AI Is Generating Journey Maps without Evidence

What the journey problem looks like and why it affects the business

AI can produce a coherent journey map from a short prompt, but coherence is not evidence. If generated stages, emotions, prevalence or causes are copied into plans, the organisation may mistake synthesis for research and expose personal information or make decisions no source can support.

Use AI as an evidence-bounded synthesis aid with provenance, minimisation, uncertainty, human review, prohibited inputs and explicit approval boundaries.

This is a governance and learning decision before it is a documentation problem. The primary reader is Customer Director, Research Lead or AI Governance Lead. The economic buyer is the customer, research or AI governance executive accountable for trustworthy insight, privacy, inclusion and decisions influenced by AI-generated journey material.

The named owner holds decision authority; specialists contribute evidence within their remit, but contribution does not replace accountable ownership.

Use AI to organise supplied journey evidence. Generated output remains material for hypotheses; it cannot create personas, research evidence, factual proof or approval.

What to expect in the working pack

Start the working pack with the charter. Use stable IDs to connect sources, tests, decisions and review records across its eight sheets. Entries need supporting evidence; the workbook alone cannot prove a conclusion:

• Start here: Read the bounded governance or learning decision, evidence labels, sequence and stop conditions.
• Journey charter: Define the user outcome, start and end, routes, exclusions, consequence and authority.
• Journey map: Connect user tasks, context, channels, systems, handoffs, owners and unresolved questions.
• Evidence register: Separate observed facts and statements from interpretations, inferences, proposals and unknowns.
• Friction and failure log: Record symptoms, affected routes, severity, recurrence, consequences and competing causes.
• Intervention and acceptance: Link a proposed governance, measurement or release change to acceptance, guardrails and rollback.
• Outcome and recovery: Observe progression, value, support, recovery, exceptions, harm and durable learning.
• Review and change log: Record versions, evidence changes, approvals, review triggers, replacement and retirement.

Use the same IDs across the workbook so a reviewer can move from a claim to source, definition, consequence, proposed change, outcome and decision. Seed rows are prompts, not local evidence; replace them with authorised organisation-specific material. Do not enter participant names, contact details, recordings, free-text case notes, AI prompts containing identifiers or unnecessary personal information. Store sensitive material in an approved controlled repository and record only a permitted reference. The workbook does not create consent, lawful authority or publication approval.

What evidence distinguishes the real cause

The visible symptom does not establish the cause. Before changing a measure, release gate, owner, definition or AI workflow, build a dated evidence set that distinguishes observed behaviour from interpretation.

Label every material statement. Use VERIFIED only when a current attributable source supports the statement and permitted use. Use OBSERVED for an event or participant statement without turning it into a population claim. Use INFERRED where evidence supports but does not prove an explanation. Use PROPOSED for a future design or decision. Use UNKNOWN when evidence is missing. Use REQUIRES VALIDATION where testing is required before investment, release, scale or public copy.

Test the three competing explanations in the authoring profile: The generated map contains plausible detail not present in the supplied evidence. Sensitive or confidential inputs are used without a permitted purpose or controlled environment. Reviewers approve a polished output without tracing claims back to sources and contrary cases. Record confirming and contrary cases with equal care. A report, metric, release note, map, generated synthesis or consensus can indicate where to investigate, but none proves why a control works or why a journey improved.

The research frame should cover the whole user problem and the routes people actually use. NIST Artificial Intelligence Risk Management Framework and GOV.UK Service Manual: User research support end-to-end research and joined-up channel thinking; GOV.UK Service Manual: Map a user's whole problem supports a clear user-needs frame where relevant. NIST Artificial Intelligence Risk Management Framework gives a primary measurement reference. Preserve interfaces across teams and dependencies instead of hiding them in one report.

Accessibility is part of the journey outcome and acceptance test, not a final compliance tick. The Web Content Accessibility Guidelines 2.2 include requirements relevant to complete processes, errors, interruptions, input assistance, redundant entry and accessible authentication. Test with disabled users and people who need assistance where the decision may affect them. Do not infer universal access from automation alone.

Personal information must follow the organisation's approved privacy and research process. POPIA, the Protection of Personal Information Act (https://www.gov.za/documents/protection-personal-information-act), is the primary South African statutory source. This guide is not legal advice; the Information Officer or qualified adviser retains legal authority. Record purpose, permitted use, access, retention and deletion arrangements before collecting identifiable evidence. Minimise workbook content and use controlled references.

Stop and re-charter when the decision is unnamed, the population cannot be bounded, the outcome is only a proxy, evidence cannot be used lawfully, assisted or excluded users cannot be represented, the owner lacks authority, or the work crosses the scope boundary. Record UNKNOWN when evidence is missing.

How to fix it step by step

Step 1: Define the journey decision

Open Start here and Journey charter. Write the decision as a choice an authorised owner must make, not an aspiration such as improve governance. Define the user outcome, the decision consequence and what evidence would permit action. Begin by testing source provenance; do not let a familiar report, release status or map stand in for the governed outcome.

For source provenance, record the affected user or business condition, source, date, owner, limitation and consequence. Test this explanation against successful and contrary routes: The generated map contains plausible detail not present in the supplied evidence. Keep prompt and output boundary separate from proxy measures and status labels. Mark incomplete evidence UNKNOWN or REQUIRES VALIDATION and name the next observation needed.

Checkpoint 1: Use E01 to answer “What evidence establishes source provenance?” For source provenance, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision.

Step 2: Collect and label the evidence

Create evidence IDs before interpretation. In Evidence register, record source, date, population, route, context, limitation, confidence, lineage and permitted use. Reconcile research, operational, support, product, analytics and commercial evidence without pasting personal data. Collect successful, assisted, interrupted, failed and contrary cases; a clean report cannot prove that governance is working.

For prompt and output boundary, record the affected user or business condition, source, date, owner, limitation and consequence. Test this explanation against successful and contrary routes: Sensitive or confidential inputs are used without a permitted purpose or controlled environment. Keep uncertainty separate from proxy measures and status labels. Mark incomplete evidence UNKNOWN or REQUIRES VALIDATION and name the next observation needed.

Checkpoint 2: Use E02 to answer “What evidence establishes prompt and output boundary?” For prompt and output boundary, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision.

Step 3: Set the population, routes and observation window

Use Journey charter to set population, routes, start, end, exclusions, observation window, review authority and decision boundary. For this guide, make prompt and output boundary visible rather than letting a dashboard, workflow state or launch date stand in for user progress. Keep unknown prevalence, unknown causality and unobserved routes explicit.

For uncertainty, record the affected user or business condition, source, date, owner, limitation and consequence. Test this explanation against successful and contrary routes: Reviewers approve a polished output without tracing claims back to sources and contrary cases. Keep consent and minimisation separate from proxy measures and status labels. Mark incomplete evidence UNKNOWN or REQUIRES VALIDATION and name the next observation needed.

Checkpoint 3: Use E03 to answer “What evidence establishes uncertainty?” For uncertainty, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision.

Step 4: Map tasks, handoffs and possible causes

In Journey map, map user tasks and context first, then add channels, systems, handoffs, measures, owners and unresolved questions. Connect each transition to evidence IDs and show where uncertainty changes the user's state or the decision's reliability. In Friction and failure log, describe symptoms separately from causes and record competing explanations.

For consent and minimisation, record the affected user or business condition, source, date, owner, limitation and consequence. Test this explanation against successful and contrary routes: The generated map contains plausible detail not present in the supplied evidence. Keep human review separate from proxy measures and status labels. Mark incomplete evidence UNKNOWN or REQUIRES VALIDATION and name the next observation needed.

Checkpoint 4: Use E04 to answer “What evidence establishes consent and minimisation?” For consent and minimisation, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision.

Step 5: Check the business consequences and competing causes

Connect the issue to value, harm, cost, delay, quality, trust, retention, capacity or decision risk only where a defensible source exists. If size or causality is unknown, label it UNKNOWN or REQUIRES VALIDATION. Test whether consent and minimisation changes the decision. Do not treat a trend, definition, AI output or stakeholder consensus as proof.

For human review, record the affected user or business condition, source, date, owner, limitation and consequence. Test this explanation against successful and contrary routes: Sensitive or confidential inputs are used without a permitted purpose or controlled environment. Keep source provenance separate from proxy measures and status labels. Mark incomplete evidence UNKNOWN or REQUIRES VALIDATION and name the next observation needed.

Checkpoint 5: Use E05 to answer “What evidence establishes human review?” For human review, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision.

Step 6: Name owners, escalation and recovery routes

In Intervention and acceptance, name the accountable owner, contributors, dependency owners, decision rights, exception route and recovery owner. State what receiving teams must accept and what qualified privacy, accessibility, security or technical authorities must decide. A role without evidence access or power to pause, reject or reverse is not sufficient governance.

Step 7: Agree acceptance, guardrails and rollback tests

Write acceptance before implementation: A human reviewer can trace each material AI-assisted statement to authorised evidence or label it PROPOSED or REQUIRES VALIDATION, while consent, minimisation, prohibited inputs, uncertainty, dissent and approval are recorded. Add accessibility, privacy, value, harm, support, quality, inclusion, trade-off and rollback guardrails. Use Outcome and recovery to record tests over a stated window. The negative case is: An AI map invents a common user emotion and a sensitive user attribute, then the team uses both to prioritise an intervention without research or lawful authority.

Step 8: Record the decision and review triggers

Record the human decision and version together. In Review and change log, state proceed, adapt, pause, stop or investigate, approver, effective date, next review, evidence change and reopen or retirement trigger. Approval permits action; it does not turn PROPOSED work into VERIFIED evidence. Observe the result, preserve dissent and capture learning.

Worked example: A hypothetical team asks an AI system to turn support notes into a journey map. The reviewer removes identifiers, supplies only coded summaries, rejects unsupported emotional claims and records questions for human research before the map can guide investment. This is a hypothetical teaching example, not an Abrasim client result and not evidence of a guaranteed outcome. Its purpose is to show how a bounded governance decision, competing explanations, responsible ownership, recovery, value, harm, trade-off and acceptance tests fit together.

How to verify the journey change

Verification asks whether the governance or learning decision improved the user outcome and control quality under stated conditions, not whether a map or workbook was completed. Check source provenance, prompt and output boundary, uncertainty, consent and minimisation, human review. Test this possible cause: the generated map contains plausible detail not present in the supplied evidence. A generic health signal cannot establish whether it is true. Keep the work within this scope: Evidence-bounded journey synthesis. AI output remains hypothesis material and does not create personas, research evidence or approval. Accept the work only when the following condition is met: A human reviewer can trace each material AI-assisted statement to authorised evidence or label it PROPOSED or REQUIRES VALIDATION, while consent, minimisation, prohibited inputs, uncertainty, dissent and approval are recorded. Challenge this contrary case: An AI map invents a common user emotion and a sensitive user attribute, then the team uses both to prioritise an intervention without research or lawful authority. Review the evidence chain in this order:

1. Recheck population, routes, start, end, exclusions and observation window in Journey charter.
2. Sample IDs from Journey map back to Evidence register, source lineage and permitted-use limits.
3. Confirm severe items in Friction and failure log have an owner, disposition, escalation and recovery route.
4. Execute the acceptance, accessibility, privacy, value, outcome, harm, support, inclusion, negative-case, trade-off and rollback tests in Intervention and acceptance and Outcome and recovery.
5. Observe progression, quality, first value, assistance, recovery, exceptions, support demand and adverse outcomes for the agreed window.
6. Reconcile route-level and contrary outcomes; do not use an aggregate to hide material variation or definition drift.
7. Ask the authorised owner to decide proceed, adapt, pause, stop or investigate.
8. Record evidence changes, decision, approver, review date and reopen or retirement trigger in Review and change log.

Use evidence appropriate to the decision. For source provenance, reconcile research, operational, analytics, reports, support and commercial sources with their population, period and limitation. Check prompt and output boundary, uncertainty and consent and minimisation and human review against the agreed acceptance conditions. Verification is incomplete when only the easiest route was tested, personal data was used beyond purpose, accessibility was checked only by automation, downstream ownership was absent, or the window ends before value or learning can be observed.

Use this evidence with AI safely. AI may organise authorised, minimised evidence, surface contradictions and draft questions. It must not invent users, prevalence, outcomes or legal conclusions, infer sensitive attributes, make the final decision or receive personal or confidential data outside an approved environment. The NIST AI Risk Management Framework (https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10) is a primary risk reference, not a substitute for local authority. Remove identifiers and use controlled references. Use this prompt:

Role: Journey governance and evidence reviewer. Decision: Use AI for structured synthesis while preserving sources, uncertainty, consent and human authority. Scope: Evidence-bounded journey synthesis; AI output remains hypothesis material and does not create personas, research evidence or approval. Focus: source provenance; prompt and output boundary; uncertainty; consent and minimisation; human review. Evidence questions: E01: What evidence establishes source provenance | E02: What evidence establishes prompt and output boundary | E03: What evidence establishes uncertainty | E04: What evidence establishes consent and minimisation | E05: What evidence establishes human review. Competing explanations: The generated map contains plausible detail not present in the supplied evidence | Sensitive or confidential inputs are used without a permitted purpose or controlled environment | Reviewers approve a polished output without tracing claims back to sources and contrary cases. Acceptance and guardrails: A human reviewer can trace each material AI-assisted statement to authorised evidence or label it PROPOSED or REQUIRES VALIDATION, while consent, minimisation, prohibited inputs, uncertainty, dissent and approval are recorded. Use only authorised, minimised summaries with evidence IDs and labels. Separate VERIFIED, OBSERVED, INFERRED, PROPOSED, UNKNOWN and REQUIRES VALIDATION; cite the supplied ID for every material statement. Identify contradictions, missing routes, definition drift, unsupported causal claims and the smallest next observation. Do not invent users, prevalence, outcomes or legal conclusions. Return an evidence table, risks, questions, decision options, rollback and human approvals required.

Review AI output against source records. If it cannot cite the supplied evidence ID, mark the statement REQUIRES VALIDATION. Human owners retain responsibility for research, privacy, accessibility, commercial, operational, release, governance and recovery decisions.

How to keep the problem from returning

Use three standing controls:

• Provide AI only minimised, authorised summaries with stable source IDs.
• Keep generated hypotheses separate from VERIFIED and OBSERVED evidence.
• Require human research, privacy, accessibility and decision owners before use.

Review after a material offer, channel, system, policy, dependency, operating-model, definition, release, population or AI workflow change; after recurring exceptions; when evidence becomes stale; or when acceptance or guardrails change. Update evidence, decision and change log together. Assign a review owner even when several teams maintain the journey. That owner must convene contributors, preserve dissent, access permitted evidence and obtain an authorised decision.

Continue with the preceding guide when earlier governance or learning evidence is unresolved, then use the next guide when that is the next question you need to answer. Use ai generated personas not customer research when the adjacent problem owns the wider diagnosis: The persona guide owns AI-generated people models; this guide owns AI-assisted synthesis of journey evidence and unknowns.

If the issue crosses teams, budgets, systems or evidence sources and cannot be resolved safely within the guide, Govern the AI-assisted journey workflow. An assessment should first agree the decision, evidence access and authority. Technology choices follow the findings.

Primary sources used

Source cutoff: 2 September 2026. Review by 2 March 2027, or sooner after a material change. These dates apply to the external references below; local organisational evidence still needs its own date and limitation.

• NIST Artificial Intelligence Risk Management Framework
• GOV.UK Service Manual: User research
• GOV.UK Service Manual: Map a user's whole problem
• W3C Web Content Accessibility Guidelines 2.2
• South African Government: Protection of Personal Information Act
• NIST Artificial Intelligence Risk Management Framework Checkpoint 6: Use E01 to answer “What evidence establishes source provenance?” For source provenance, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision. Checkpoint 7: Use E02 to answer “What evidence establishes prompt and output boundary?” For prompt and output boundary, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision. Checkpoint 8: Use E03 to answer “What evidence establishes uncertainty?” For uncertainty, link the controlled source, owner, evidence status and next action. Preserve contrary findings and reopen the charter if they change the decision.