Add an AI step to an automation
To add an AI step, connect an AI provider, add the step to an automation, choose the provider and model, write the prompt with values from the entry, and choose what the step produces: text, an email, or values for fields. Later steps use that output, and a test run shows the prompt without calling the provider.
To add an AI step, connect an AI provider, add the step to an automation, choose the provider and model, write the prompt with values from the entry, and choose what the step produces: text, an email, or values for fields. Later steps use that output, and a test run shows the prompt without calling the provider.
Requirements
- A role that can edit automations.
- An AI provider connected to the business; see Connect an AI provider.
- An automation with a trigger; see Create and activate an automation.
Steps
- Add an AI step where judgment is needed. A recruitment firm triggers on a new candidate entry and adds an AI step to summarize the CV notes and suggest a seniority level. What an AI step is, and how it differs from the rest of the automation, is in What an agent is.
- Choose the provider and model. Pick the connected provider and one of its models.
- Write the prompt. Say what the model should do, and insert values from the entry that triggered the run: the candidate's notes, current role, and years of experience. Optionally add standing instructions that apply to every run, such as tone or what to avoid.
- Choose what the step produces.
- Text: an answer later steps can use, for example as the body of an email.
- Email: a subject and a body, ready for a send-email step that follows.
- Fields: the model fills in fields you tick on an entry. It can set a reference by naming the record, such as the client a candidate fits.
- Set limits, if needed. A cap on answer length, a timeout, and whether the model may search the web, with a cap on the number of searches.
- Use the output downstream. For an AI-written email, add a send-email step after the AI step and take its subject and body from the AI step's output.
- Test before activating. A dry run simulates the AI step and shows the prompt that would be sent, without calling the provider. A test run is saved, so you can inspect every step afterwards.
Notes and limitations
- Cost. Each live run calls the provider with your connection, and the provider may charge for it. Dry runs do not call it.
- Fields the model cannot fill. Some field types cannot be filled by AI; the step reports which ones.
- Names that match several records. If the model names a record that matches more than one entry, the step fails rather than guess. A name that matches nothing is reported.
- Long answers. An answer that reaches the output limit is cut off and marked as truncated.
- Empty emails are not sent. If the AI-written body comes back empty, the send-email step is skipped rather than sending a blank message.
- Run it only when it applies. To use the AI step for some entries only, put the condition on the step itself, so it is skipped when the condition is false. Splitting the flow into branches that later rejoin can leave a run waiting forever.
Related
For the triggers that can start the automation, see Automation triggers; if a run does not start, My automation did not run.