The demos are always customer-facing chatbots. The savings are almost always somewhere else — in the unglamorous work that happens after a customer has gone.
Where the hours actually go
- Reading documents a human typed, and re-typing them into a core system
- Reconciling two ledgers that should agree, and finding the twelve rows that don't
- Assembling the same compliance report from the same four sources every month
- Routing an exception to whoever happens to know how to handle that particular exception
Each of these is high-volume, rule-shaped at the edges and judgement-shaped in the middle — precisely the profile where a language model paired with a deterministic pipeline outperforms either one alone.
The pattern that works
Automate extraction and classification with the model; keep the decision and the write in your own code. The model reads the document and proposes structured fields. Your system validates them against business rules and only then commits. Anything the validator rejects goes to a human, with the model's reasoning attached.
Accuracy stays auditable, and the human queue shrinks month over month instead of disappearing on day one and reappearing in the incident log.
Measure the right thing
Not “percentage automated”. Measure handling time per case, error rate after commit, and the size of the exception queue. A process that is 60% automated with a shrinking exception queue is worth more than one that is 95% automated and generates rework.
What to do first
Pick one document type, one team and one measurable number. Run it in parallel with the existing manual process for a month. If the numbers hold, expand. If they don't, you have lost a month rather than a programme.