LinkedIn·Monday, 24 August 2026·2d ago

I've been reviewing some RPAs for a client. It's an absolute graveyard and easily upgraded: 1. What RPA does → An RPA bot is a recorded…

Assem Chammah

CEO @ Nexus | AI transformation for enterprises | Clients inc. Orange Group & Lambda

I've been reviewing some RPAs for a client. It's an absolute graveyard and easily upgraded: 1. What RPA does → An RPA bot is a recorded script. It clicks buttons and types into fields the way a person would. → It operates at the screen layer, so it depends on the button staying put. 2. Process drift → Never have I seen a company that follows its SOP perfectly. There's the formal version, and there's what people do. → You build good automation on A, B and C. Six months later nobody is using it, because the process changed and the script didn't. → Building the thing is easy. Keeping it current as reality drifts is what no company does. 3. Maintenance tax → 70-75% of RPA total cost of ownership goes on building, fixing and maintaining them. Licensing is the small part. → About 45% of companies report robots breaking weekly (Forrester). 4. Three zones → Think about the task you need solving. The process solving it can either be very reliable and the output looks standardized, or it can be very flexible but not very reliable to always give the answer you want. → Automation sits at the reliable end. → Open-ended agents sit at the flexible end. You hand over an objective and the agent picks its own steps. It will invent things with confidence. We don't want this. → Agentic workflows sit between them. This is when you begin to interject AI into RPA workflows. 5. When automation wins → Payroll, tax and accounting are 90-95% automation. → The rules are written down and the output has to be good enough that it passes any sort of financial audit. 6. When agents win → A call agent has to decide what to offer while the customer is still talking. → Every call arrives different. What the customer opens with rarely matches what they actually need. → No decision tree holds that. The agent reads the account and the history, takes what's said on the call, then picks the offer. → The agent converts better than the people who used to do the job. Cheatsheet: - Map what the process actually is before automating any of it - Send fixed rules with auditable output to plain automation - Send branching judgment calls to an agent - Go and open what's already deployed, then sort every bot into one of the three zones How old are the rules running your automation stack?