Source: https://www.svs-digital.com/guides/map-process-data-dependencies-before-ai-pilot
Language: en
Content updated: 2026-11-19
Description: Trace each required input, event, transformation and output to its source, owner, freshness rule and fallback before an AI pilot connects to business systems.

AI process reviews

# Map process data dependencies before an AI pilot

A workflow can look suitable for AI until the team asks where its facts come from. A customer status may be copied across three systems, an approval time may exist only in email, and a spreadsheet may contain a corrected value that never returned to the system of record. A process data-dependency map exposes those gaps before a pilot starts reading or writing live data.

By SVS Digital · Published 19.11.2026

## Start here

Choose one bounded case from trigger to final state. Create one row for every input, event, derived value and output the proposed AI step needs. Record its business purpose, authoritative source, identifier, transformation, freshness, permitted access, receiving step, owner, retention, failure behaviour and evidence. Remove data that is not necessary, resolve conflicting sources, and turn stale, missing, duplicated or unavailable data into acceptance tests and stop conditions before connecting live actions.

## 1. Bound one process and one proposed AI task

Name the item moving through the workflow: one enquiry, order, invoice, service request or approval. Define the trigger, successful end state and current process owner. Then state the proposed AI contribution in one verb phrase, such as classify an enquiry, extract order fields or draft a reply. Do not map every database in the company.

NIST's AI RMF Map playbook says the intended purpose, deployment context, users, assumptions, limitations, human-AI roles and non-AI alternatives should be documented. Apply that discipline before collecting fields. A required form value or deterministic integration may solve the problem without a model, while an unclear outcome means the process is not ready for either.

- One case type with a clear trigger and end state.

- One proposed AI task and its receiving role.

- Current manual or rules-based alternative recorded.

- Process owner and people affected by the output named.

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

## 2. Create one row for each data element and event

Walk through several suitably redacted completed cases with the people who perform and receive the work. Record the data actually used: identifiers, statuses, dates, free text, attached documents, policy references, approval events and final outputs. Keep raw inputs separate from values that are looked up, calculated, inferred or drafted. A label such as customer data is too broad to test.

For each row, write why the element is needed and which step consumes it. If the workflow is analysed from event histories, verify that the source can tie activities to a case and time. Microsoft documents case identifiers, activity names and event or activity timestamps as core process-mining inputs. A workshop table or spreadsheet is useful evidence, but it is not automatically a reliable event log.

- Field or event name stated in business language.

- Raw, looked-up, calculated, inferred and drafted values distinguished.

- Case identifier and relevant timestamp available where sequence matters.

- Purpose and receiving process step recorded for every row.

[Microsoft Learn: Prepare processes and data](https://learn.microsoft.com/en-us/power-automate/process-mining-processes-and-data)

## Need to trace process data before an AI pilot?

Bring one recurring workflow, a few anonymised cases and the systems involved. SVS Digital can help map the data dependencies, identify source gaps and define a bounded, testable next step.

## 3. Trace source, transformation and destination

Name the authoritative source and every copy used in the workflow. Record the lookup key, transformation and destination: for example, a web form creates an enquiry ID, a CRM adds customer status, an employee corrects the service category, and the approved result creates a task. Mark where values can diverge and which system wins when they conflict.

Do not treat an API response as self-explanatory. Record the source field, units, allowed values, timezone, null meaning and update direction. NIST's Manage playbook includes documentation prompts for data provenance, origins, transformations, dependencies, constraints and obsolescence. This map should make a wrong or old value traceable without copying unnecessary content into a new store.

- Authoritative source and any working copies.

- Identifier or matching rule between systems.

- Transformation, units and meaning of empty values.

- Destination, write direction and reconciliation owner.

[NIST AIRC AI RMF Playbook: Manage](https://airc.nist.gov/airmf-resources/playbook/manage)

## 4. Test necessity, access, freshness and retention

Challenge every row. Could the task work without this element, with a less detailed value or with access only at the moment it is needed? Record whether it may contain personal, confidential, contractual or regulated information and which role may see it. Set a freshness rule based on the business decision: an address used for dispatch and a policy used for drafting may have different tolerances.

The ICO's AI audit framework recommends mapping where personal information is used, checking that it is adequate, relevant and limited to the purpose, reviewing retention and identifying unnecessary duplicates. Its page is under review and this guide is not legal advice, but the operational questions remain useful. In Slovenia, Serbia or another market, ask the appropriate privacy, security, legal or sector specialist to confirm applicable duties.

- Necessary, optional or prohibited for this task.

- Access role and permitted environment.

- Freshness threshold and method for detecting stale data.

- Retention or deletion owner and duplicate copies identified.

[ICO: Data minimisation for AI](https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/data-minimisation/)

## 5. Give every dependency a failure behaviour

For each critical row, ask what happens when the value is missing, conflicting, stale, malformed or unavailable. The safe result may be a request for information, a manual queue, a deterministic default, a retry with a limit or a full pause. Do not let a model invent a substitute for an absent customer identifier, approval, price or policy source.

Separate read failure from write failure. A useful draft can still become a lost case if the CRM update times out; a retry can create duplicates if the external action succeeded but the acknowledgement was lost. Define an idempotency key or duplicate check where actions can repeat, keep the pending state visible, and reconcile the destination before declaring completion.

- Missing, conflicting, stale and malformed input routes.

- Unavailable source, model and destination routes.

- Pending, approved, failed and reconciled states kept distinct.

- No consequential action from invented or untraceable evidence.

## 6. Convert the map into pilot tests and ownership

Select ordinary cases and deliberate data failures. For each test, define the source state, expected AI role, evidence shown to the reviewer, permitted output, prohibited action and final system state. Include a duplicate event, stale value, conflicting sources, missing identifier, revoked access and unavailable destination. Verify the whole business state, not only plausible wording from the model.

Finish with a decision for each dependency: ready, repair the source, reduce the data, add a rule, require human review or keep it outside the pilot. Name who owns source quality, access, the AI output and reconciliation. NIST notes that AI may not be the right solution for a task; a data map is successful when it reveals that the process should be repaired before automation as clearly as when it supports a bounded pilot.

- Acceptance test for every critical dependency and failure route.

- Stop condition for missing evidence, lost cases or uncontrolled writes.

- Source, process, review and reconciliation owners named.

- Map reviewed after source, schema, policy, model or integration changes.

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

## Need to trace process data before an AI pilot?

Bring one recurring workflow, a few anonymised cases and the systems involved. SVS Digital can help map the data dependencies, identify source gaps and define a bounded, testable next step.

## Sources and further reading

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

- [NIST AIRC AI RMF Playbook: Manage](https://airc.nist.gov/airmf-resources/playbook/manage)

- [Microsoft Learn: Prepare processes and data](https://learn.microsoft.com/en-us/power-automate/process-mining-processes-and-data)

- [ICO: Data minimisation for AI](https://ico.org.uk/for-organisations/advice-and-services/audits/data-protection-audit-framework/toolkits/artificial-intelligence/data-minimisation/)

## More guides

AI process reviews

## [Map decision rights before an AI pilot](https://www.svs-digital.com/guides/map-decision-rights-before-ai-pilot)

Review one real workflow decision by decision, then define what AI may retrieve, draft, recommend, route or execute and who remains accountable.

14.11.2026

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

---

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)
