Menu
← Blog

How to choose your first AI pilot use case

Innokenty Bodrov ·

When a team brings me ten tasks it wants to automate, I ask to see a handful of recent cases from the two most promising workflows. Who picked up the task? What did they have to look up? Where did they wait for someone else? Who could say the work was done? A ticket history, a conversation thread and the final reply often tell you more than the process diagram.

For a first pilot, I look for repeatable work with accessible inputs, a result the team can verify, and a chance to correct mistakes before they affect a customer.

One workflow, two different pilots

Imagine a support team handling payment enquiries. It could ask AI to draft replies for a support agent to review and send. It could also give AI responsibility for deciding who qualifies for a refund.

The difference shows up when a customer reports a duplicate charge. One transaction has settled and the other is still pending. A draft can report both statuses for a support agent to verify. A refund decision also requires confirmed policy, clear authority, and a rule for what happens if the transaction status changes after the check.

I would start with drafts. The team can take closed tickets, reconstruct the information available when the agent replied, and review the results without sending anything to customers. If historical payment states cannot be reconstructed, collect fresh examples with the AI running in observation mode. Both approaches need access to the same facts.

Compare the work with and without AI

Measure time to a reply that is ready to send. Include reading the ticket, looking up the status, reviewing the draft and making corrections. If the same agent tries both approaches, use different but comparable cases so they cannot answer the second time from memory.

Pilot decisionSupport example
AI taskDraft a reply when payment status is available
Human decisionConfirm the status, next step and final wording
Test casesRoutine charge, two transactions with different statuses, unavailable data source
MetricTime to a send-ready reply, including review and corrections
Critical errorThe draft promises an unauthorised refund or reports the wrong status
Rollout criterionAgree in advance on the minimum time saved and the acceptable error limit

Suppose a draft brings a routine five-minute reply down to four minutes. On tickets with incomplete data, it takes seven minutes because the agent has to verify the claims again. I would check whether the system can identify tickets with a confirmed status up front. If it can, narrow the pilot to those cases and recalculate the benefit using their share of all incoming tickets.

Discuss a workflow for an AI pilot

Back to top