Does part of your work still depend on someone moving data between two systems by hand?
Automation is worth doing when the work is repetitive, the rules are knowable, and the cost of a mistake is real. We start by checking whether all three are true.
The problem
The usual pattern is a task that was never designed — it accumulated. Someone exports a report, reformats it, checks it against a second system, and forwards the result. It works because that person knows the exceptions. It breaks when they are on leave, and nobody can say exactly what the rules were. The first problem is not the automation; it is that the process was never written down.
How we approach it
We map the process as it is actually performed, not as the documentation describes it, and separate the deterministic parts from the parts that genuinely need judgement. Deterministic steps become ordinary code and scheduled workflows, because they have to be auditable. Judgement steps get a language model only where its output can be checked — and where a wrong answer is caught rather than filed. Where the data cannot leave your infrastructure, the model runs on your own hardware.
Worked example
Supplier invoices, 300 a week
A distributor received around 300 supplier invoices a week as PDFs, in a dozen different layouts, and two people re-keyed them into the accounting system. We built a pipeline that extracts the fields, matches each invoice against the purchase order, and posts the clean ones automatically. Anything that fails a check — totals that do not reconcile, an unknown supplier, a duplicate number — goes to a review queue with the reason attached. About 80% now post without a human touch, and the remaining 20% arrive already triaged. The extraction model runs on the company's own server; no invoice leaves the building.
What you get
- A written map of the process, including the exceptions nobody had documented
- The running automation, deployed on infrastructure you control
- A review queue for everything the system should not decide alone
- Handover documentation, so the system is maintainable without us