The difference is who chooses the path. In an agentic workflow, people design the steps in advance and a model handles some of them, such as reading a document. An AI agent is given a goal and tools and chooses its own steps. So control sits in the design for a workflow, and at each action for an agent.
What is an agentic workflow?
An agentic workflow is a process with fixed stages, some of which use a model. For example: receive an invoice, extract its fields with a model, match against the purchase order with rules, route mismatches to a person. The model does real work inside a step, but it does not decide what the next step is.
Because the path is fixed, you can review it before it runs. Every possible route through the workflow is visible on a diagram, and controls can be placed at known points.
What is an AI agent?
An AI agent receives a goal, such as "resolve this payment exception", and a set of tools. It decides what to look at, what to try and when it is done. Two exceptions of the same type may be resolved by different sequences of steps, because the agent responds to what it finds.
That flexibility is why agents handle variety better than workflows. It is also why their routes cannot all be reviewed in advance.
How does each one change governance?
| Aspect | Agentic workflow | AI agent |
|---|---|---|
| Who chooses the steps | The designer | The agent, at run time |
| Where controls sit | At known points in the design | At every action, wherever it happens |
| What gets tested | Each step and each route | Outcomes across many real cases |
| Typical failure | A step mishandles an unusual input | The agent takes an unexpected route |
| Human gates | Placed in the flow | Applied by the platform to any action that needs one |
The last row is the practical consequence. In a workflow, you add a gate at the step that needs it; see how to add a human gate to an agent workflow. For an agent, you cannot know which step will need the gate, so the gate must sit in the platform and apply to the action, wherever in the run it occurs.
What does this look like in practice?
Take a lender's direct debit failures. As a workflow: a failed collection triggers a fixed sequence. Check the reason code; if insufficient funds, schedule a retry in three days; if account closed, draft a letter to the borrower; route anything else to a person. A model reads the bank's free-text reason when the code is missing. Every path is known.
As an agent: the goal is "get this collection resolved". The agent reads the account history, notices the borrower has called twice this month, sees a pending address change, and decides a retry is unlikely to succeed. It drafts a message offering a payment plan. That is a route no designer drew. Whether the agent may send that message, or set up the plan, is decided by the platform at the moment it tries: depth, limits and the source of the facts it relied on.
Both are useful. The workflow is cheaper to run and easier to assure for common, predictable cases. The agent handles the cases the workflow would have routed to a person.
When should you use which?
Use a workflow where the process is stable and the variations are known. Use an agent where cases vary enough that a fixed path would send most of them to a person anyway. Many processes use both: a workflow for the main path, and an agent for its exceptions.
Rules-based automation without any model is a third option, compared in how an AI agent differs from workflow automation.
What stays with a person in either case?
Designing the workflow, or defining the agent's role. Setting limits. Approving actions that are irreversible, above the business's limits or based on unverified claims. In a workflow, people sit at designed points. With an agent, they sit wherever the platform holds an action.
The governing layer that holds those actions is the subject of the governed agentic operating system guide. Where agentic workflows meet agent platforms, see how an agentic operating system differs from an agent platform.
Questions people ask
Is an agentic workflow safer than an agent?
It is easier to assure in advance, because every path is known. It is not automatically safer; a model step inside a workflow can still mishandle an input, and the workflow will pass the mistake along.
Can an agentic workflow contain agents?
Yes. A workflow step can hand a case to an agent with a goal, then resume when the agent finishes. The agent's actions are governed in the usual way during that step.
Which is easier to explain to a regulator?
A workflow's design is easier to show on paper. An agent's behaviour is explained through its role, its certification on real cases and the record of its verdicts, which can be just as clear if the record is good.