SVS Digital

AI process reviews

Build a process exception register before an AI pilot

A process map shows the usual route, but an AI pilot often fails in the side paths: an order without a customer number, a request that matches two categories, a document in an unexpected format or a connected system that is unavailable. A process exception register turns those cases into design decisions instead of leaving them to be discovered after launch.

By SVS Digital · Published

Start here

Choose one bounded workflow and review representative completed cases. Record each meaningful exception with an anonymised case reference, trigger, expected path, actual variation, likely cause, consequence, current workaround, evidence, owner and observed count. Then assign one treatment: repair the input, add a deterministic rule, let AI assist under review, send directly to a person, or keep the case out of scope. Convert the approved exceptions into pilot test cases and pause conditions before connecting live actions.

1. Define an exception without treating every variation as a defect

Start with one case type, such as a service enquiry, supplier invoice or order change. Write its expected start, end and normal path. An exception is a case that needs different information, responsibility, approval or recovery from that normal path. A legitimate express order may be a valid variant; a skipped approval, repeated correction or unknown route may be a defect. Keep those meanings separate.

Microsoft defines a process variant as cases that follow the same sequence of activities and recommends examining variants to find frequent and problematic scenarios. UiPath's conformance guidance distinguishes retries, skipped steps, invalid activities, unexpected starts and loop-backs. These tool categories are useful prompts, but the process owner still has to decide what is acceptable in the business context.

  • One case type with a clear trigger and end state.
  • Normal path confirmed by the people who do and receive the work.
  • Valid business variants kept separate from defects and failures.
  • Rare high-consequence cases retained even when they are not frequent.
Microsoft Learn: Analyze processes with variantsUiPath: Conformance checking

2. Build the register from real cases, not workshop memory

Review a representative sample with the employees involved. For each exception, record a short label, anonymised case reference, date, point in the workflow, expected event, actual event, input condition, likely cause, consequence, current workaround, role that resolved it and evidence source. Record an observed count only for the stated sample and period; do not turn a handful of examples into an annual rate.

Where systems keep event histories, use a small export to check the interviews. Microsoft explains that process mining extracts event data from systems of record to show how work is actually executed, compare paths and investigate inefficiencies. A spreadsheet register is enough for a first review when event data is incomplete. Store only what the review needs: use case IDs or redacted examples instead of copying message bodies, documents, passwords or unnecessary personal data.

  • Expected path and actual variation.
  • Evidence source and sample period.
  • Likely cause labelled as a hypothesis until confirmed.
  • Operational consequence: delay, rework, wrong output, missed action or compliance concern.
  • Current owner and workaround without unnecessary sensitive data.
Microsoft Learn: Overview of process miningMicrosoft Learn: Analyze processes with variants

3. Classify cause and consequence before discussing AI

Use a small, practical taxonomy. Input exceptions include missing, conflicting or unreadable information. Process exceptions include a skipped approval, wrong owner or work returning to an earlier step. System exceptions include timeouts, duplicate records and unavailable integrations. Decision exceptions include ambiguity or a case outside the agreed policy. AI-specific exceptions include unsupported output, insufficient evidence or a result that a reviewer cannot verify.

Then score consequence separately from frequency. Ask what happens if the case is processed incorrectly: a minor correction, customer delay, financial loss, unauthorised commitment, safety concern or effect on a person's rights. UiPath compares conforming and non-conforming cases and shows that deviations can differ in throughput and cost. Your register should similarly preserve impact and context rather than rank exceptions only by count.

  • Cause category: input, process, system, decision or AI output.
  • Consequence described in business terms.
  • Detectability: who can notice the problem and from which evidence?
  • Reversibility: can the action be corrected before it affects another system or person?
  • Required specialist review flagged for legal, privacy, security, safety or sector questions.
UiPath: Conformance checking

4. Give every in-scope exception a handling route

Choose the simplest treatment that addresses the cause. A required field may prevent missing information. A deterministic rule may route a known code. AI may help extract, classify or draft when inputs vary, provided an authorised person can verify important outputs. Ambiguous, prohibited or high-consequence cases may need immediate human handling. A system outage needs a queue, retry rule and reconciliation, not a more persuasive model response.

Write the route as an operational rule: detection signal, permitted action, receiving role, information shown, response time, fallback and final record. NIST's Manage playbook says AI may not be the right solution for a task, calls for risks to be prioritised by likelihood and consequence, and recognises mitigation, transfer, avoidance or acceptance as possible responses. The register should therefore support a no-AI or no-go decision as readily as an AI pilot.

  • Prevent at source, deterministic rule, AI with review, direct human route or out of scope.
  • Named owner with authority and the source information needed to decide.
  • No customer message, booking, payment or record change on an uncertain result unless separately approved and tested.
  • Fallback for unavailable models, integrations and staff.
  • Reconciliation step for actions that may fail or be duplicated.
NIST AI RMF Playbook: Manage

5. Turn exception rows into acceptance tests and stop conditions

Select ordinary cases, frequent exceptions and rare cases with serious consequences. For each test, define the input, expected route, acceptable output, evidence a reviewer must see, prohibited action and final system state. Test missing fields, conflicting instructions, duplicate events, unavailable systems and cases that should be rejected. Keep test records clearly labelled and isolated from production where possible.

Do not hide a critical failure inside an average score. Set separate acceptance checks for correct routing, output quality, reviewer effort, integration state and recovery. A pilot should pause when it performs a prohibited action, loses a case, creates an untraceable duplicate or repeatedly sends uncertain work down the normal path. NIST recommends evaluating performance in the real use context, documenting risk responses and monitoring whether the system still achieves its intended purpose.

  • A test case for every approved exception route.
  • Expected business state, not only expected model wording.
  • Critical failures and the person authorised to pause the pilot.
  • Test evidence retained without unnecessary personal data.
  • Retest required after rule, prompt, model, integration or source-data changes.
NIST AI RMF Playbook: Manage

6. Keep the exception register alive during the bounded pilot

During the pilot, log new exception types, routing mistakes, overrides, recovery time and cases that employees handle outside the designed path. Review the register on a fixed cadence with the process owner and users. Merge duplicate labels, confirm causes with evidence and update a rule only through the pilot's change process. A growing register is not automatically failure; it may show that the team is finally seeing the real workflow.

At the review, decide which exceptions are prevented, handled reliably, still require a person or make the use case unsuitable. Compare only with the documented sample and baseline. This guide does not replace a data-protection, employment, security, legal, safety or sector-specific assessment, and an exception log does not prove that an AI system is compliant or ready for production. Its purpose is to make scope, tests, ownership and limits explicit before a larger investment.

  • Process owner and review cadence.
  • New exception, override and recovery fields included in pilot logging.
  • Changes versioned and retested before wider use.
  • End decision: proceed, narrow, repair, collect more evidence or stop.

Sources and further reading

More guides

What would you like to build for your business?

Tell us where you need a hand. We’ll email you a brief and follow up about the next step.

Uroš Jovanović

Founder · SVS Digital

Hi, I’m Uroš.

I founded SVS Digital. We design and build custom digital solutions for businesses.

What kind of solution do you need?

Your reply