A client paid me to rip the AI out of the tool I built them.

I build automations and AI agents for companies. Done it for about forty clients at this point, mostly small and mid-size teams. This one from earlier this year still bugs me. Built a ticket routing tool for a support team. About fifteen people, maybe 90 to 100 tickets a day coming in through Zendesk. They needed each ticket tagged by category and priority so it could land in the right queue. I built it with an LLM doing the classification. Seemed like the obvious call. Feed it the ticket text, get back a category and priority score, route it automatically. Worked well in testing. Client was happy during the demo. In production it was right about 92% of the time. Which sounds fine until you do the math. At their volume that's roughly 7 or 8 misrouted tickets a day. Not a disaster, but enough that the team noticed. And when a ticket ended up in the wrong queue, nobody could explain why. The model just decided. There was no rule to point at, no logic to trace. It just got it wrong sometimes and you had to accept that. Within a couple weeks the team started spot checking every classification before they trusted it. Which meant they were basically doing the work twice. Once by the agent and once by a human making sure the agent didn't screw up. The client called me and said something I didn't expect. He said the tool felt like a black box and his team didn't trust it. He asked if I could make it dumber. So I ripped out the LLM and replaced it with a keyword matcher and a short rules engine. If the ticket mentions billing or invoice or charge, it goes to billing. If it mentions login or password or access, it goes to account. About thirty rules total. For anything that didn't match, the system just surfaced a dropdown and let the rep pick manually. Took me three days to rebuild. Accuracy went up to basically 99% because the rules were transparent and the team could see exactly why a ticket went where it went. When something was wrong they could tell me which rule was off and I'd fix it in ten minutes. Latency went from two to three seconds per ticket down to instant. Monthly API costs went from around $180 to zero. The client told me it was the best money he'd spent on the project. Paying me to take the AI out. I think about this one a lot because it would've been easy to just tune the prompt and push for more accuracy and try to get the team to trust it over time. That's what most of us would do. The model just needs better instructions, right. But the problem was never accuracy. The problem was that people need to understand why a system does what it does or they'll work around it. Same thing happens with agents that make decisions in CRMs or qualify leads or triage anything. If the people using it can't trace the logic they'll build a shadow process next to it and your tool becomes expensive decoration. Not everything needs an LLM. Sometimes thirty rules and a dropdown will outperform a model because the team actually trusts it enough to stop checking its work. After forty-something builds I've learned that the right answer is sometimes less AI, not more. Weird thing to say in this sub but it's true.