Find the work, then find the pressure
Look at intake and triage, transactions, customer cases, quoting and proposals, and knowledge or review work. These are places to investigate, not categories that automatically need AI.
Useful signals include backlogs, slow response, rework, rising cost-to-serve and effort spent moving information between systems. Confirm the pressure is current. An old vendor case may describe a problem management has already solved.
Ask six questions
What consequence would improve? Name the customer, financial, service, capacity or risk outcome. “Save time” is incomplete without identifying whose time and what happens next.
Where is the constraint? Follow the whole process. Automating upstream work may fill a downstream queue faster.
Is the work understood? Clear inputs, rules and exceptions support assessment. Unresolved policy disagreements need management decisions before automation.
What information is required? Check quality, access, currency and appropriate use. Fluent summaries cannot make conflicting records authoritative.
How will quality be judged? Define acceptance and the consequences of error. Reviewer cost and capacity belong in the case.
Who can make the change happen? Identify the workflow owner, implementation capability and benefit owner.
Compare alternatives in a sensible order
First consider whether the work should exist. A report nobody uses may need removal rather than faster production.
Then consider simpler forms, clearer rules, fewer hand-offs or a better approval process. Existing software and conventional automation may solve the remaining task more reliably than AI.
Use AI where handling language, variation or inference creates a defensible advantage within acceptable limits. Retain human judgement where required.
This sequence, remove the work first, then automate what survives, then reallocate the time, is used as a set of questions, not a rule that every human task should disappear.
Three illustrative candidates
Enquiry routing: investigate whether classification is the constraint, how errors affect service and whether receiving teams can handle the volume.
Management reporting: establish which decisions the pack should support before automating its existing slides.
Customer eligibility decisions: repetition alone does not justify automation. Assess affected people, obligations, data, review and specialist controls.
These examples illustrate selection, not client results or deployment recommendations.
Rank by decision value
Compare consequence, evidence, feasibility, dependencies and risk. A simple priority judgement with reasons is more useful than a complex score based on unsupported estimates.
A high-potential but uncertain idea may deserve a small experiment. A moderate-value improvement with strong evidence may be the better first implementation. Some uses need specialist work before any pilot.
For the strongest candidate, define the question to resolve, evidence required, owner and decision at the end. Do not commission a transformation only because one workflow looks promising.