Accurate enough to survive, too fragile to scale
A fintech operations team ran critical daily workflows across spreadsheets and email. Importantly, it worked — the numbers were right, the customers were served, and the people involved were good at their jobs.
That is what makes this pattern dangerous. Nothing is visibly broken, so nothing forces a decision, and the cost of the current approach only appears as a ceiling on growth: every additional unit of volume needs another pair of hands, and every new hire has to learn conventions that exist only in someone's memory.
Why not a big-bang migration
The obvious plan is to build the platform and switch. For daily operations handling real money, that concentrates all the risk into one morning.
Instead workflows moved one at a time, each running alongside the spreadsheet it replaced until it had proven itself over a full cycle. Slower, and considerably less likely to produce the failure where an operations team loses its working process on a Monday and has no fallback.
It also meant each migration was small enough to be reversed by one person, which is the property that makes people willing to try it.
The hard part
Encoding the exceptions. The documented process covered the ordinary case; the value the team added lay in knowing what to do when a payment arrived short, when a reference did not match, when a customer paid twice.
None of that was written down anywhere, and none of it surfaced when we asked how the process worked — it surfaced only by watching the work. Any system that automates the documented path and ignores the exceptions gets abandoned within weeks, because the exceptions are where the day actually goes.
What transfers to other operations teams
- A process that works but does not scale creates no urgency, which is exactly why it persists for years.
- Migrate one workflow at a time and keep the old path until the new one has survived a full cycle.
- Watch the work; do not just ask about it. The exceptions are never in the documentation.
- The system should catch errors before a customer does — that is the return, more than the hours saved.