Around 800 hours of staff time saved every month.
Context
A large London accounting firm ran billing and reconciliation by hand: a core team of twelve doing the work, five approvers signing it off, and around ten more people contributing pieces along the way. Month-end pulled qualified staff into copy-paste work.
The problem
The volume was not the issue. The variation was. Every client had exceptions, and the exceptions lived in the heads of the people doing the work, spread across nearly thirty of them. The firm had already tried to solve it internally, with a change management lead and an AI technician, and that attempt had failed.
The key decision
I spent the first two weeks on process mapping and planning before writing a line of code, and modelled the exceptions as first-class rules rather than edge cases. Shipping a happy-path automation quickly is exactly how the internal attempt died. Slower start, but the system had to be trusted by the people whose job it changed.
What I built
A production pipeline on the Claude API with Make.com orchestration and custom Python: document ingestion, rule-based and LLM-assisted matching, human review queues for genuine judgement calls, and full audit logging.
The outcome
Around 800 hours of staff time saved every month. The system is live, monitored, and has survived a year of month-ends, which is the test that matters.
Stack
Claude API, Make.com, Python, custom integrations.
Have a similar problem?
Tell me what you are trying to move, and I will tell you straight whether I can move it.