded experiment, or proceed within named conditions. It should also say what would invalidate the decision.

Worked artifact: checking a regional training partnership assessment

A synthetic training provider is considering a regional delivery alliance. Its assessment recommends a six-month partnership because the prospective partner reports local customer access, bilingual facilitators, and public-sector delivery experience.

The CHECK review finds three decision-changing issues. First, the customer pipeline figure mixes signed work with unqualified prospects. The reviewer relabels the categories and the downside case falls below the staffing break-even point. Second, two facilitator profiles are current, but the proposed certification is still pending. Third, the draft says customer data will remain in-region, while the proposed learning platform uses a subprocessor elsewhere.

The review does not replace the recommendation with a negative verdict. It records a conditional outcome: run one synthetic delivery exercise only after the pipeline categories are corrected, facilitator availability is evidenced, the data route receives specialist review, and the parties agree an information-sharing boundary. The stop conditions are a failed delivery rehearsal, unresolved data transfer issue, or inability to staff the agreed service window.

This artifact demonstrates the checklist's purpose. It reveals where evidence quality changes the available next step without pretending to predict the partnership's success.

Verification and Two Prune boundary

Verification is complete when all checklist items have a disposition, material failures are reflected in the decision, calculations can be reproduced, conditions have owners and dates, and the record identifies what remains unknown. Sampling only favourable claims or treating missing evidence as a pass is a failed review.

Two Prune can place a participant in a synthetic partnership-assessment task and capture evidence of research, source use, calculation checks, contradiction handling, judgment, and communication. It does not perform corporate due diligence, approve a partner, provide legal or investment advice, predict later workplace performance, or make employment decisions. Assessment evidence remains task-bounded and requires human interpretation.

Primary sources

UK Competition and Markets Authority, Joint Venture Business Advice

  • Business.gov.uk, How to set up joint ventures abroad
  • World Bank, Partnership Strategies

Sources: [1] [2] [3]