01

Map the task, inputs, outputs and tools

Name one specific task and its purpose. Record who starts it, each main step, systems used, input formats, expected output and recipient. Note handoffs, approvals, integrations and any unofficial spreadsheets or workarounds. Keep the boundary narrow enough to assess independently.

02

Record bottlenecks and establish a baseline

Track a representative sample of current work. Capture volume, elapsed time, active effort, queue delays, corrections, rework and escalation frequency. Separate observed measurements from estimates. Ask where work stalls, which errors recur and what acceptable performance looks like before considering a different approach.

03

Compare rules-only and AI-assisted options

List steps that follow stable, explicit rules and test whether templates, validation, workflow automation or system configuration could address them. Then identify steps involving interpretation, classification, extraction or drafting. Compare options using the same criteria: reliability, maintenance, review effort, exception handling and operational fit.

04

Check data permissions and required human decisions

Inventory the information the task uses, including its source, owner, sensitivity, permitted uses, retention expectations and access controls. Mark data that is incomplete or inconsistent. Identify decisions that must remain with an authorised person, who reviews AI-assisted output and what evidence that reviewer needs.

05

Test normal, difficult and exceptional examples

Assemble representative completed cases without exposing data unnecessarily. Include routine work, ambiguous inputs, missing fields, unusual formats and high-impact exceptions. Define expected results and unacceptable outcomes before testing. Fictional example: an invoice-coding task includes standard invoices, blurred scans, duplicate references and an unfamiliar supplier category.

06

Choose simplification, cleanup, pilot or deferral

Review the evidence and select one next step. Simplify if the process itself causes friction; clean data if inputs are unreliable; pilot only when the task is bounded, measurable and reviewable; defer when permissions, ownership or safeguards are unresolved. Record the decision, rationale, assumptions and conditions for reconsideration.