The challenge
Automation has a long and mixed history in most organisations. Some RPA is pure gold; the rest is fragile, untouched and expensive to maintain. Meanwhile vendors now promise "agents" that solve everything — while analyst firms report that a large share of agentic projects get cancelled before reaching production.
The cause is rarely the technology. It's that processes were never mapped, exceptions were never designed for, and nobody owned the question of what happens when the agent is wrong.
Our approach
We start from the process, not the tool:
- Mapping. We walk the workflows — with process mining where data exists — and identify steps that are high-volume, regular and error-prone.
- Design with boundaries. Every agent gets defined tools, least privilege, cost and step limits, and decision points where a human approves. Logging and traceability are there from day one.
- Run and improve. The automation is monitored against time saved, quality and throughput. Exceptions become lessons — not midnight incidents.
We are equally happy to conclude that a simple script or classic RPA is the right answer. Cheap and robust beats modern and fragile.
Frequently asked questions
Is agentic AI ready for production?
For bounded workflows with clear limits — yes. For open-ended agents roaming freely across the business — no, and anyone claiming otherwise is selling something. Our line: agents do what they are good at, people keep the decisions where errors are expensive.
We already have RPA. What does agentic AI add?
Resilience to variation. Classic RPA excels at regular, unchanging flows and breaks when documents, formats and exceptions vary. Language models handle exactly that variation — the two often complement each other in the same process.
How do you make sure agents don't do stupid things?
The same way as in all other secure engineering: least privilege, defined tools, cost and step limits, complete logging, and human approval at defined decision points. We design for failure, not on the assumption it won't happen.