Agentic AI is a process decision, not a model decision
Whether an agent pays depends on which process it joins and who owns that process. The model comes later.
Every few years the industry renames the same promise: data warehouses, machine learning, digital twins, generative AI, and now agents. I have shipped every one of them, and the pattern that separates the ones that paid from the ones that did not has barely changed. The winners were chosen for the process they changed, not for the technology inside.
Agents raise the stakes, because for the first time the software does not just advise. It acts: it raises the purchase order, books the maintenance slot, answers the supplier, updates the record.
Three tests for a process
Before choosing a model or a platform, put the candidate process through three tests:
- Are the inputs digital? Emails, tickets, sensor events, ERP transactions. If the work starts on paper or in someone’s head, digitise that first.
- Are the rules known? Not perfectly, but well enough that two experienced people would usually make the same call.
- Can a mistake be undone? Reversible actions (a draft, a booking, a flagged ticket) are where agents should start. Irreversible ones come later, if ever.
A process that passes all three is a candidate. A process that passes one is a research project, and should be funded like one.
Then the owner
The second decision is who owns the process the agent joins. Not who owns the agent: who owns the outcome. In factories that is usually obvious. In shared services, sales operations and procurement it often is not, and an agent dropped into a process with no owner becomes everyone’s experiment and nobody’s responsibility.
The model comes last
Only then does the technology choice matter: which model, which platform, which cloud and, for anything touching China, which of those can run where the data must stay. Those are real decisions with real trade-offs. They are also the ones vendors are most eager to help with, which is one more reason to make the other two first.
The fastest route to an agent in production is a boring, digital, reversible process with a clear owner, made measurably better. Then the next one.