Somewhere in your company, someone is about to suggest putting an AI agent on a process that a simple automation would have handled for a tenth of the cost. Or the opposite: forcing a rigid script onto work that changes every time it shows up. Both mistakes are common. Both are avoidable with one honest question.
The one question that decides it
Can you write down every step and every exception in advance? If yes, you need automation. If the right next step depends on reading something messy and making a call, you may need an agent. Most processes are the first kind, and they get sold as the second.
What automation is good at
Rules-based automation does the same thing the same way every time. A form comes in, the data moves to the right system, an approval gets requested, a follow-up goes out on day three. It is fast, cheap to run, and completely predictable. When it fails, it fails in a way you can trace. For anything repeatable with clean inputs, nothing beats it.
What an agent is good at
An agent earns its cost when the input is unstructured and the path is not fixed. A customer email that could be a refund request, a complaint, or a sales lead. A contract that needs checking against your policy. A support ticket that needs context pulled from three systems before anyone can respond. The agent reads, decides, acts, and hands off to a person when it is unsure.
The cost nobody puts on the slide
Agents are harder to run than automations. They need rules, escalation logic, monitoring, and a way to find out when they are wrong. Gartner predicted in 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value, and weak risk controls. The same research estimated that only about 130 of the thousands of vendors calling themselves agentic were genuinely delivering it. The technology is not the problem. Reaching for it too early is.
How to decide in ten minutes
Take one process and score it. Are the inputs structured or messy? Are the rules fixed or do they depend on judgment? What does a wrong answer cost you? Fixed rules and clean inputs mean automation. Messy inputs and judgment calls mean an agent, with a person reviewing the exceptions. Most real workflows are a mix, and the best systems use both: automation for the predictable steps, an agent for the one step that needs reading.
Start with the cheaper one
If you are unsure, build the automation first. It will show you which step keeps breaking, and that step is usually where an agent belongs. You learn where the judgment lives before you pay for it.