AI Workflow or AI Agent? Look at the Shape of Your Process First

The same automation project, two vendor pitches: one proposes a workflow — fixed process, stable and controllable; the other proposes an agent — smarter, it figures things out by itself. Both sound reasonable, and the quotes are far apart. What actually matters here is not which word sounds more advanced, but what the process you want to automate looks like.

This article skips the spec comparison and does one thing: explain the real difference between the two forms, then give you a set of questions you can hold up against your own process.

One runs on rails. One navigates.

A workflow is a path drawn in advance. Which step receives the document, which step extracts the fields, above what amount a human takes over — every node and every branch is fixed at design time, and the AI only does its assigned job at its assigned spot. The payoff is that every step is predictable: the same input takes the same route, and when something breaks, you know which node broke.

An agent works the other way round: you hand it a goal, and it plans its own steps, calls tools, and adjusts based on intermediate results. For the full picture, see What Is an AI Agent; here one point is enough: the path is not drawn by a person — the agent decides it on the fly. The same task may run differently today than it did yesterday.

So the difference is not degree of intelligence, but whether the path is fixed in advance. That single difference determines what each form is good for.

Processes that are born for workflows

The first profile is high determinism: inputs arrive in reasonably stable formats, the rules can be written down in plain language, and the exceptions can be enumerated. Expense pre-checks, contract field extraction, high-frequency support questions, weekly report consolidation — all of this qualifies. These processes usually share another trait: high volume and repetition, which is exactly where automation pays back fastest.

The second profile is anything with compliance or audit requirements. For finance, contracts and external commitments, you must be able to answer how the system decided at the time — and a workflow satisfies this by construction, because the flowchart is the audit trail. If you find you cannot even draw the flowchart, the problem is not tool selection; the process itself has not been mapped yet. What to Map Before You Automate covers that step.

Where an agent actually earns its place

Invert those conditions and you have the agent's home ground: paths that cannot be enumerated, inputs that are wildly irregular, work that requires judgment along the way. Open-ended research, where what you search next depends on what you just found. Tracing a mismatched figure across several systems. Turning a completely free-form inquiry email into a structured requirement. These jobs cannot be written as a fixed flow — forcing them yields a flowchart that is never finished.

But an agent's flexibility has a price. You need permission boundaries — which systems it may read, what data it may change, how much a single task may consume. And you need to budget supervision: its output needs human review, and it tends to fail in ways nobody predicted. Deploy without settling those two bills, and flexibility turns into loss of control.

The cost structures differ in kind

A workflow's cost is clean: build once, then each run costs roughly the same, so even at high volume the spend is predictable. Errors surface during design and flatten out after launch. Auditing is simple too — every step leaves a record.

An agent's ledger is messier. Each run takes a variable number of steps, so consumption floats with the task. The bigger item is hidden cost: review time, cleanup after failures, and the upkeep of permission management itself. None of this shows up on the invoice, and all of it is real. It is not money you cannot spend — it is money you must plan for at approval time.

The sensible default: workflow as skeleton, agents for the elastic parts

In practice this is rarely an either-or decision. The common, sensible shape is hybrid: the trunk is a workflow — intake, classification, routing and archiving all run on rails; the one node that genuinely needs understanding and judgment is delegated to an agent; its output returns to the workflow, which carries on with the fixed steps.

Take inbound quote requests: mail arrives, the customer is identified, a record is created, the case is assigned — workflow. Turning a rambling paragraph of requirements into structured fields — agent. A salesperson confirms the result, and it flows back into the quoting process. The whole remains predictable and auditable, while the agent's permissions stay fenced inside a single node.

Five questions to answer before you commit

1. Can the path be drawn as a flowchart? If yes, use a workflow. If no, then consider an agent.
2. Can you enumerate the exceptions? If yes, they become workflow branches. If not, hand them to an agent — or to a person.
3. What does an error cost? Where money, compliance or external commitments are involved, you need auditability: lean towards a workflow with human confirmation points.
4. What is the volume? Large and steady favours the workflow's cost profile; rare and different every time favours the agent.
5. Who will supervise? If nobody has capacity to review the agent's output, do not deploy an agent yet.

If the five questions still leave you undecided, step one level up: should this project exist at all? Six Questions to Ask Before Approving an AI Project is the checklist for that layer.

Don't use architecture to look advanced

One selection motive deserves to be named: choosing an agent because it sounds more cutting-edge. It does demo better, and it does make a better slide. But the value of automation never lived in architectural novelty — it lives in whether the thing reliably does the job. A workflow that has run quietly for a year is worth more than an agent that needs debugging every week.

Map the process first, then pick the form. Most of the time, the shape of the process answers the question for you.