The workflow is yours. AI handles the interpretation.

Turn unstructured input into a process you control.

Classify an email. Extract the facts. Look up a record. Return a useful result. Build it as a typed Python workflow with explicit routes and clear review outcomes.

Build your first workflow Follow the support-email tutorial

Async PythonConfiguration as filesEvaluations included

AI interprets. Your application decides what happens next.

Emails and documents contain requests, uncertainty, and conflicting information. Use a model for the parts that need interpretation. Keep sequence, permissions, routing, and final disposition in your authored process and application code.

Explore an authored processIllustration · no model calls

Choose an input to see the intended route. These are reviewed illustrative outcomes, not live predictions.

A duplicate charge Cancel a renewal An unclear request

“Please review the duplicate charge on invoice INV-7 for account A-100.”

1. 01Interpretbilling
2. 02Follow a ruleBilling flow
3. 03Return a resultDraft for an agent to review

The message explicitly identifies a duplicate charge. The workflow selects the billing flow; nothing is sent or changed automatically.

Unresolved input is a normal business outcome. A valid assessment can say there is not enough evidence, information conflicts, or several categories fit. Your process can return needs_review instead of forcing a guess.

Deterministic routing means the same validated result follows the same rules. It does not make a model deterministic or correct. Measure that with reviewed evaluations.

Find the capability you need

DECIDE

Make a bounded decision

Yes or no, one category, several labels, or an urgency level. Get a typed answer with a reason and evidence strength.

EXTRACT

Turn content into useful data

Extract JSON fields or separate distinct requests from a message. Validate the shape before the next operation uses it.

CONNECT

Use tools with clear boundaries

Call a known MCP tool, let a model choose tools, or run a Python function. Keep tool access explicit.

DELIVER

Put the result in your application

Embed in Python, serve an HTTP endpoint, and handle review or failure. Your host owns the final action.

Three building blocks

You define	Its responsibility	Support-email example
Workflow	The complete process: input, routes, and final output	Handle an incoming email
Flow	A sequence of related operations	Classify, or prepare a billing response
Step	One operation with selected input and a validated result	Extract an account reference

Start with one workflow, one flow, and one step. Add a branch or tool only when your process needs it. Configuration lives in readable files; Python owns integration and deterministic business functions.

AI interpretation inside an explicit process. Drafting is separate from sending or changing an account.

Understand workflows, flows, and steps →

From first request to a deployed application

1. Install and run — start manually or with Claude or Codex.
2. Define the process — learn the folder layout, data bindings, and routes.
3. Add a capability — choose the task by the result you need.
4. Connect your provider — local models, OpenAI, Azure, Anthropic, Google, or Bedrock.
5. Evaluate real requirements — create reviewed cases and inspect disagreements.
6. Integrate and deploy — handle requests, results, errors, secrets, and operational limits.

The Guide explains each task and its options. Tutorials build a support-email assistant step by step. Evaluation helps you measure and improve it. Detailed field and CLI lookups sit inside Guide when you need an exact contract.

A library inside your application

Foliqant runs in your Python process. Your application owns inbound transport, identity verification, durable jobs, result storage, and any business writes. The core does not install a queue or a database.