Alta Flows ship with a built-in AI integration — a set of five no-auth, no-setup AI actions you can drop into any flow to analyze, transform, or decide based on the data flowing through. No API key, no provider account, no separate connector. The credits run on Alta.

This article walks through all five actions, the fields you'll see in the step settings, and the patterns that get the most out of them.

Who this is for

Anyone building or editing an Alta Flow who needs the flow to read, classify, summarize, or reason over text — for example, scoring inbound replies before routing them, pulling structured fields out of a free-text email, summarizing a long meeting note before pushing it to a CRM, or running a multi-step agent that calls other tools.

Before you start

You need access to Flows in Alta (left nav → Flows).
You need at least one step before the AI action whose output you want to feed in — a trigger, or any earlier step.
No external auth needed. The AI integration uses Alta's models and runs on your workspace credits.

Add an AI action to a flow

Open the flow in the builder, or create a new one.
Click the + between any two steps to add a step.
In the step picker, choose the AI integration.
Pick one of the five actions: Ask AI, Summarize Text, Classify Text, Extract Structured Data, or Run Agent.
Fill in the fields described below.
Click Test step to dry-run with sample data from earlier steps.
When the whole flow is ready, hit Publish.

Every AI action lets you pick a Model from this list: GPT-5 Mini, GPT-5.4 Nano, GPT-5.4 Mini, Gemini 2.5 Flash, Gemini 3 Flash, Gemini 3.1 Flash Lite, Claude Haiku 4.5. These are the fast, affordable models tuned for inline-flow work — pick the lighter ones (Nano, Haiku, Flash Lite) for high-volume classification and extraction, and the larger ones (5.4 Mini, Gemini 3 Flash) when reasoning quality matters more than speed.

1. Ask AI

A flexible AI step. Ask it to analyze data, explain, draft, or decide based on your flow's data.

Fields

Prompt (required) — long-text instruction. Reference earlier-step values with the variable picker.
Model (required) — see model list above.
Creativity (optional, default 100) — controls how creative vs. deterministic the response is. Lower values (e.g. 20) make output more predictable; higher values (e.g. 150) make it more varied.
Max Tokens (optional, default 2000) — caps the response length. Bump it up if you're asking for long-form output.
Web Search (optional, default off) — when on, the AI can search the web for current information.
Web Search Options — appears once Web Search is on. Lets you cap searches per run (Max Web Search Uses, default 5), toggle Include Sources to return the URLs the AI consulted, and (for Anthropic models) set Allowed Domains / Blocked Domains and a User Location for localized results.

Returns the AI's text response. If Include Sources is on, returns { text, sources } instead.

Use it for: reply triage ("Is this an objection, interest, or out-of-office?"), drafting a custom note before pushing to CRM, deciding which branch of the flow to take based on free-text context.

2. Summarize Text

Summarize long emails, articles, or documents into what matters.

Fields

Text (required) — the long-form input to summarize. Map it from an earlier step.
Prompt (required, with default) — the instruction the model follows. Default: "Summarize the following text in a clear and concise manner, capturing the key points and main ideas while keeping the summary brief and informative." Override it to change tone or length (e.g. "Summarize as three bullet points for a sales rep").
Model (required).
Max Tokens (optional, default 2000).

Returns the summary as plain text.

Use it for: compressing a long inbound reply before logging it to your CRM, condensing call transcripts, turning a long article into a one-paragraph summary for a campaign signal.

3. Classify Text

Categorize any text input using custom labels, so your flow knows what to do next.

Fields

Text to Classify (required) — the input.
Model (required).
Categories (required) — an array of category names. Add one entry per category. The AI is forced to return exactly one of these.

Returns the chosen category as a single string. If the model can't fit the text into any of your categories, the step fails with: Unable to classify the text into the provided categories.

Use it for: branching logic ("Interested" / "Not now" / "Wrong person" / "Unsubscribe"), tagging support tickets by topic, routing a reply to a specific Slack channel.

Tip: Pair this with a Router step downstream, with one branch per category. Keep your category list short and mutually exclusive — 3 to 6 options works best.

4. Extract Structured Data

Accurately pull names, amounts, and other structured data from emails, invoices, and scanned documents.

Fields

Text — the unstructured input. Optional individually, but you must provide either text or the future image/PDF input (text-only today).
Guide Prompt (optional, default "Extract the following data from the provided data.") — extra context for the AI on what to look for.
Model (required).
Data Schema Type (required, default Simple) — choose Simple for a no-code field list, or Advanced if you need a full JSON Schema.
Data Definition — in Simple mode, you add fields one by one with Name, Description, Data Type (Text, Number, or Boolean), and a Fail if Not present? checkbox. In Advanced mode, you paste a JSON Schema object directly.
Max Tokens (optional, default 2000).

Returns a JSON object with one key per field you defined.

Use it for: pulling email, company_name, budget out of an inbound message; extracting line items from a forwarded invoice; turning a free-text intent reply into { wantsDemo: true, timing: "next quarter" }.

Tips and pitfalls

Field names are sanitized — non-alphanumeric characters (besides _, ., -) are replaced with _ before being sent to the model, then restored to your original name in the output. Stick to plain names like budget or contact_email to avoid surprises.
Use the Description field on each property generously — a one-line hint like "the dollar amount the prospect mentioned, as a number" dramatically improves accuracy.
Mark a field as Fail if Not present? only when the downstream step truly can't tolerate a missing value. Otherwise, leave it off and let the field come back empty.
If the model can't produce a valid extraction at all, the step fails with: Failed to extract structured data: ....

5. Run Agent

Handles complex, multi-step tasks by reasoning through problems, using tools accurately, and iterating until the job is done.

Run Agent is the heaviest action — use it when a single LLM call isn't enough and the task needs the model to plan, call tools, and check its work.

Fields

Prompt (required) — describe what you want the assistant to do.
Model (required).
Agent Tools (optional) — array of tools the agent can call. Each entry has Tool Type, Tool Name, optional Flow External ID (to call another Alta flow as a tool), optional MCP Server URL + Protocol + Auth Configuration (to bring in an external MCP server).
Structured Output (optional) — array of Display Name / Description / Type entries that force the agent to return a structured object instead of free text.
Max steps (required, default 20) — the maximum number of reasoning / tool-call iterations the agent is allowed before it stops.
Web Search (optional, default off) and Web Search Options — same as Ask AI, with the exception that Include Sources isn't available here.

Returns a structured agent output with the final text, the tool-call timeline, and a status (completed or failed).

Tips and pitfalls

If you don't need multi-step reasoning or tool use, prefer Ask AI — it's cheaper and faster.
Keep Max steps conservative. The default 20 is plenty for most tasks; raising it past 30 burns credits quickly on dead-end paths.
If the agent errors out partway through a tool call, the step returns the partial output with a failed status and an Errors encountered block — useful for debugging in Runs.

Common pitfalls across all AI actions

Variable references must be mapped, not pasted. Use the variable picker in each field to reference earlier steps — pasting {{step1.output}} as raw text won't resolve.
Test step before publishing. Each action runs against real sample data and shows you the returned shape, which you'll want to know before mapping downstream steps.
Watch your credits on Run Agent. Multi-step agents with web search and tool use are the most expensive AI actions in Flows. Check Settings → Credit Usage after your first few production runs.
Pick the right model for the job. Classification and short extraction work fine on the Nano / Flash Lite / Haiku tier. Save the 5.4 Mini and Gemini 3 Flash models for Ask AI prompts that need reasoning quality.