Choosing an AI Step
Four steps in Glow use AI. Picking the wrong one is the most common reason an automation ends up slower, dearer or less predictable than it needed to be, so this page sets out the trade behind each.
The short version: most jobs are an AI Prompt. Reach past it only when you can say what the prompt cannot do.
Start here
Answer these in order and stop at the first yes.
| Ask yourself | If yes |
|---|---|
| Am I trying to build the workflow itself, rather than run something? | Workflow Assistant |
| Do I need up to three ordered values from one piece of text? | AI Data Transform |
| Does it have to act in another app, or decide its own next move? | AI Agent |
| Anything else | AI Prompt |
AI Prompt
One question in, one answer out. You write what you want in plain language, reference the data it should use, and the step returns a single result.
What it is good at
- Text work. Classifying, summarising, rewriting, drafting, translating.
- Volume. One model call, so it is the cheapest and fastest of the four.
- Repeatability. The same input gives the same shape of answer.
- Structured output. Describe the shape and a later step reads
{{ 2.result.status }}knowing it is there.
What you give up
- It cannot act. No sending, no writing to a CRM, no fetching a page.
- No second attempt. It cannot check its own work or try another approach.
- It sees only what you reference. An earlier step’s value has to be written in as a placeholder.
Reach for it when the job is one judgement on one piece of text. That covers far more than people expect.
AI Agent
A goal in, and it works out the steps. You say what you want achieved and switch on the accounts it may act through. It decides its own next move, uses tools in sequence, and stops when it judges the goal met.
What it is good at
- Unplannable work. Research a company, triage against several sources, chase a thread.
- Acting in other systems, through the accounts you connect.
- Several tools in one step. Search the web, read a page, calculate, write a record.
- Context across runs, if you turn Memory on.
What you give up
- Predictability. Two runs on the same input can take different routes.
- A forecastable bill. It is the dearest of the four and the hardest to estimate.
- Traceability. More moves means more places to go astray, so put Human Review in front of anything irreversible.
Reach for it when you cannot write down the steps yourself, or when the work has to touch another system to finish.
The most common mistake is starting here. An agent looks like the powerful choice, so it gets picked for jobs a prompt would do more cheaply and more predictably. If you can describe the steps, you do not need something that works them out.
AI Data Transform
One text in, up to three ordered values out. Write one instruction per value — “the customer’s full name”, then “the plan they asked about” — and read them by position as {{ N.result.0 }}, {{ N.result.1 }} and {{ N.result.2 }}.
What it is good at
- Values out of mess. A raw email, form or document becomes an ordered result you can use downstream.
- No prompt to write. One line per value, and no JSON shape to describe.
- Speed. The instructions run at the same time, so three is not three times the wait.
What you give up
- Three instructions at most. For a fourth value, use AI Prompt with a response shape.
- No control over wording. It returns the value, not a sentence you styled.
- It is not in the AI group. Find it under Tools → Data → Change.
Reach for it when you want fields rather than prose, and three of them will do.
Workflow Assistant
Describe the automation and it builds it. This one is different in kind: it does not run inside a workflow, it writes one.
What it is good at
- A working draft in minutes, from nothing.
- The fiddly configuration. Field mappings, channel names, references between steps.
- Explaining a canvas somebody else built.
- Working beside you. The canvas stays yours while it builds.
What you give up
- A draft, not a finished workflow. Read what it made before you publish.
- Precision. It works from what you wrote, so a vague request gets a vague skeleton.
Reach for it when you are starting out, inheriting a canvas, or facing a screen of fields you would rather not fill in by hand.
The comparison, side by side
| AI Prompt | AI Agent | AI Data Transform | |
|---|---|---|---|
| Model calls | One per run | As many as it decides | One per instruction |
| Cost | Lowest | Highest | Low |
| Predictable | Yes | Less so | Yes |
| Acts in your apps | No | Yes | No |
| Picks its model | No | Yes | No |
| Remembers | No | Optional | No |
| Main limit | One answer | Time and cost vary | Three instructions |
Two habits worth having
Start simpler than you think you need. Try the prompt first. If it gets you eighty per cent of the way, the remaining twenty is usually a better prompt rather than a bigger step.
Test before you trust. Run the step on five awkward real cases: the email with no signature, the form where somebody typed “n/a”. Read the output rather than the green tick. See Try it before you trust it.