AI workflow or AI agent? How to choose for a business process

The short answer

Choose an AI workflow when the steps are known in advance and only some of them need language understanding. Choose an AI agent when the steps depend on what the agent finds along the way. In our experience, most processes that people first describe as “agent” work turn out to be workflows with one or two judgement calls.

What’s the difference between an AI workflow and an AI agent?

A workflow is code that calls a model at specific points: read this invoice, classify this email, draft this summary. The order of steps never changes. An agent is a model running in a loop: it plans a step, calls a tool, looks at the result and decides what to do next, until it reaches the goal or a limit.

Five questions to decide

1. Can you draw it as a flowchart? If yes, it’s a workflow.
2. Does each case need different tools, in a different order? If yes, lean toward an agent.
3. Is there a right answer you can check? If not, neither option is ready yet. Define one first.
4. What does a mistake cost? The higher the cost, the more approval steps, and the more a workflow makes sense.
5. How often does it run? High volume favours workflows, because per-run agent costs add up quickly.

When an agent is the wrong choice

An agent given a fixed process will still try to plan it, which adds cost, latency and new ways to fail. The same job written as a workflow is a few lines of ordinary code with one model call:

const invoice = await extractFields(pdf)      // model call
const match   = await findPurchaseOrder(invoice)
if (!match || diff(invoice, match) > limit) {
  await sendForApproval(invoice, match)       // a person decides
} else {
  await postToLedger(invoice)
}

If your process looks like this, an agent will do the same work at several times the cost, with less predictable results.