Don't pay for the agent. Pay from what it returns.

An AI agent is software that pursues a defined outcome by choosing among tools it has been given permission to use, rather than following a fixed script. In practice that means four things are written down before it runs: the tools it may call, the actions it may take, the decisions it must escalate to a person, and the record it keeps of what it did. We describe it as a domain expert you can audit — expert because it is scoped to real work someone owns, auditable because every one of those boundaries is explicit enough to show an assessor.

A chatbot answers questions inside a conversation; the person remains responsible for acting on the answer. An agent is permitted to act — to call systems, retrieve records, and produce or change work product within limits set in advance. The distinction is standing permission to touch a system of record. A chatbot that can only talk carries a very different risk profile from an agent that can write.

Automation executes a fixed sequence identically every time, which is exactly what you want when the process really is fixed. It cannot handle a case nobody wrote a rule for. An agent reasons about the unfamiliar case instead of halting on it. That flexibility is the value and the risk in the same breath, which is why an agent needs governance that a deterministic script does not. Automation is something an agent can do, not something agents replace — we still build and maintain workflow automation where it is the right tool.

That its boundaries exist as artefacts rather than intentions. A written tool list, separately granted read/write/send permissions, escalation rules defined before deployment, and a log tying each output to the inputs and approvals behind it. If you cannot produce those on request, you do not have an auditable agent — you have a capable one, which is not the same thing and does not survive an assessment.

If the process is stable, the rules are written down, and the exceptions are rare and well understood, automation is cheaper, faster and easier to defend. Agents earn their cost where the work requires judgement across unstructured input — reading a document nobody templated, reconciling records that disagree, triaging cases that do not fit a category. We will tell you when the answer is automation; that is usually the shorter engagement.