> For the complete documentation index, see [llms.txt](https://docs.siit.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.siit.io/workflow/ai-actions.md).

# AI Actions

Let AI read the request inside a workflow: summarize or translate the conversation, route the workflow on a question, pick a priority, status, inbox, tags, service or applications, or write the reply

#### Why it matters

Some workflow steps need judgement rather than a fixed rule: is the employee blocked, which inbox should own this, what should the first reply say? Conditions on fields can't answer those questions, so the request waits for an agent to read it.

AI actions put that judgement inside the workflow. The AI reads the request when the workflow runs and acts on what it finds, with the same audit trail as any other step.

AI actions are available on the Trial and Pro plans. Your plan decides: a workspace that moves to another plan loses them at its plan change.

{% embed url="<https://www.youtube.com/watch?v=0D__CXtR95U>" %}

#### Where to find them

Open the action picker in the workflow editor. **AI actions** is the first group in the list.

The group holds two kinds of steps:

* **AI modes on existing actions**: **Set request status**, **Set priority**, **Assign to inbox**, **Associate to service**, **Associate to applications** and **Tags** run in **AI decides** mode, and **Quick reply** runs in **AI writes it** mode.
* **AI-only actions**: **Summarize thread**, **Translate** and **AI routing**.

AI actions are offered in [request](https://docs.siit.io/workflow/request-triggers.md) and [app access](https://docs.siit.io/workflow/app-access-triggers.md) workflows. **Summarize thread** is offered in request workflows only.

On a plan without AI actions, the cards still show in the picker, but they are disabled and link to **Upgrade plan**. The classic versions of each action stay available.

#### Summarize thread

Summarizes the conversation in a few lines and posts it as an internal note headed **AI summary**. There is nothing to configure.

The summary is written in the language of the conversation. It is also available as an output, so you can reuse it in a later step, for example in a Slack message to the assignee or as the text of a **Translate** step.

#### Translate

Translates a text and posts the translation as an internal note headed **AI translation**.

* **Text to translate**: **Description**, **Whole conversation**, **Message content** (the message that triggered the workflow, on reply triggers), or the text output of an earlier step, such as a summary.
* **Translate into**: Arabic, English (UK), English (US), French, German, Hebrew, Italian, Japanese, Portuguese, Romanian, Spanish or Swedish.

Internal notes are never translated. The text can be up to 20,000 characters; for the whole conversation, the oldest messages are left out first.

When the text is already in the target language, no note is posted. The step still returns the text as its **Translation** output, so later steps always have something to use.

#### AI routing

Asks the AI a question about the request and continues the workflow in the branch that matches its answer. Use it where a [branch condition](https://docs.siit.io/workflow/branching-and-approvals.md) can't express the rule, because the answer lives in what the employee wrote.

{% stepper %}
{% step %}

#### Write the routing question

Ask one question with a clear answer, for example *Does the employee say they cannot work at all?* The question can be up to 2,000 characters.
{% endstep %}

{% step %}

#### Choose what the AI returns

* **Yes or no**: two branches, one for each answer.
* **A value from a list**: between 2 and 50 branches. Give each branch a unique name and describe, under **When AI picks it**, when the AI should choose it. For example, *Blocked* when the employee cannot work at all, *Slowed down* when they can still work but a tool or task is affected, *Not impacted* when the request is a question or a non-urgent ask.
  {% endstep %}

{% step %}

#### Add actions in each branch

The AI always returns exactly one answer, so exactly one branch runs. Steps placed after an AI routing step belong to a branch; there is no shared continuation after the branches.
{% endstep %}
{% endstepper %}

#### AI decides

**Set priority**, **Set request status**, **Assign to inbox**, **Tags**, **Associate to service** and **Associate to applications** have two modes: **Choose a value**, the classic step, and **AI decides**. In **AI decides**, the AI reads the request and picks the value.

* **Restrict allowed values**: limit what the AI can pick from, for example only the inboxes of your IT team.
* **Guidance (optional)**: tell the AI how to choose, for example *Use Urgent only when several employees are blocked*.

Without a restriction, the AI considers the first 250 eligible values ordered by name. Restrict the allowed values when the one you want sits beyond that group.

The AI picks one value for priority, status, inbox and service, and can add several tags or applications. For services and applications, the AI also reads their description, so a clear description helps it choose.

{% hint style="info" %}
When the AI is not sure, nothing changes: the field keeps its current value. If an agent changes the same field while the AI is deciding, the agent's change is kept. Tags and applications are only ever added, never removed.
{% endhint %}

#### AI writes it

**Quick reply** has two modes: **Write the message**, the classic step, and **AI writes it**. In **AI writes it**, you give **Instructions** and the AI writes a new reply for each request, then sends it to the employee when the workflow runs.

For example: *Acknowledge the request, say which team will handle it, and ask for the device's serial number if it is missing.*

* The reply is written in the requester's language, unless your **Instructions** ask for another one.
* The AI never sees internal notes when it writes a reply, since the reply goes to the employee.
* If the AI can't write a reply, the employee still gets a short acknowledgement, so the step never leaves them without an answer.
* Switching between the two modes clears the settings of the other mode, including the fixed message and its attachments.

#### What the AI reads

The AI works from the request itself: its fields, its latest messages and internal notes, and the names of the people in the conversation. In app access workflows, it also reads the access request: the application, the role, the status and the business reason.

Attachments and deleted messages are not sent to the AI. **Translate** only sends the text you chose.

The AI doesn't call your integrations or take any action beyond the step it runs in.

#### Following what happened

AI actions run in the background. The step waits for the AI's answer, then the workflow resumes.

The workflow's run history shows what each AI step did: the summary, the translation, the value it applied, the reply as sent, or the branch it took. **AI decides** and **AI routing** also log the AI's reason and confidence, and say when nothing was applied and why.

When an AI step fails, for example because AI actions are no longer on your plan, the run history shows the step in error with the reason.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://docs.siit.io/workflow/ai-actions.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
