Failure PatternDecision layer
The Demo-to-Retainer Gap: Why RPA Tools Collapse at the Second Client Site
Symptom: The first automation runs clean in a sandbox, then breaks within two weeks of touching a client's live login, VPN, or legacy ERP screen. Root cause: Bots are scoped to a single client's UI surface, so every engagement is bespoke labor dressed up as a productized retainer.
By InnovaAI ResearchPublished
How do you recognize it?
- •The first automation runs clean in a sandbox, then breaks within two weeks of touching a client's live login, VPN, or legacy ERP screen
- •Each new client engagement starts from zero because the bot logic was built for one client's screen layout and never generalized
- •Delivery leads quietly stop quoting automation work after the second project because margins vanished in rework
- •Nobody can say which bots are still running, on whose credentials, or what they touch, until a client asks
- •Change requests pile up for selector fixes and CAPTCHA handling rather than new process coverage
Why does it happen?
- •Bots are scoped to a single client's UI surface, so every engagement is bespoke labor dressed up as a productized retainer
- •No environment contract exists between agency and client: credentials, screen versions, and app updates are unmanaged variables that silently invalidate recorded flows
- •Agent permissions are granted broadly to make the demo work, and no one revisits them, so a single misbehaving worker can touch data it was never scoped to
- •Pricing is set per bot rather than per process outcome, which means maintenance hours are absorbed by the agency instead of billed
How do you fix it?
- •Freeze new bot builds for a week and inventory every live automation by client, credential owner, and last successful run
- •Rewrite the two highest-maintenance flows as prompt-driven agents (BrowserAct for scraping, Robin for multi-site browser journeys) so selector drift stops consuming delivery hours
- •Add a written access clause to each automation retainer limiting each worker to the minimum permissions its task requires, then re-issue credentials
- •Reprice the next three engagements on process outcomes with a named maintenance line item, not a flat bot fee