We get asked about AI agents more than anything else. That is understandable — it is what everyone is talking about. But in roughly one conversation in three we end up recommending something else, and it is worth explaining why.
An AI agent is good at tasks that need judgement. It can read an enquiry and work out what it is about, write a reply that fits the situation, and handle the cases nobody anticipated. That is a real strength. But it is also a cost: a model is slower, more expensive per run and less predictable than an ordinary rule.
Sign one: the task can be described as a rule
If you can say "when X happens, do Y" without caveats, you do not need a model. You need an automation. It is cheaper to build, costs almost nothing to run, and does the same thing every time. A new order that should create a task, a submitted form that should land in the CRM, an invoice that should be forwarded — those are rules, not judgement.
Sign two: a mistake is expensive and hard to spot
Models make mistakes. Not often, but often enough to matter when the consequence is large. If the agent is to send money, change a contract or promise a customer something, there has to be a person in the approval at minimum. If it is a task where the mistake cannot be spotted until it is too late, a model may be the wrong tool however well it performs in testing.
Money that moves
Payments, credit notes and refunds. Here the approval step stays permanently.
Promises to customers
Delivery times, prices and commitments. An agent may draft the wording, but not send it.
Changes to agreements
Terminations, renewals and terms. The consequence is legal, not practical.
Data that gets deleted
Anything where undo is not an option. Let the agent flag rather than remove.
Messages that cannot be recalled
Sends to many recipients at once. The mistake is spotted only once it is out.
Sign three: the volume is small
An agent pays off when it replaces hours, not minutes. If the task comes up three times a week and takes five minutes, the total gain is fifteen minutes — and that justifies neither the build cost nor the running cost. Then it is cheaper to let a person do it, or to remove the task altogether.
What we do instead
We always start by mapping the task before choosing the tool. Often it turns out that half of what looked like an AI problem is a missing integration, and the other half is a decision nobody has made. What is left is what an agent should be used for — and there, it is genuinely good.