Source: https://www.svs-digital.com/guides/map-decision-rights-before-ai-pilot
Language: en
Content updated: 2026-11-19
Description: Review one real workflow decision by decision, then define what AI may retrieve, draft, recommend, route or execute and who remains accountable.

AI process reviews

# Map decision rights before an AI pilot

An AI pilot changes who sees evidence, proposes an outcome, approves an action and recovers when a result is wrong. A decision-rights map makes those boundaries explicit before prompts, integrations or model choices distract the team.

By SVS Digital · Published 14.11.2026

## Start here

Choose one bounded workflow and mark every point where work can change route, create a commitment, update a record or affect a person. Classify the AI role as information retrieval, drafting, recommendation, routing or execution. For each point, record the human decision owner, evidence, consequence of error, reversibility, approval or override rule, fallback and audit record. Start with the least authority needed. Permit execution only when the action, limits, recovery and accountable owner are explicit and testable.

## 1. Trace one real workflow and mark its decision points

Set a clear trigger and end state, such as receiving a supplier invoice and posting an approved entry. Walk through completed cases with the people who perform, approve and receive the work. Mark every question that changes the next step: Is the supplier known? Does the amount need approval? Which cost centre applies? Map decisions inside one case type, not every departmental task.

Microsoft explains that process mining uses event data from systems of record to visualise actual execution. Process maps display activities, transitions and measures such as frequency, duration and rework. They show what happened, but not who should hold authority or whether a decision is acceptable. Business owners must interpret the evidence and define the intended process.

- One trigger, one end state and one named process owner.

- Ordinary, ambiguous and high-consequence cases represented.

- Decision points separated from data entry and routine handoffs.

[Microsoft Learn: Overview of process mining](https://learn.microsoft.com/en-us/power-automate/process-mining-overview)[Microsoft Learn: Process map overview](https://learn.microsoft.com/en-us/power-automate/minit/process-map)

## 2. Classify the AI role before assigning authority

Use five labels. Retrieval finds information without changing it. Drafting prepares text or fields for review. Recommendation proposes an outcome. Routing chooses the next queue or reviewer. Execution changes a system, sends a message, commits money or grants access. Classify each decision point rather than calling the whole pilot an assistant or agent.

A retrieved policy can be incomplete; a draft can misstate facts; a recommendation can anchor a reviewer; wrong routing can hide urgent work; execution can create an external effect. NIST's Map playbook calls for the task, intended use, deployment context and human oversight to be documented. It also says non-AI alternatives should be considered. A required field or fixed rule may be the better control.

- One role label for each proposed AI contribution.

- Input, output and permitted downstream use stated plainly.

- Non-AI option considered before model judgement.

[NIST AI RMF Playbook: Map](https://airc.nist.gov/airmf-resources/playbook/map/)

## Need to define decision rights before an AI pilot?

Bring one recurring workflow, a few anonymised completed cases and the roles involved. SVS Digital can help map the decisions, set human and AI boundaries and define a testable pilot scope.

## 3. Complete one decision-rights row for every marked point

Create one row per decision. Record the question, AI role, accountable decision owner, required evidence, consequence, reversibility window, approval rule, override authority, fallback and audit record. The owner must have authority to accept the consequence, not merely be the person who clicks approve. Name the evidence source, date and any missing or conflicting information.

Describe consequence in operational terms: delay, incorrect message, unauthorised payment, changed access, regulatory concern or effect on rights. State whether the result can be corrected before another person or system relies on it. High consequence, weak evidence or poor reversibility should narrow the AI role, require specialist approval or keep the decision outside the pilot.

- Decision owner has authority, time and relevant domain knowledge.

- Required evidence is visible at the moment of approval.

- Consequence and reversibility are assessed separately.

- Approval, override, fallback and audit rules are operational.

## 4. Design approval, override and fallback as operating rules

Define which cases require review, what the reviewer sees, what counts as approval, what happens after rejection and how long work may wait. An override should require a reason and authorised role. An approve button is not meaningful oversight if reviewers lack time, training, source evidence or authority to challenge the output.

NIST's Govern playbook calls for clear roles, responsibilities, communication and accountability supported by training. Differentiate human and AI roles: AI may propose a route, an operations lead may approve routine cases, and a specialist may decide exceptions. If the model, integration or reviewer is unavailable, use a manual queue, deterministic rule or pause that preserves cases and prevents silent action.

- Review threshold and receiving role.

- Evidence displayed, response window and rejection path.

- Override reason, authorised role and later review.

- Fallback covers model, system or staffing failure.

[NIST AI RMF Playbook: Govern](https://airc.nist.gov/airmf-resources/playbook/govern/)

## 5. Specify the audit record before live actions

For each decision, define the minimum record needed to reconstruct what happened: case reference, time, model or rule version, input sources, proposed output, uncertainty signal if used, reviewer action, override reason, final state and recovery. Use links or hashes where suitable instead of copying unnecessary personal or confidential content into a new log.

The audit record should answer who decided, from which evidence, under which rule and with what result. It is not proof that a decision was correct, fair or compliant. Access, retention and monitoring need their own controls. Ask privacy, security, legal, employment, safety or sector specialists where the workflow affects protected data, rights or regulated outcomes.

- Record supports reconstruction without collecting unnecessary data.

- Final action and downstream system state are captured.

- Access and retention owners are named.

- Missing logs or mismatched states trigger reconciliation or pause.

## 6. Approve the pilot boundary decision by decision

Review the map in sequence. Confirm the intended use and non-AI option, then the AI role, owner and evidence. Assess consequence and reversibility. Test approval, override, fallback and audit behaviour with ordinary and difficult cases. Grant only the authority supported by evidence. A pilot may retrieve and draft while recommendation, routing or execution remain out of scope.

Set stop conditions for prohibited action, missing evidence, lost cases, untraceable overrides, failed recovery or repeated reviewer disagreement. Revisit the map when the workflow, model, prompt, data source, policy or integration changes. This review does not replace a formal risk, data-protection, legal, security or sector assessment. Its purpose is to make operational authority visible before the pilot changes real work.

- Each decision is approved, narrowed, deferred or kept human-only.

- Test cases cover missing, conflicting and outdated evidence.

- A named person can pause the pilot.

- Execution requires fresh review of consequence and recovery.

## Need to define decision rights before an AI pilot?

Bring one recurring workflow, a few anonymised completed cases and the roles involved. SVS Digital can help map the decisions, set human and AI boundaries and define a testable pilot scope.

## Sources and further reading

- [NIST AI RMF Playbook: Map](https://airc.nist.gov/airmf-resources/playbook/map/)

- [NIST AI RMF Playbook: Govern](https://airc.nist.gov/airmf-resources/playbook/govern/)

- [Microsoft Learn: Overview of process mining](https://learn.microsoft.com/en-us/power-automate/process-mining-overview)

- [Microsoft Learn: Process map overview](https://learn.microsoft.com/en-us/power-automate/minit/process-map)

## More guides

AI process reviews

## [Build a process exception register before an AI pilot](https://www.svs-digital.com/guides/process-exception-register-before-ai-pilot)

Record real process exceptions, their causes, consequences, owners and handling rules before deciding what an AI workflow may process, escalate or reject.

10.11.2026

AI process reviews

## [Build a task inventory before an AI process review](https://www.svs-digital.com/guides/task-inventory-before-ai-process-review)

Collect a representative sample of recurring work, delays, exceptions and data boundaries so an AI process review starts with evidence rather than guesses.

05.11.2026

AI adoption and team workflows

## [AI meeting notes: review decisions before creating tasks](https://www.svs-digital.com/guides/ai-meeting-notes-review-action-items-workflow)

Set practical rules for AI meeting notes: eligible meetings, participant notice, access, retention, human review and a reliable handoff from approved action items to owned tasks.

13.11.2026

---

SVS Digital · Dunajska cesta 5, 1000 Ljubljana, Slovenia
Contact: [uros@svs-digital.com](mailto:uros@svs-digital.com)
[All resources](https://www.svs-digital.com/ai)
