Practical guide
How to create an AI-aware process map
An AI process map should show more than a sequence of boxes. It should reveal where evidence enters, where AI transforms it, who remains accountable, which exceptions can change the route, and what record proves each control operated. The useful artifact is a testable operating model: someone unfamiliar with the workflow should be able to follow one normal case and one exception without.
An AI process map should show more than a sequence of boxes. It should reveal where evidence enters, where AI transforms it, who remains accountable, which exceptions can change the route, and what record proves each control operated. The useful artifact is a testable operating model: someone unfamiliar with the workflow should be able to follow one normal case and one exception without inventing a missing handoff.
Start with a bounded mapping question
Name the trigger, end condition, unit of work, and decision the map must support. “Map marketing” is too broad. “Map how a product claim moves from campaign brief to a published landing page” is bounded enough to test.
Collect evidence before running a workshop: the current brief, source-of-truth product material, approval policy, recent examples, system timestamps, exception logs, and the names of people who perform and receive the work. Ask each contributor what they receive, what they change, what they decide, what they pass on, and what makes them stop or escalate. Observed records should take precedence over an idealized procedure.
NIST’s AI Risk Management Framework treats risk management as continuous across the lifecycle and calls for documented roles, monitoring, human review, and multidisciplinary perspectives. Those ideas are useful prompts for an AI-assisted process map, but using them does not establish that a workflow is effective or low-risk. NIST AI RMF Core
Sources: [1]
Use the LANES method
LANES turns a flowchart into a reviewable evidence artifact.
### L — Limit the journey
Write the start event, finish event, customer or internal outcome, in-scope channels, and excluded adjacent processes. Choose one unit that can be traced end to end, such as one landing-page claim. Record the version date and map owner so an old map cannot silently become the current procedure.
### A — Assign actors and systems
Give every step one performing role and, where needed, a separate accountable decision owner. Put systems, AI services, and external providers in their own lanes. An AI system can draft, classify, retrieve, or compare; it cannot absorb an unnamed person’s accountability. Mark where data crosses a tool or organizational boundary.
### N — Name nodes and evidence
Each node should use a verb and object: “verify claim against approved specification,” not “quality.” Add its input, output, source of truth, and retained record. A prompt and model response may be evidence that AI was used, but neither proves the underlying claim was correct. Link verification to the authoritative source and its version.
### E — Expose exceptions and controls
Draw decision diamonds for conditions that actually alter the path. Include missing evidence, conflicting sources, sensitive data, a stale approval, tool failure, and a requested override. For each control, state who performs it, what constitutes a pass, where the result is recorded, and what happens on failure. Do not hide rework in a backwards arrow labelled “fix.”
### S — Set signals and change rules
Choose signals that reveal whether the mapped process works: end-to-end time, waiting time, first-pass acceptance, exception frequency, rework, unsupported-claim catches, and overrides. Define the denominator and owner. The GOV.UK Service Manual recommends combining performance metrics with user research and using multiple data sources rather than relying only on digital analytics. That supports a mixed evidence plan, not a universal metric set. Measuring the success of your service
Set a review trigger, such as a source-system change, new AI model, revised approval policy, material incident, or sustained shift in exception rate.
Sources: [2]
Worked artifact: a campaign-claim approval map
This synthetic example maps one factual claim proposed for a campaign landing page.
| Step | Performing lane | Input and action | Evidence artifact | Exception and route | | --- | --- | --- | --- | --- | | 1. Register claim | Campaign manager | Enter audience, region, proposed claim, launch date, and brief ID | Claim record C-104 | Missing region: return to brief owner | | 2. Retrieve support | Content specialist + approved AI tool | Search the approved product library and list candidate passages with document versions | Retrieval log and cited passages | No current support: mark “unsupported” and stop | | 3. Draft wording | Content specialist + AI tool | Produce variants constrained to the cited facts | Prompt, response, selected draft | AI adds an uncited comparison: discard variant | | 4. Verify claim | Product evidence owner | Compare subject, scope, quantities, date, and region against the source | Five-field verification record | Source conflicts with release notes: escalate to product owner | | 5. Review presentation | Brand reviewer | Check meaning is preserved in headline, body, footnote, and visual | Review decision and comments | Qualification becomes unreadable: revise layout | | 6. Authorize publication | Campaign owner | Confirm evidence and required reviews are current | Timestamped authorization | Approval older than policy window: re-verify | | 7. Monitor | Campaign operations | Watch complaints, corrections, source changes, and overrides | Weekly exception register | Material source change: unpublish or amend |
Trace two tokens through the map. The normal token has a current regional specification and reaches publication with a complete verification record. The exception token relies on an old global claim while the regional release note narrows the feature. It must stop at step 4, preserve the conflict, and return to the product owner. If the map cannot show that second route, it is not ready for operational use.
Acceptance test for the map
Before approving the artifact, ask a reviewer who did not draw it to perform a tabletop walk-through.
Reject a map that omits waiting states, collapses several decisions into “review,” assigns steps to a department rather than a responsible role, or treats the happy path as the whole process. Also reject decorative precision: a large diagram is not stronger merely because it has more boxes.
- Can they identify the current source of truth at every consequential transformation?
- Can they distinguish AI assistance from human authorization?
- Can they locate the record showing a control passed?
- Can they route a missing, conflicting, or sensitive input without guessing?
- Can they tell which signal would trigger a map revision?
How this connects to an assessment
In a Two Prune assessment, a participant can be asked to construct or critique a process map from synthetic workplace materials. The evidence may include which sources they inspect, whether they distinguish transformation from authorization, how they handle a planted conflict, and whether their final artifact exposes controls and exceptions. Two Prune reports observed assessment performance and evidence strength. It does not certify the mapped process or make an employment decision.
Sources and limits
These sources provide useful public frameworks, not evidence that LANES or any mapped workflow will produce a particular result. The example is synthetic and should be adapted to the organization’s actual policy, systems, risks, and records.
- NIST AI RMF Core — lifecycle governance, documentation, roles, monitoring, and human review considerations.
- GOV.UK: Measuring the success of your service — guidance on combining performance measures, user research, and multiple data sources.
- GAO: Artificial Intelligence — Key Practices to Help Ensure Accountability in Federal Use — an official accountability framing organized around governance, data, performance, and monitoring.
Sources: [1] [2] [3]
Sources
- 1.NIST AI RMF Core · National Institute of Standards and Technology
- 2.Measuring the success of your service · UK Government
- 3.GAO: Artificial Intelligence — Key Practices to Help Ensure Accountability in Federal Use · U.S. Government Accountability Office