AI Workflows
AI outputs your workflow can actually use.
Most workflow tools have an LLM step: prompt in, text out. The catch is the text. The next step can't act on a prose paragraph without parsing it first. Formt's AI steps produce typed, structured output shaped to what the next step needs. And any AI step can pause for your review before its work goes anywhere, so the monotonous parts run themselves and the judgment stays yours.
AI Generate · Text mode
Prose for humans to read, or for another step to format.
Text mode produces one written output as plain text, HTML, or Markdown. Reach for it when a person reads the result directly, or when it feeds a formatting step like PDF Generate or a Slack message.
Writing an email? Email mode generates a subject and body as separate outputs, with purpose-specific prompting. It lives on the Email Platform page.
The team moved $340K from proposal to negotiation, led by the Meridian Accounting deal advancing after three weeks stalled. Two deals slipped back to discovery on pricing objections…
→ flows into PDF Generate or Slack
AI Generate · Structured mode
Typed fields the next step uses immediately.
In Structured mode you define the exact output fields: named, typed, and optionally constrained. The model fills them from your prompt and context, and the output arrives as data, not prose. The next step reads each field directly.
Oversight
Hand off the monotonous work. Keep exactly as much oversight as you want.
The point of AI in a workflow isn't replacing your judgment. It's not spending it on the hundredth identical record. You decide which steps need review, and Formt pauses each one at the right moment, so a single workflow can run some steps automatically and hold others for your sign-off.
Let it run
Enrichment, categorisation, summaries, formatting: high-volume work where a miss is cheap and the value is in not doing it yourself. The step runs, the run log records everything it did, and you look when you want to.
Check what it produced
The AI step runs, then the workflow pauses with the actual output in front of you. Approve it, override a field, or reject it before anything downstream depends on it. Formt does this for generated content and scoring that feeds real records.
Approve the inputs first
Put review on the step that touches the outside world: the send, the CRM write, the post. AI does the drafting, and nothing leaves without your sign-off. When the batch looks right, approve all of it in one action.
Every review lands in the Task Manager, one inbox across all your workflows, with a record of who approved what and when.
See the Task ManagerAt scale
One brief, every record.
Chat AI answers one prompt at a time. Workflow AI takes the brief you wrote once and runs it across everything the workflow finds, each pass working from that record's own data.
The two AI steps
Research Agent gathers. AI Generate produces.
Formt's other AI step, the Research Agent, collects information from the web and returns it as typed, cited fields. AI Generate turns that (or any upstream data) into the outputs your workflow acts on. The two are built to chain.
Account research to CRM-ready scoring
The Research Agent returns a company snapshot as structured fields: overview, funding, hiring signals. AI Generate scores the lead and assigns a tier using a HubSpot-bound picklist. The score lands on the contact record, and the email step writes an opener that references what the research actually found.
Filing analysis to client-ready brief
The filing supplies the numbers. The Research Agent adds what the numbers leave out: recent news, analyst commentary, competitive moves. AI Generate synthesises both into a typed brief (growth, risk factors, outlook), and PDF Generate renders it as a document ready to share.
Built in
The models come with Formt
No API key of your own, no separate usage account to fund and monitor. Every AI step just picks its model from OpenAI and Anthropic: quick ones for high-volume steps, heavyweight reasoning for the calls that decide outcomes.
GPT-4o Mini
Fast and efficient, the right call for most steps: email drafts, summaries, straightforward field extraction.
DefaultGPT-4o
A step up in reasoning for nuanced generation and the extractions that take real reading comprehension.
o1 Mini
A reasoning model for steps with layered logic: scoring rubrics, multi-condition rules, calculations described in words.
Claude Sonnet 4.5
Anthropic's balanced model. Strong structured output and instruction following on complex prompts.
Claude Haiku 4.5
Anthropic's fast model. Built for high-volume steps that need to move quickly through a long list.
GPT-4 Turbo
The previous-generation heavyweight, available where its behaviour is what your prompts were tuned against.
How it compares
More than a prompt step
Most automation tools have an AI step now. The difference is everything around it: what comes out, what it's allowed to say, who signs off, and how it handles a list.
| Capability | Other tools | Formt |
|---|---|---|
| Setup | Bring your own model API key and fund its usage account, or settle for a bundled model you don't choose. | Included. Models from OpenAI and Anthropic, chosen per step, no keys to manage. |
| Output | A block of text for you to parse in a later step. | Typed fields the next step reads directly, or clean prose where prose is the point. |
| Valid values | Whatever the model decides to write. | Picklists and live HubSpot enums the model can't step outside. |
| Oversight | Fully automatic once it runs. Checking the output is another workflow you build. | Review any step, before it acts or after it generates, plus bulk approval and a full run history. |
| Across a list | Loops you assemble and babysit yourself. | For Each runs the step once per record, with that record's own data. |
Use cases
Where structured output earns its keep
Anywhere AI output feeds a system rather than a person: a CRM field, a formula, a report template, a tracker. Typed fields make those handoffs boring, which is the point.
Metrics out of filing text
SEC filings arrive as long documents. AI Generate in Structured mode pulls the numbers and statements you care about into typed fields your formulas and reports can use.
See templateClause data across a contract batch
PDF Extract reads each contract, AI Generate structures the clause data (parties, terms, obligations, red flags), and the batch lands in one tracker instead of forty open tabs.
See templateAccount tiering with CRM-safe values
AI scores and tiers each account, with the tier field bound to your live HubSpot picklist. Every value it writes is one HubSpot accepts, so the write-back never bounces.
See templateWritten reports from raw numbers
AI Generate in Text mode turns a metrics sheet into readable commentary, formatted as HTML that PDF Generate renders into a styled report, delivered on a schedule.
See templateTry it first