Workflow automation follows rules someone wrote: if this, then that, the same way every time. It is predictable and stops when something unexpected arrives. An AI agent handles variation, reading unstructured inputs and choosing steps toward a goal, but its behaviour is probabilistic. Automation is controlled by its design; an agent must be controlled at each action.
What does workflow automation do well?
Rules-based automation is excellent at stable, high-volume work with clear inputs: post this file to the ledger every night, send a reminder seven days before a due date, move an approved invoice to the payment run. It is cheap, fast and fully predictable. Given the same input, it does the same thing.
Its weakness is the exception. An invoice with a missing purchase order number, an email that does not match any template, a customer whose situation fits no rule: automation stops and routes it to a person. In many operations, those exceptions are where most of the human effort goes.
What does an agent add?
An agent reads the messy input, works out what it means and decides what to do. It can find the purchase order from the line items, understand an unusual email, or notice that a customer's situation resembles a known case. That is why agents are usually introduced at the exceptions automation leaves behind.
The cost is predictability. An agent's choices come from a model, so two similar cases can be handled slightly differently. That is fine when its actions are bounded; it is a problem when they are not.
How do testing and control differ?
| Aspect | Workflow automation | AI agent |
|---|---|---|
| Behaviour | Deterministic | Probabilistic |
| How you test it | Every rule and path | Outcomes across many real cases |
| What failure looks like | Stops at an unexpected input | Proceeds with a plausible but wrong action |
| Where controls sit | In the rules themselves | At every action, outside the agent |
| Human gates | Placed at designed points | Applied wherever an action needs one |
| Evidence of control | The rule definitions | Certification results and a record of verdicts |
The failure row is the one to remember. Automation fails loudly. Agents can fail quietly, producing an action that looks reasonable and is wrong. That is why agents need certification on real cases and checks at the moment of action, not just a review of their design.
Where should human gates live?
In automation, gates are part of each workflow: an approval step after the third rule. With agents, you do not know in advance which step will need a gate, so gates are better placed in the platform and applied to actions, wherever they occur. The argument is made in human gates belong in the platform, not in each workflow, and the workflow-level technique is in how to add a human gate to an agent workflow.
What does this look like in practice?
Take an accounts receivable team. Automation applies incoming payments to invoices when the remittance reference matches exactly. That handles the bulk of payments, at almost no cost. Payments with missing or garbled references fall out to a queue.
An agent works the queue. It reads remittance emails, compares amounts to open invoices, spots that one payment covers three invoices less a credit note, and proposes the allocation. Allocations within tolerance, and resting on the business's own ledger data, are within its depth and go through. Allocations that rely only on what the customer's remittance email says, or that would write off a difference, are held for a person.
Automation and agent do different jobs on the same process. Neither replaces the other.
What stays with a person?
With automation: writing and changing the rules, and handling everything that falls out. With an agent: defining its role, setting limits, approving actions above its depth or resting on unverified claims, and deciding when it may do more. In both cases, a person answers for the outcome.
Related distinctions are covered in how an agentic workflow differs from an AI agent and how an AI agent differs from an AI assistant. For the governing layer agents need, see the governed agentic operating system guide.
Questions people ask
Should we replace our automation with agents?
No. Where rules work, keep them: they are cheaper and more predictable. Use agents for the work rules cannot handle.
Can an agent write automation rules?
It can draft them from patterns it sees, which is a useful way to shrink an exception queue. A person should review and approve each rule before it runs, because the rule will act on every future case.
Is robotic process automation the same as workflow automation?
It is a form of it, focused on driving existing screens. It shares the same strengths and weaknesses: predictable on the expected path, stuck on anything else.