AI was entering the support operation from three directions at once. Performance and customer-feedback data was too voluminous to review by hand, so it was not being turned into specific coaching. Some agents were pasting customer questions into an AI chatbot and sending the answer back unchanged, which put filler, and at times inaccurate or off-policy answers, in front of customers. And there was no agreed view on which AI tools support experts could use, with or without customer data.

Data analysis and coaching

Using only approved tooling, and with customer identifiers removed or anonymised, I used AI to group recurring issues into themes by error type, show where coaching was needed, and spot inconsistencies in how cases were assessed. The findings went to the outsourced partner's training team as a short summary, which used them to build coaching content. A person reviewed and approved every item.

A clear rule for customer replies

We set one rule that everyone could follow: AI may be used to search knowledge, on specific approved sources with a specific tool, and never to write any part of a reply to a customer. No customer personal data goes into AI tools that are not approved. Agents were trained on the rule rather than simply told it.

Working with the AI tools steering committee

I worked with a steering committee that identified and evaluated which AI tools support experts could use, in general and with or without customer data. I carried out the review of each tool, kept the list of approved tools, and drove the requests and exceptions.

Impact:

• Coaching became specific to the real error patterns instead of generic, and the partner's training team adopted the approach as a standard way of working.
• Agents were trained on the AI-use rule, which kept unchecked AI-written replies, filler and inaccuracies away from customers.
• A governed route for deciding which AI tools experts could use, with an approved list and a way to handle exceptions.
• Delivered on approved tooling and existing teams, with no new software spend.

This project reflects how I approach AI in operations: AI accelerates drafting and analysis, people keep the judgment, and governance decides where the line sits. Set the line clearly, teach it to the people who use it, and give them a route for exceptions.