Skip to main content
AI Automation With Zapier, Make and n8n

LESSON 3 OF 6

Which step should be a model

BY THE END OF THIS LESSON

Use a model for the step that needs judgement, and nothing else.

Most steps should not be a model

The weak version and the strong one.
AvoidA model formatting a dateslower, and occasionally wrong
Do thisThe direct methodnever gets them wrong

Moving a file, looking up a record, formatting a date, posting to a channel — these are exact operations with one right answer. A model doing them is slower, costs more, and occasionally gets them wrong, which the direct method never does.

Reach for a model only where the step genuinely needs judgement about language or meaning.

What a model is genuinely good at here

3 options, side by side.
Classifyinginto one of your categories
Extractingspecific fields out of messy text
Draftinglanguage in a consistent tone

Three shapes come up constantly: classifying something written by a person into one of your categories, extracting specific fields out of messy text, and drafting language in a consistent tone.

All three share a property — the input is unpredictable human text, and the output is something you defined. That is exactly the gap a rule cannot cover.

Constrain the output or you cannot use it

Cannot be routed anywhere against One of these five words and nothing else, drawn at their real relative sizes.

Cannot be routed anywhere

Back It depends, but probably billing

One of these five words and nothing else

Back billingEscape unclear

A model step feeds the next step, which is a machine. "It depends, but probably billing" cannot be routed anywhere.

Say exactly what may come back: one of these five words and nothing else, or JSON with these keys. Add an explicit escape — a category called "unclear" — because otherwise it will pick the closest wrong answer rather than admit it does not know, and you will never see that happen.

WORKED EXAMPLE

UNUSABLE OUTPUT

Read this support email and tell me what it is about.

ROUTABLE OUTPUT

Classify this support email into exactly one of: billing, bug, feature-request, account, unclear. Reply with the single word and nothing else. Use "unclear" if it does not clearly fit one — do not guess the closest.

The next step is a machine deciding where to send this. It can act on "billing" and it can act on "unclear". It cannot act on a paragraph, and without the escape hatch every ambiguous email is silently filed as whatever was nearest.

YOUR TURN

Find the one step in your automation that actually needs a model.

RUN THIS

Here are the steps in an automation I am building: [list them]. For each, tell me whether it needs a model or whether it is an exact operation that should be done directly. Be strict — most steps are exact. For the ones that do need a model, write the instruction, say exactly what the output may be, and include an escape value for when the input does not fit.

What a good result looks like

Expect one or two model steps at most. If everything comes back needing a model, ask which of them have exactly one right answer — those should be direct.

KNOWLEDGE CHECK

Answer all 3 correctly to complete this lesson.

  1. 1. Why not use a model for looking up a record or formatting a date?

  2. 2. Why does a model step need an "unclear" option?

  3. 3. What do classifying, extracting and drafting have in common?

REMEMBER THIS

A model for judgement about language; direct operations for everything else — and always constrain the output.