The situation
A European energy supplier is moving its customers onto a new platform in groups, and every group that goes live creates a stream of problems that the supplier and its platform partner need to sort, route and fix quickly. (The client is unnamed here: the engagement terms restrict it.)
The problem
- Issues were raised in a shared chat channel. They only became tickets if someone remembered.
- Working out what an issue was about meant reading threads 30 to 50 messages long.
- Issues on the supplier's own side weren't tracked anywhere.
What I did
- Built an assistant that turns every reported issue into a fully filled-in ticket within 15 minutes, with no duplicates.
- It tags the right owner on both sides, writes a summary of the whole discussion, and posts updates back to the original thread so the person who raised it always knows where it stands.
- Every hour it re-reads every open ticket and marks it as one of ten states, such as waiting on someone, fix due, or workaround in place, quoting the messages it based that on.
- Split the work into plain code for predictable steps and AI only for judgement, and added a free check so the AI only starts when there's something new. That cut the monthly AI cost from about £490 to £39.
- Designed it to catch up on its own after outages (tested on a 45-minute and a 15-hour gap) and to never drop a ticket to save money.
- Moved it onto company infrastructure with a dashboard the whole team uses, and wrote up the lessons so the next system starts from them.
The jobs that depend on someone remembering are the first jobs to give a machine.