Workflow or AI agent?
Choose predictable automation where the steps are known, and use an agent only where the task needs judgment and controlled tool access.
A contact form does not need an agent to copy an email address into a project record. A request that arrives as a messy paragraph may benefit from AI to identify what the person wants. Those two jobs can sit inside the same workflow without handing the whole process to a model.
As of October 2026, platforms including n8n and Make combine conventional automation with AI steps and agents. The first decision is which parts should be predictable and which parts need interpretation.
Use a fixed workflow when the route is known
A workflow follows steps you define. When a valid form arrives, store it, notify the team, and acknowledge receipt. You can specify that route without asking a model to choose it.
Use ordinary rules for known categories, required fields, numerical limits, duplicate detection, and assigning work from an agreed lookup table. These steps are easier to inspect when they remain explicit.
Use an AI step before you reach for an agent
An AI step can interpret text or draft a response inside a controlled sequence. It does not have to plan its next action or select its own tools.
Consider this fictional enquiry: “We’re launching a furniture range in November and need product images plus a simple landing page. We have photography but no campaign direction yet.” A model could extract services, launch timing, supplied assets, and missing information. The workflow then checks the fields and gives a person the original message alongside the summary.
That is useful interpretation within a defined process. It does not require an autonomous agent.
Use an agent when the next step cannot be fixed in advance
An agent can choose actions based on the task and what it learns. For example, preparing an internal brief may require looking up an existing project, finding supplied documents, and deciding which questions remain unanswered.
That flexibility adds choices you must control. Give the agent a narrow objective, only the tools it needs, and a clear stopping condition. Limit attempts and cost. A task should stop when it has produced an accepted result or reached a defined point for human review.
Do not let a model’s own confidence score stand in for evaluation. Test it against representative enquiries with known expected outcomes, including ambiguous messages and missing details.
A useful intake design
For the fictional furniture enquiry, a manageable flow looks like this:
- Validate: check required fields and assign a unique enquiry ID.
- Interpret: use an AI step to extract requested services and timing into a fixed structure.
- Check: reject invalid output and flag ambiguity instead of filling gaps with guesses.
- Route: assign the brief using explicit rules and show the original message.
- Review: let a person approve the proposed scope and reply.
- Record: keep the status, attempts and outcome so failures can be found.
The AI helps prepare the work. The workflow controls where it goes. A human commits to the client.
Design the failure path before connecting everything
A temporary service failure may justify a bounded retry. Missing information may require a question. An ambiguous request may belong in a review queue. These are different situations, not one generic “try again” condition.
Before retrying any action that changes a record or sends a message, check whether it already happened. Use the enquiry ID to prevent duplicate tasks or replies. When retries are exhausted, preserve the input and tell the owner what failed.
n8n’s current platform description includes human approvals, explicit logic and execution monitoring. The availability of those controls is useful; you still need to configure and test the route for your task.
Measure the job, not the novelty
Compare the result with the manual process. Look at correct routing, time spent reviewing, duplicate actions, and failures caught before a client sees them. Keep the version that reduces work without hiding mistakes.
Choose the smallest amount of autonomy that gets the job done.
Start with a workflow. Add interpretation where it earns its place. Introduce an agent when choosing the next step is part of the actual problem.
Sources and further reading
Product documentation checked on 2 October 2026: n8n: AI workflows and controls and Make: AI automation. The enquiry and intake sequence are illustrative, not reported client results.