How to identify AI automation opportunities

Start with the actual work

Trace a real workflow and separate information handling from decisions requiring judgement.

Find repeated work before choosing an AI tool

Choose a workflow that occurs often enough to examine and has a clear accountable owner. Write down the input, required output, source of truth and next business action. Repeated copying of accepted order details may be a candidate; deciding whether to accept an unusual order is a different problem.

Observe the work as it happens. Include time spent correcting inputs, checking output and recovering from mistakes. An impressive demonstration of generated text does not establish that the whole workflow will become easier or more reliable.

Separate assistance from authority to act

Mark the point where information becomes a business commitment: sending a customer message, changing a record, booking capacity or approving payment. A system that drafts an answer and a system that sends it have different consequences. Agree which actions, if any, are authorised before implementation.

For a first investigation, consider a bounded proposal in which an appropriate person reviews output before use. Human review still needs clear criteria, sufficient time and access to the source. A nominal approval button is not evidence that mistakes will be detected.

Compare failure cost with review effort

Collect representative examples, including incomplete inputs and exceptions. Define what acceptable output means and record errors by type and consequence. Look at missed details, invented facts, duplicated actions and cases the system should decline or escalate.

Evaluate the review burden alongside speed. If an operator must reconstruct every source record to trust the output, the proposed automation may simply move the work. Avoid putting live sensitive information into a provider until the relevant data, permission and supplier checks have been completed.

Design the pilot and the route back

Specify the workflow owner, permitted inputs, action limits, success evidence and stop conditions. Keep a practical manual route available. Decide how to identify failed or incomplete runs and who will resolve them without creating duplicate customer commitments.

Begin only after the relevant technical and organisational controls have been agreed. Compare suitable cases with the starting condition and account for review, error and recovery effort. The free AI Automation Audit helps organise the opportunity and control questions; it does not guarantee a saving, compliance outcome or implementation result.

What the example does not establish

Automation is a candidate to investigate, not an assumed saving.

Look for a case that challenges your first explanation as well as one that supports it. Agree which part of the situation is observed, which is an assumption and which needs another person’s evidence. Keep any proposed change within the authority and commitments already in place.

How do you know whether a change helped?

Record the starting condition and a suitable review point before the change. Compare real cases with similar constraints, including exceptions and any extra work created elsewhere. Separate a prepared document, an implemented change and a tested operating result. If the evidence is incomplete, name the gap rather than treating it as success.