AI process reviews
Build a task inventory before an AI process review
A workshop often begins with the same claims: email takes too long, reporting is manual and someone should add AI. Those claims may be true, but they do not show which step repeats, what a good result looks like or where judgement is essential. A short task inventory gives employees and managers a shared view of real work before they compare AI, ordinary automation or a process change.
By SVS Digital · Published
Start here
Choose one team and a representative work period. For each recurring task, record the trigger, expected output, frequency, hands-on time, waiting, systems, data, common exceptions, reviewer and consequence of an error. Use role names and anonymised examples, not copied customer or employee data. Group the entries into workflows, validate them with the people doing the work, and take only two or three evidence-backed candidates into the AI process review.
1. Define the question before collecting tasks
Do not ask everyone to list everything they do. Pick a boundary such as sales administration, invoice intake or service enquiries, name the manager who owns the review and state the question: where does this team repeat work, wait, correct information or make the same judgement? Include the start and end of the selected workflow so a faster step does not merely move work to the next person.
The UK Government's AI Knowledge Hub recommends breaking work into workflows, tasks and steps before deciding where AI can help. NIST's AI Risk Management Framework likewise starts by mapping the context, intended purpose, users and affected people; it says that this context should support an initial decision about whether an AI solution is appropriate at all. The inventory is therefore evidence for a decision, not a promise to automate.
- One team or workflow boundary, not the whole company.
- A named process owner and participating roles.
- A representative period that includes ordinary work.
- A separate note for month-end, seasonal or rare critical cases.
2. Capture one row per task without turning it into surveillance
Use a small table. Record the event that starts the task, the result that marks it complete, how often it occurs, approximate hands-on time, typical waiting, applications used, input and output, common exceptions, who checks it and where it goes next. Add one short note about why the task is difficult: missing information, retyping, searching, inconsistent rules, a queue or a decision that needs experience.
Ask employees to record work as participants in the review, not as monitored subjects. Use role names instead of personal performance comparisons and explain who can see the inventory. If screen or task recording is considered, agree the purpose and access first. Microsoft notes that process recordings can expose information to owners and contributors and provides controls to pause recording and remove sensitive steps or personal data. A plain manual log is often enough for the first review.
- Trigger and finished output.
- Frequency, hands-on time and waiting kept separate.
- Systems, input data and next recipient.
- Exception, correction and review path.
- No passwords, message bodies or unnecessary personal data.
3. Sample enough variation to see the real task
A practical starting point is five consecutive working days, or another short period that represents the team's normal load. This is a sample, not a precise time-and-motion study. If the workflow changes at month-end, during a campaign or when a supplier file fails, add examples from those conditions rather than pretending the ordinary week covers them.
Do not merge every entry into an average too early. Keep the routine path, frequent exception and rare high-consequence case visible. Microsoft describes task mining as a way to understand what employees actually do and identify common mistakes and automation opportunities; its preparation guidance also recommends looking for repeated patterns, frequency and dependencies, then checking the grouped activities with employees and supervisors. The same discipline improves a manual inventory.
- At least two people where a role has different working patterns.
- Ordinary cases plus frequent exceptions.
- Waiting caused by another person or system marked explicitly.
- Uncertain estimates labelled as estimates.
- Critical but rare work reviewed separately.
4. Turn task rows into workflow evidence
Group rows by the result they help produce, not by the application that was open. Copying order details from email, checking a customer number in the ERP and correcting a spreadsheet may be three steps in one order-entry workflow. Draw the trigger, ordinary path, handoffs, waiting points, exceptions and final output. Ask the people who send and receive the work to confirm the map separately.
Then calculate only what the sample can support: observed occurrences, approximate hands-on effort, waiting, rework and exception count. High volume can matter, but so can a lower-volume task with repeated errors or a long queue. Keep estimates in ranges and record the source. The aim is to locate a problem worth investigating, not to manufacture an annual saving from a few diary entries.
- Tasks grouped around one business result.
- Handoffs and queues visible.
- Manual copying and repeated searches marked.
- Exception paths retained instead of averaged away.
- Observed facts separated from assumptions.
5. Screen candidates for AI, ordinary automation or process repair
For each problem, ask what kind of change fits. A required form field may fix missing input. A deterministic integration may remove copying between systems. AI may assist when inputs vary and the task involves extraction, classification, summarisation or drafting, provided a person can review important outputs. The UK AI Knowledge Hub says automation is better suited to common, repetitive, well-defined tasks with clear rules, while judgement-heavy work needs human control and may be better augmented than automated.
Reject or pause a candidate when the expected output is unclear, examples cannot be used lawfully, important errors would be hard to detect, the reviewer lacks time or authority, or the process changes every time. NIST treats mapping as support for a go or no-go decision and also asks organisations to consider viable non-AI approaches. A process review succeeds when it finds the right intervention, even when that intervention is not AI.
- Clear input, output and owner.
- Enough representative examples to test.
- A reviewer who can detect and correct failure.
- An ordinary rule, integration or process change considered.
- Legal, security or domain review flagged where needed.
6. Bring a compact evidence pack to the process review
Finish with four items: the task inventory, one current-state workflow map, two or three candidate cards and a list of open questions. Each candidate card should state the problem, affected role, observed frequency or delay, example inputs and outputs, exceptions, data boundary, possible change and the person responsible for the result. Include a candidate that may need only a process fix so the review does not begin with a predetermined technology.
The next step is discovery, not procurement. Validate the sample, inspect permitted examples, compare tool and integration options, and define a bounded pilot only when the task is suitable. This method does not replace a data-protection, employment, security, legal or sector-specific assessment. It reduces guesswork and gives an AI consultant or internal team a concrete basis for deciding what to investigate next.
- Inventory with sources and confidence labels.
- Current workflow including exceptions and waiting.
- Two or three candidates, not a long wish list.
- Open questions and required specialist checks.
- A decision owner for the next step.

