AI in workflows

Who this is for: admins, approvers and security reviewers whose company uses, or may use, an AI model in its workflows.

Before you start: nothing. To add an AI step, see Connect an AI model.

A workflow can ask your own AI model to classify, summarise or extract data. That is your choice, and Liitora doesn’t block it. What Liitora does is make that choice visible, have a person confirm it, and let your company set limits.

Where AI steps come from

A step counts as an AI step when it sends data to an AI model:

it uses the built-in action ai.generate;
it uses a connection made from an AI template (OpenAI, Azure OpenAI, Anthropic, Mistral or an OpenAI-compatible model);
it calls the address of a known AI provider in another way, for example with http.request.

The model, the account and the API key are yours. Liitora runs no model of its own.

Where you see it

A Uses AI mark shows up in these places:

the Workflows list, next to the status of a workflow whose newest or active version has an AI step;
the Runs list and the run page, for a run of a version with an AI step, and for a run that called an AI model in another way;
the action Library, on actions that call an AI model;
the approval page, with a section that lists the AI steps.

What the approver confirms

On the approval page, AI in this workflow lists each AI step:

the AI provider and the connection it uses;
what it Sends: the data the step passes to the model, for example trigger.body or steps.fetch.outputs.rows.

Below the list is a short note. It says that AI answers can differ between runs and that the data the workflow reads can influence them. To approve, tick I have read this and confirm the AI use. Liitora refuses the approval without the tick, also over the API, and the audit log records it.

When an AI answer decides what happens

Some inputs decide where data goes or what is changed. We call these write targets:

the recipients of an email (to, cc, bcc);
the address of a web request or a webhook (url, and the request method);
a file path (path, and from and to when a file is moved);
a stored state key (key);
every input of an action, script or called workflow that can change data in another system.

If an AI answer reaches one of these, for example the model picks who gets an email, the approval page lists it under AI answers that decide what is sent or changed. The approver ticks a second box to confirm it. When you save the workflow, it also shows a warning (ai_output_to_write_target) with the place in the YAML.

This also covers an AI answer that passes through the workflow’s saved state. If a step saves an AI answer with platform/state.set, or a script or action that got one could save it, then every value read back from state counts as an AI answer too. That includes values read in a later run.

An answer that can only be one of the values you wrote doesn’t count. For example, a JSON Schema field with enum: [it, hr] can only be it or hr. Use a fixed list like that when you can.

Turn AI off

AI steps are allowed by default. An admin can turn them off for the whole company:

Open Settings, then Members.
Under AI, clear Allow AI steps.

With AI off:

workflows with AI steps can’t be activated, started, retried or replayed (error code ai_disabled);
a run of such a workflow that was already waiting fails before its first step;
every call to an AI model fails, wherever it comes from.

You can still save, test (with mocked answers) and approve workflows with AI steps, so they are ready when AI is turned back on.

Strict mode

Strict mode is off by default. We recommend it for banks and other regulated companies. An admin turns it on under Settings, Members, AI, with Strict mode.

With strict mode on, an AI answer may not decide a write target. A workflow where it does:

can’t be saved: the warning becomes an error;
can’t be approved or activated if it was saved before (error code ai_strict_mode).

AI can still pick from a fixed list you wrote, summarise for a person, or fill fields that don’t decide where data goes.

Workflows that are already active keep running when you turn strict mode on. Their next version meets the new rule. To stop AI at once, turn AI off instead.

Only admins change these settings. Everyone can see them, and every change is in the audit log.

Your data and your AI provider

When a workflow sends data to your AI model, it does so on your company’s instruction, to your company’s own provider. That provider is your processor, not ours. Check your agreement with the provider and where it processes data. See Connect an AI model.