An agent is software that is allowed to act on your behalf: read a record, create an entry, send a message, move a task forward. That permission is what makes it useful and what makes the decision worth slowing down for. Five questions usually settle it.
One: can you describe the job in a sentence a new starter would understand? "Check whether the customer has an open order, and if they do, reply with its current status" is a job. "Handle customer queries" is a category. Agents work against jobs. If the sentence needs three clauses and two exceptions, the work is not yet ready to be delegated to anything, human or otherwise.
Two: does the task involve more than one system? The strongest case for an agent is work that spans tools — a message arrives in one place, the answer lives in a second, and the record has to be updated in a third. If everything happens inside one system, that system's own rules or a straightforward automation will be cheaper, faster and easier to reason about.
Three: does the work require a decision, or just a sequence? A fixed sequence is a workflow. Write it as a workflow. An agent earns its cost when the next step genuinely depends on what was found: an order exists or it does not, the request is in scope or it is not, the document is complete or something is missing.
Four: what is the worst thing it could do? Agents write, not just read. Before building, list the destructive actions in reach — issuing a refund, cancelling a booking, emailing a customer the wrong balance — and decide which ones require a human confirmation step. Read-only agents are a legitimate first version, and often the right one.
Five: who reviews what it did? Every action should be logged in a form somebody can read on a Monday morning. If no one owns that review, the system will keep operating long after its assumptions stop matching reality, and the first sign of trouble will be a customer complaint rather than a log entry.
If most answers point away from an agent That is a useful result, not a failure. A well-grounded assistant that answers questions accurately, plus a handful of ordinary automations, solves a surprising share of the problems that agents get proposed for — with less to secure and less to maintain.
Our AI agent work always begins with these questions rather than with a platform choice. Contact us and describe the task you are considering handing over.
AI Studio · Published 26 Jan 2026
← All insights