The 30% of tickets that never should have been ours

When a client headcount reduction left us absorbing work outside our contractual scope, the fix wasn't just automating — it was automating first and renegotiating scope after. The result: a 70-75% reduction in configuration time, and 40% fewer tickets once we handed the task back to the client.

The problem: work that never should have been in our scope

There was a period, in a banking operation I led, where a headcount reduction on the client's side left us with more work than we had capacity to sustain. The problem wasn't general volume — it was that between 30% and 40% of the tickets arriving were user configuration requests: initial setup, office changes, approval-level changes. Work that, to begin with, was never part of our contract's original scope. Our CEO had personally taken it on as a temporary commitment, while we finished the platform rollout.

The operational problem was concrete: each of those tickets consumed close to half an FTE during the first 5-6 days — preparing files by hand, running validations, loading data. Every time one of these requests came in, the rest of the ticket queue fell behind on its diagnostic SLAs.

Two decisions, not one

I analyzed the most frequent ticket types, classified the root causes, and reached two decisions, not one. First: with no dedicated development support, my team built an AI tool that read the source documents directly and wrote the configuration both to the platform engine and the decision-rule system, via their APIs, with integrated validation to ensure each configuration met our operability standards. Second, and what actually solved the underlying problem: once the tool matured, I negotiated with the client to formally transfer that task to their own Operations team — because it should never have been in our scope to begin with.

The result, measurable

The result: a 70-75% reduction in configuration time while we still had it, and after the transfer, our internal team received 40% fewer tickets overall and consistently met the diagnostic SLAs again — not just faster, but sustainably, because the root cause of the overload disappeared; it wasn't masked by a tool that made it faster to tolerate.

To put that number in context: the 2026 enterprise median for AI deflection rate is around 41%, according to production benchmarks from the industry — and the general deflection average in the tech sector is barely 23%. The 70-75% reduction we achieved in configuration time wasn't generic self-service automation — it was a tool built specifically for a real bottleneck, without waiting for an external vendor to solve it.

The lesson I take away

Automating work that shouldn't exist in your scope is only half the solution. The other half is having the uncomfortable conversation to hand it back to whom it actually belongs.