Many RPA business cases lose their footing in the year after launch. The build came in on budget, the bot saved the hours the deck promised, and then a UI changed, the queue backed up, an exception lane fell silent, and somebody on the operations team started doing the work by hand again. The original ROI math never accounted for any of that.
The four costs that show up after launch
- Drift maintenance: the upstream system changes, the bot breaks, and the fix consumes days of senior time.
- Exception handling: every bot leaves some share of volume unprocessed, and somebody has to action it. Budget this as a fractional FTE, not as zero.
- Process change: the underlying business process changes faster than the bot does. Each material change triggers a re-spec.
- Sprawl tax: once the program has dozens of bots, finding which one to fix takes an hour. Without a catalog, time-to-fix grows with every new workflow.
A maintenance budget that survives audit
Our default model: budget maintenance as a recurring line item per bot, sized from the exception rate and change frequency you observe in production, not from a flat guess. It covers monitoring, exception triage, small change requests, and platform upgrades. It does not cover net-new functionality.
The bot catalog is the single most valuable artifact
We require every engagement to ship a bot catalog: name, owner, business process, system dependencies, schedule, last-changed date, and one paragraph of business context. Six months later, when something breaks at 7am, the on-call engineer can answer 'what should this be doing?' in seconds instead of after digging through code.
When to retire a bot
RPA programs accumulate bots that nobody runs. We sweep the catalog on a regular schedule and retire any bot with little or no recent activity. The exception is monthly close bots: those have to be measured against close-week activity, not a rolling window.
Sweeping is the only way to keep an RPA program honest. The cost of a retired bot is zero. The cost of a forgotten bot is the time it takes the next engineer to figure out whether it still matters.