AI agent or fixed workflow: which to build first
Use a workflow when the steps are known and an agent when the input is free text. How each one fails, what each costs to run and how to combine them.
- 5 min
The rule of thumb
Write the steps on paper. If you can list them before the data arrives, build a workflow: a web form that creates a contact, a call that is logged on the right record, an invoice that goes to accounting. A workflow does the same thing every time and costs little to run. If the input is free text and the right action depends on what it says, an agent can help. Two examples are an email that asks three things at once and supplier messages that arrive in different layouts. The agent reads the text and picks an action from a short list.
Side by side
The table compares the two on the points that change a project.
| Question | Fixed workflow | AI agent |
|---|---|---|
| What it suits | Known steps in a known order. | Open text, where the meaning decides the next step. |
| Cost to run | Small and predictable for each run, at a given volume. | Varies with the length of the input and the number of steps the agent takes. |
| How it fails | Stops at the failed step and sends an alert with the data it received. | Can choose a wrong action that looks reasonable, so it needs limits and review. |
| Who approves | A person approves the design once, when the flow is built. | A person approves each action above a limit, and every write at first. |
| How you test it | Same input, same output, so a fixed set of test cases is enough. | A sample of real messages, checked by a person. |
The hybrid: an agent inside one step
Most useful setups combine the two. The workflow keeps the path and the rules. The agent handles the single step that needs judgment. Take an inbound email to a sales team:
- The workflow receives the email and finds the contact in the CRM.
- The agent reads the text and proposes one action from a short list: log a note, create a task, draft a reply or pass it to a person.
- The workflow checks the proposal against fixed rules. An agent may draft an email and may not send one.
- A person approves anything that leaves the company.
- The workflow writes the result to the CRM and records the call in a log.
Only the second step uses an agent. The other four are the same every time, so they are cheap to run and easy to test.
A short checklist
Answer these before choosing.
- Can you list the steps before the data arrives? If so, start with a workflow.
- Does the input vary in layout or wording from one item to the next? That is where an agent earns its place.
- How costly is a wrong action? If it is hard to undo, put a person in front of it.
- Who reviews the first outputs, and where do the corrections go?
- Where do the keys and the log live? They belong in your own accounts.
What to build first
Build the workflow first. It pays back sooner, it is easier to test, and it shows which inputs are messy enough to need an agent. When that step appears, add the agent to it with a scoped key, an approval and a log, as described in connect an AI assistant to your CRM safely. The idea behind the server that gives the agent its access is in what an MCP server is. The n8n automation agency page shows the workflows AINS builds. The free 30-minute review shows which of your tasks fit which pattern: see the services or book the review.