Failure PatternDecision layer
The Bot Nobody Owns: Why RPA Tools Stall After the First Client Handoff
Symptom: Bots built during the pilot keep running, but no one on the agency side can explain what a given bot does without opening it and reading the step log line by line. Root cause: UI-level bots encode the surface of a process, not its logic, so every cosmetic change in a client application is a silent breaking change with no compiler to catch it.
By InnovaAI ResearchPublished
How do you recognize it?
- •Bots built during the pilot keep running, but no one on the agency side can explain what a given bot does without opening it and reading the step log line by line.
- •A client-side process change (new invoice template, renamed CRM field, shifted approval step) breaks a workflow and the agency finds out from the client, not from monitoring.
- •The retainer line item reads "automation support" with no hours budgeted, so every fix becomes an unpriced favor that eats into the account margin.
- •Two bots in the same client account touch the same record and produce duplicate entries that surface weeks later in a finance reconciliation.
- •The original builder has moved to another account, and the handover document is a screen recording nobody has watched.
Why does it happen?
- •UI-level bots encode the surface of a process, not its logic, so every cosmetic change in a client application is a silent breaking change with no compiler to catch it.
- •Agencies scope automation engagements as one-time builds and price them like projects, which leaves no funded owner for the maintenance work that rule-based workflows demand every month.
- •Permission sprawl accumulates during the pilot: each bot gets whatever access made the demo work, and nobody revisits those grants once the workflow goes live.
- •The agentic layer on top of the bot adds a second failure surface, and multi-step chains propagate a bad input downstream before any human reviews the output.
How do you fix it?
- •Inventory every live bot per client with a named internal owner, the trigger condition, the systems it touches, and the last date someone verified it end to end.
- •Reprice automation retainers to include a maintenance block (start at 15 to 20 percent of build value per quarter) so breakage is funded work rather than goodwill.
- •Strip each bot's credentials to the minimum scope its defined task requires, then log every permission change against the client account.
- •Add a weekly canary run per bot against a test record and route failures to a shared channel the account lead actually reads.