AI account plan verification checklist

What this checklist verifies

Salesforce describes an account plan as a living record that develops as teams research an account, connect with stakeholders, define objectives, and monitor progress. Its account-planning guidance also distinguishes hard facts from informed ideas that must be refined through customer conversations. NIST's AI Risk Management Framework adds a second requirement for AI-assisted work: document intended use, knowledge limits, human oversight, risk controls, tests, and residual risk.

Together, those sources support a practical verification rule: treat the account plan as a controlled decision record. Confirm what is known, expose what is inferred, and keep the evidence needed to reproduce material conclusions.

This checklist is educational. It does not validate a CRM configuration, certify an AI system, establish legal compliance, or prove future commercial outcomes.

The TRACE check

Run five passes in order. A failed earlier pass should not be hidden by a strong later one.

### 1. Trace every material statement

Record verified, contradicted, stale, hypothesis, or missing beside each material statement. A source link alone is not enough when the source does not support the wording used.

### 2. Recalculate numbers

The check is reconciliation, not plausibility. A number that looks reasonable still fails when its inputs cannot be found.

### 3. Assign ownership and approval

### 4. Challenge completeness and contradictions

### 5. Expose residual uncertainty

• Is each customer fact linked to an approved source, observation date, and responsible recorder?
• Are customer statements distinguishable from internal interpretations?
• Are stakeholder roles supported by a current record or direct conversation?
• Are AI-generated summaries checked against the retained input?
• Are stale or conflicting records visible rather than silently reconciled?
• Can another reviewer reproduce every total, percentage, forecast input, and date calculation?
• Does the plan identify the system of record for commercial numbers?
• Are units, currencies, time periods, exclusions, and rounding rules stated?
• Are model-produced estimates visibly separate from approved figures?
• Has a human owner reviewed any number that will be shared externally?
• Does every next action have one named owner, a due date, and completion evidence?
• Is the person approving external use identified?
• Are pricing, procurement, privacy, security, and contractual questions routed to the appropriate owner?
• Is any approval implied from silence or a job title?
• Does the plan say what must happen before an unverified claim can be promoted to fact?
• Does the plan answer the decision it was created to support?
• Are missing stakeholders, documents, and customer confirmations listed?
• Do objectives connect to actions and observable evidence?
• Are disagreements between meeting notes, CRM records, and generated text resolved or disclosed?
• Is irrelevant personal or confidential information excluded from the AI workflow?
• Which material questions remain unanswered?
• What consequence could follow if each uncertain item is wrong?
• What evidence-producing action will reduce the uncertainty?
• What date or event triggers the next review?
• Which parts of the plan must not be used until that review occurs?

Downloadable checklist

Copy this table into a document or spreadsheet. Add rows when the plan contains several objectives or calculations.

| Check ID | Item under review | Required evidence | Result | Owner | Resolution or approval | Review trigger | | --- | --- | --- | --- | --- | --- | --- | | T1 | Material customer claim | Source, date, exact supporting passage | Pass / fail / hold | Account owner | Correct wording or obtain confirmation | New customer evidence | | R1 | Commercial calculation | Inputs, formula, units, system of record | Pass / fail / hold | Commercial owner | Recalculate and approve | Input or scope changes | | A1 | External action | Named owner, due date, completion evidence | Pass / fail / hold | Action owner | Assign or remove | Due date or dependency | | C1 | Contradictory record | Both sources, dates, resolution rationale | Pass / fail / hold | Evidence owner | Resolve or disclose | Source update | | E1 | Residual uncertainty | Consequence, evidence route, use restriction | Pass / fail / hold | Decision owner | Investigate, qualify, or stop | Evidence arrives |

Use three release dispositions:

• Pass: evidence supports the item and required approval is recorded.
• Repair: the issue is bounded and an owner can correct it before use.
• Hold: the missing or contradictory evidence could change the decision, calculation, or external claim.

Worked example: verify a synthetic renewal plan

An original synthetic plan for Northbank Learning says: "The customer will expand to three regions, creating a $240,000 opportunity. The chief learning officer approved a November workshop."

The reviewer completes TRACE:

| Check | Finding | Disposition | | --- | --- | --- | | Trace | Three-region expansion appears only in an internal brainstorm. The workshop request appears in a dated customer note. | Remove the expansion statement; retain the workshop request as verified. | | Recalculate | $240,000 equals three assumed regions multiplied by an unapproved $80,000 package. Neither input is approved. | Hold the value and route pricing to the commercial owner. | | Assign | The plan names a chief learning officer but contains no approval record or budget role. | Replace "approved" with "stakeholder to confirm authority and scope." | | Challenge | CRM lists a 30 November renewal while a meeting note says 15 December. | Show the conflict and ask the account owner to reconcile it with the system of record. | | Expose | Procurement steps and permitted data handling are unknown. | Add evidence-producing actions and prohibit external sharing until resolved. |

The corrected plan does not merely sound less confident. It changes the artifact: unsupported revenue is removed, approval is no longer invented, the date conflict becomes visible, and each gap has an owner and review trigger.

Every organization, person, value, and event in this example is synthetic. It is not customer proof.

Common checklist failures

A checklist fails when it becomes ceremonial. Warning signs include every row marked pass without evidence, the same person drafting and approving sensitive claims, calculations reviewed only for arithmetic, and uncertainty recorded without a consequence or resolution route.

Another failure is verifying prose while ignoring provenance. A concise account plan with a source map is stronger evidence than a detailed plan assembled from unapproved exports and unverifiable AI summaries.

How to assess the work

In a Two Prune assessment, provide a participant with a CRM extract, meeting notes, an old organization chart, a pricing table, and an AI-generated draft containing one unsupported opportunity, one date conflict, and one implied approval. Ask for the completed TRACE table, a corrected account plan, and a short change log.

Observe whether the participant finds the unsupported claim, reproduces the calculation, distinguishes a title from approval authority, retains the contradiction, and routes missing evidence. These artifacts can provide task-bounded evidence about Context Framing, Evidence Navigation, AI Orchestration, Verification & Risk Control, Judgment & Synthesis, and Communication & Delivery.

Two Prune assesses performance in a realistic work simulation. It does not certify account plans, predict commercial performance, rank people, or make employment decisions.

Sources

Salesforce Trailhead: Explore and create account plans

• Salesforce Trailhead: Set objectives and take action
• NIST: AI Risk Management Framework Core
• NIST: Generative Artificial Intelligence Profile

Sources: [1] [2] [3] [4]