Practical guide

AI process map verification checklist

A process map is ready for use only when another person can trace real work through it, find the evidence behind consequential steps, route exceptions without guessing, and identify who can approve or stop the process. Use this checklist after a draft map exists. It tests the artifact rather than rewarding visual polish, and it deliberately includes normal, exception, and contradictory cases.

By Two Prune

A process map is ready for use only when another person can trace real work through it, find the evidence behind consequential steps, route exceptions without guessing, and identify who can approve or stop the process. Use this checklist after a draft map exists. It tests the artifact rather than rewarding visual polish, and it deliberately includes normal, exception, and contradictory cases.

Prepare the verification packet

Gather the map version, stated purpose, process owner, source procedures, representative records, system screenshots or logs, decision rules, exception records, measures, approval history, and change log. Use records from the period the map claims to describe. If the map is a proposed future state, label assumptions and keep them separate from observed current-state evidence.

Select at least three test tokens:

The GOV.UK Service Manual recommends mapping the wider journey, including online and offline touchpoints, backend processes, involved organizations, and evidence users must provide. It also highlights broken handoffs and dead ends as problems a map should reveal. That guidance supports a whole-journey check; it does not determine the correct process for a particular organization. Map and understand a user’s whole problem

  • a normal case that should reach the intended end state;
  • an exception that should stop, loop, or escalate; and
  • a contradictory case in which two sources or instructions disagree.

Sources: [1]

Run the eight-check verification ledger

Copy the following ledger into the review record. Mark each item pass, fail, not observed, or not applicable. “Not observed” is a missing-evidence finding, not a pass.

### 1. Boundary and purpose

Failure test: ask two reviewers to state what one token represents and where the process ends. If their answers differ materially, the boundary fails.

### 2. Source traceability

NIST’s AI Risk Management Framework calls for documentation, defined roles, human oversight, measurement, and monitoring across the AI lifecycle. Use those outcomes as prompts for traceability and ownership, not as evidence that a mapped workflow is effective or low-risk. NIST AI RMF Core

### 3. Actor and handoff integrity

Failure test: remove one expected input at a handoff. The map should show who detects the absence and what they do next.

### 4. Decision and exception logic

### 5. Calculation and transformation checks

For a process without arithmetic, test the transformation rule instead. “AI summarizes request” is incomplete unless the map says what source material is allowed, how omissions are checked, and what happens when the summary conflicts with the original.

### 6. Control operation

### 7. Measurement and feedback

NIST’s Baldrige commentary treats operations as processes to design, manage, measure, and improve, and notes process mapping, error proofing, performance review, and corrective action as possible improvement practices. It is an improvement framework, not a certification of this checklist. Baldrige Criteria Commentary

### 8. Approval and residual uncertainty

  • The trigger, finish condition, unit of work, audience, and decision supported by the map are explicit.
  • In-scope channels, teams, systems, and adjacent processes are named.
  • Current state, proposed state, and aspiration are visually and textually distinguishable.
  • The version date and person responsible for maintaining the map are present.
  • Each consequential rule points to an authoritative source, owner, and version.
  • Inputs and outputs at each transformation are identifiable.
  • AI prompts, model responses, and retrieved passages are not treated as proof of the underlying fact.
  • A reviewer can move from a mapped decision to the retained record showing what happened.
  • Every step has a performing role; consequential decisions have an accountable owner.
  • Systems and AI services occupy distinct lanes from people.
  • Every handoff states what moves, in what form, and what confirms receipt.
  • Waiting, queues, offline activity, external parties, and rework are visible.
  • A failed handoff has an owner and recovery route.
  • Each decision diamond is phrased as a testable condition.
  • Every branch has a destination, including “no,” “unknown,” and tool-unavailable routes.
  • Missing, conflicting, stale, sensitive, or unauthorized inputs have explicit treatment.
  • Overrides require named authority and a retained reason.
  • Rework returns to a specific step with a closure condition.
  • Any calculation names its inputs, units, formula or rule, rounding, and output.
  • Counts and percentages have denominators.
  • Classification, summarization, extraction, or ranking steps identify their acceptance check.
  • A changed input can be traced to every affected downstream value or decision.
  • Each control names its performer, timing, pass criterion, retained evidence, and failure action.
  • Human review describes what the reviewer checks rather than merely saying “approve.”
  • Critical failures remain visible and cannot be offset by success elsewhere.
  • The person authorized to pause or stop the process is identifiable.
  • Measures relate to the map’s purpose and include definitions, denominators, owners, and review cadence.
  • End-to-end time is separated from waiting and active work where that distinction matters.
  • Quality, rework, exceptions, overrides, and complaints are not hidden behind speed.
  • A signal or event triggers review of the map itself.
  • The map records who reviewed content, control logic, data handling, accessibility, and operating ownership where relevant.
  • Unresolved assumptions and missing evidence are listed beside their consequence.
  • Approval has a date, scope, and expiry or change trigger.
  • The map does not imply that approval removes residual risk.

Sources: [2] [3]

Worked verification: purchase-order exception map

This synthetic map covers a purchase order whose price differs from an approved supplier catalogue. The normal token matches the catalogue, passes budget validation, and reaches release. The exception token differs by 7%, and the contradictory token has a current catalogue price but an older email authorizing a temporary increase.

| Check | Test performed | Result | Repair required | | --- | --- | --- | --- | | Boundary | Reviewers trace one purchase-order line from intake to release | Pass | None | | Source | Price decision cites “supplier data” without version | Fail | Name catalogue ID, effective date, and owner | | Handoff | Buyer sends exception to finance; receipt is not recorded | Fail | Add queue acknowledgement and ageing owner | | Decision | “Price acceptable?” has no unknown route | Fail | Add unavailable/conflicting-source branch | | Calculation | Variance is shown but denominator is absent | Fail | Specify (quoted - approved) / approved and rounding | | Control | Finance “reviews” but test and evidence are unspecified | Fail | Check authority, calculation, source versions, and retained decision | | Feedback | Exception count is tracked without total orders | Fail | Add eligible-order denominator and reporting period | | Approval | Map owner is named; unresolved email authority is not | Fail | Record policy owner’s decision and residual uncertainty |

The normal token alone would make the map look adequate. The contradictory token exposes the real weakness: the artifact does not say whether the catalogue or temporary authorization governs, who resolves the conflict, or how the decision is retained. Verification ends with a repair list, not a cosmetic redrawing exercise.

Acceptance rule

Accept the map only when all critical items pass and every non-critical failure has an owner and due date. Treat source authority, decision ownership, exception routing, sensitive-data handling, stop authority, and critical calculations as non-compensatory. A map with one of those missing is not ready merely because most boxes are correct.

Repeat verification after a material change in policy, model, tool, data source, team, volume, control, or consequence. Sample a recent real case where permitted; tabletop testing cannot prove that the process operates as drawn.

How this fits a Two Prune assessment

A Two Prune assessment can ask a participant to verify a synthetic process map against supplied workplace evidence. The assessment can record source navigation, contradiction handling, calculation checks, AI use, escalation, and the quality of the repair ledger. Two Prune reports observed assessment evidence and score confidence. It does not certify the customer’s process or make an employment decision.

Sources and limits

The checklist and purchase-order example are original and synthetic. These sources support review questions; they do not establish that passing the checklist produces a particular operational outcome.

  • GOV.UK: Map and understand a user’s whole problem — whole-journey mapping, organizational boundaries, touchpoints, backend processes, and evidence.
  • NIST AI RMF Core — documentation, roles, human oversight, measurement, and monitoring considerations.
  • NIST Baldrige Criteria Commentary — process design, operational measures, review, and improvement prompts.

Sources: [1] [2] [3]

Sources

  1. 1.Map and understand a user’s whole problem · UK Government
  2. 2.NIST AI RMF Core · National Institute of Standards and Technology
  3. 3.Baldrige Criteria Commentary · National Institute of Standards and Technology