Workflow or AI agent? How to decide which one a task needs.
Many tasks do not need an agent. This is the short test we use to decide, with two of our own builds as examples: one workflow that routes invoices with a plain date rule, and one that matches bank transfers and sends what it cannot read to a person.
Who chooses the next step?
In a standard workflow, a developer fixes every step in advance, so the same input always takes the same path.
In an AI agent, a language model decides the next step, calls tools such as an inbox, a CRM or a database, and keeps going until the task is finished or it needs a person. The model chooses the steps within limits you set.
That flexibility helps when the input varies and the path is not known beforehand. It also adds cost, variability and risk, which is why agents need limits, logging and tests.
Which approach fits which kind of step.
Work through the task one step at a time. A single process often mixes several of these.
| If the step is... | Use | Why |
|---|---|---|
| Known in advance, with structured input | A plain workflow | A rule is cheaper to run, faster and much easier to test than a model. |
| Known in advance, but the input is free text, a document or an inconsistent reference | A workflow with a language model for the reading step only | The model reads the unstructured input and hands a structured result to the next step, or to a person to review. Deterministic code does everything else. |
| Not known in advance: it needs judgment on varied input, and a mistake can be caught or reversed | An agent, within limits | The model chooses the steps, with permissions, logging, tests and a human checkpoint around it. |
| Hard to undo if it goes wrong | A person approves it first | An agent's permissions name the actions that need a person's approval before they run. |
A starting point for the conversation, not a rule that fits every business.
What the plain version looks like.
Both of these are workflows built with ordinary code. Neither is an agent.
AI for the messy part, a person for the decision.
The step that can use a model is the one that reads input a rule cannot, such as free text or documents. In the illustrative hotel invoice reconciliation example, a model suggests candidate invoices for a payment with no usable reference and a person confirms; it does not mark anything paid by itself.
Tasks that suit an agent are the ones where the path itself varies, such as triage and routing, research and first drafts, or cross-system updates. The task types on AI agents for business are illustrative, not client results.
What goes around it.
Agents get the same engineering discipline as any production system. In the architecture stage we settle each of these in writing:
- Permissions: which tools and data the agent can touch, and which actions need a person's approval before they run.
- Review: outputs that matter go to a human checkpoint until you decide the agent has earned more autonomy.
- Logging: we build in tracing and logging for reasoning steps and tool calls, with cost and latency per run.
- Testing: we build in regression tests that fail the build when behavior drifts.
- Rollback and cost: prompts, tools and policies are versioned so you can roll back, and token budgets, rate limits and circuit breakers cap spend.
Let the ranking decide.
The AI readiness assessment produces an operations map and an ROI model that show whether an agent, a workflow or neither is the right answer, including the opportunities not worth doing.
Questions about workflows and AI agents.
What is the difference between an AI agent and a workflow?
In a standard workflow, a developer fixes every step in advance. In an agent, a language model chooses the next step and calls tools such as an inbox, a CRM or a database, within limits you set, until the task is finished or it needs a person.
When should we not build an agent?
When every step is known in advance. A deterministic workflow is cheaper to run, faster and much easier to test. Reach for an agent when the task needs judgment on varied input and a mistake can be caught or reversed.
Can a workflow use a language model?
Yes. We use a language model for the parts that read unstructured input, such as free text, documents or inconsistent references, and deterministic code for everything else.
What happens when the workflow cannot decide?
The step goes to a person. In our Wise payment gateway, a payment without a recognisable invoice reference lands on a manual matching screen instead of being guessed.
Tell us the task you are weighing up.
Discovery calls are free and last 30 minutes, with someone who will actually do the work. If a simpler tool solves it, we will say so.
The button opens our enquiry form; we reply by email to agree a time for the call.
Coverage: Europe, Asia and remote. Email [email protected].