Refunds and terminations, from paper to straight-through
Three years and forty sprints to take a manual, paper-driven process to zero-touch processing for most requests, with turnaround from days to minutes.
The constraint
When a member leaves, their money has to leave with them, correctly, and the record has to show who decided what. For decades that ran on paper and people. Every refund and termination passed through hands, and every pair of hands was both a control and a delay. Turnaround was measured in days. The rules lived in the people who applied them, which meant they'd never been written down anywhere a machine could read.
The decision
Straight-through processing. The system decides what it can prove and people handle the rest. Every rule that a person applied by habit got written down, argued over with the business and audit, and put in code with a test. A request that clears every rule goes all the way through without a human touch. A request that trips one lands in a queue with the reason attached. Segregation of duties stayed, enforced by the system instead of by the seating chart.
What it cost
Three years and forty sprints. Most of the first year went into the rules, not the code. Getting the business, risk and audit to agree on what "correct" means for every case is slow, and it's the whole job. Skipping it would've shipped a faster version of the old ambiguity.
What shipped
Zero-touch processing for more than 95 percent of requests. Turnaround went from days to minutes. The exceptions reach a person with the rule that stopped them, so the review starts from the reason instead of from scratch.
What it showed
The design lives in the five percent. The happy path is easy once the rules exist. The exception path is the product, because that's where money and judgement meet. I build agent systems the same way now: the model proposes, code checks, and a human sees only what the system couldn't prove.