How to brief an n8n workflow that includes AI steps

The short answer

A normal automation step either works or fails. An AI step can do a third thing: finish successfully and hand back an answer that is wrong. It summarizes the wrong part, invents a detail, drops a field, or answers in a slightly different format than yesterday. The workflow carries on, because nothing errored.

So a brief for an n8n workflow with AI steps needs everything a normal automation brief has, plus five lines about the AI step itself: its one job, real examples, what it must never do, a check that runs after it, and a spending limit. Those five lines are what let you test it before you pay.

Why an AI step needs its own brief

Most of an n8n workflow is plumbing: a trigger, a few app connections, a row written somewhere. You can describe that plumbing in the usual fields: trigger, inputs, output, rules, failure behaviour. The freelancer builds it, you run it three times, and it either does the job or it does not.

The AI step is different in three ways:

• Its output varies. The same input can produce a slightly different answer tomorrow. “It worked in the demo” proves less than it does for a normal step.
• Its failures look like success. A made-up sentence or a wrong category has no error code.
• It costs money per run. The model provider bills by usage, so volume, retries and long inputs show up on an invoice every month. The price guide lists AI steps as one of the seven things that move a quote for this reason.

None of this is a reason to avoid AI steps. It is a reason to brief them precisely.

What running a nightly AI job taught me

I run an overnight job that asks an AI model to write drafts from a topic list. Everything I would now put in a brief for an AI workflow comes from building and running it.

The AI step needs a plain gate after it. After the model writes, a separate script checks each draft against fixed rules: required fields, a placeholder date so nothing can publish by accident, a minimum length, enough headings and working links, no forbidden phrases, and a successful build. A draft that fails the date rule is forced back to the placeholder. None of those checks is clever. They are just not done by the AI.

A rule-based gate does not catch invented facts. My first test draft passed every rule and still contained one first-person sentence with no source attached. The incident turned out to be real, but my first search missed the record, which is the point: a rule gate cannot tell a sourced claim from a plausible one. The fix was a second, independent check: a separate read-only session that compares every number and every first-person claim in a draft against the source files. On the same draft it checked four figures and three anecdotes against the records before passing it.

What you send to the model can arrive broken. At one point the command that started the job cut the instructions at the first double quote, so the model received only the first five words. The model noticed and said so, which is how I found it. The fix was to pass the instructions a different way. The lesson for a buyer: ask how the freelancer checks what the model actually received.

Usage limits stop jobs. A different overnight job hit a monthly spending limit on the account it ran on. It produced its usual output and then ended with an error. A second job would have drawn on the same limit, so I kept it switched off.

Silent fallbacks are accidents. In my own rendering jobs, an unknown option now stops the run with an error instead of quietly falling back to a default, and output that does not fit stops the run instead of shrinking. An unattended job that quietly does something slightly different is worse than one that stops.

The five lines to add to your brief

Add these under the normal fields of the workflow automation brief:

Line 4 is the one most briefs leave out, and the one that saves you. A freelancer who can describe the check in one sentence has thought about what the model will get wrong.

Agent or single AI step?

n8n lets a builder use a single AI step (send text, get an answer) or an agent that picks its own tools and next steps. An agent is more flexible and harder to test, because the path through the workflow can differ from run to run.

For a first build, ask for the simplest version that does the job: a fixed workflow with one or two AI steps in known places. Move to an agent only when the choice of next action is the point. If you are briefing something that talks to customers rather than processing data, the chatbot brief guide covers the extra fields for knowledge, handoff and tone.

Sellers who build this kind of workflow include Abdullah Ramzan, whose gig focuses on AI agents inside n8n workflows that also take actions, and Ali Raza, who lists n8n workflows with AI steps along with documentation and a video walkthrough.

How to test it before you pay

Use the examples from line 2. Run all of them through the finished workflow, then:

Check	Pass looks like
Right answers	Your pass mark, decided before testing (for example 9 of 10)
“Unsure” example	Routed to a person, not guessed
Never-do list	Five attempts, zero violations
Broken output	Force a bad answer (or a bad input); the check after the step catches it
Limit reached	Lower the spending limit for a test; you get an alert, the workflow stops cleanly
Unattended	Three scheduled runs untouched, per the acceptance checklist

Run the examples again a day later. If answers change between runs, that is normal for AI; what matters is that the changed answers are still inside the check.

Where it goes wrong anyway

• The inputs change. Your emails, forms or documents drift over months. Keep adding real examples to the test set and ask the freelancer to re-run it after changes.
• The model changes. Providers update and retire models. Put “re-run the test set after a model change” in the maintenance terms.
• Nobody looks at the rejected answers. The check after the step routes failures somewhere. Somebody on your side has to read that pile, at least weekly at first.

Wrap-up

Brief the plumbing as usual, then treat each AI step as a small job of its own: one task, real examples, a never-do list, a plain check after it, and a spending limit on your account. That turns “it seemed to work” into a test you can run, and it tells the freelancer exactly what they are being paid to guarantee.