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

# Actions

Actions let Aissist's AI look up, create and update records in your business systems during a conversation.

*Last updated: October 25, 2026*

Actions let the AI work in your systems, not just talk about them: check an order, look up a subscription, create a refund ticket, cancel an unshipped order. This page explains what actions are, how the AI decides to call one, and how actions that change data are kept safe. For the steps, see [Add Actions](https://doc.aissist.io/tutorial/add-actions.md).

### What an action is

An action is one operation in one of your systems, run through a [connector](https://doc.aissist.io/capabilities/connectors.md). You connect the system once; each action uses that connection. For example, connect Shopify, then add actions such as "look up an order by number" or "cancel an unshipped order".

Actions fall into three groups:

* **Look up** - read data the AI needs to answer: order status, tracking, plan and renewal date, stock.
* **Create** - create a ticket in Zendesk, Intercom or Freshdesk, or a record in your own system through its API.
* **Update** - cancel an order or change an address, in a system that offers it.

You find them under **Workspace → Actions**. Each action shows its **Connector**, **Kind**, **Direction** and the sub agents that use it. **Kind** tells you where it came from: **Linked** (a ready-made action from the connector's catalog), **Instanced** (a catalog action with your own wording or scenario), or **Customized** (one you defined yourself).

### How the AI uses an action

Each action has a **Trigger Scenario**: when to call it, in plain words. During a conversation the AI decides whether the scenario applies, fills in the parameters from what the customer said, calls the action and reads the result.

* **Run at interaction** - the AI may call it while replying to a message.
* **Run at start** - it runs when a conversation starts, for example to load the customer's account.
* **Response Processing** - a **Summarization** and an **Instruction** tell the AI how to read the result and what to tell the customer.

Link an action to a [sub agent](https://doc.aissist.io/capabilities/sub-agents.md) to make it available only in that scenario. An action linked to no sub agent is available everywhere in the workspace.

### Safety for actions that change data

An action that creates, updates, cancels, refunds or deletes something changes an outcome for the customer. Aissist treats these differently from look-ups:

* **Extra checks before it runs.** The decision to call it is made twice, independently. It runs only when both agree on the call and its details. When they disagree, a further check decides; it may ask the customer to confirm the details instead, or not call it at all.
* **Never added automatically.** Aissist can work out on its own which look-up actions a sub agent is likely to need. Actions that change data are never added that way: they are available only where you put them, on a sub agent or workspace-wide.
* **Outbound actions use only what you attached.** Actions that run on the AI's own reply (outbound), rather than on the customer's request, are chosen only from those attached to the workspace or the active sub agents.

See [Reliability](https://doc.aissist.io/capabilities/reliability.md) and [Reliable AI](https://aissist.io/technology/reliable-ai).

### Test with the action debugger

On **Workspace → Actions**, click **Debug** to open the **Actions Debug Studio**. Type a message that should trigger an action. You see which action was chosen, the parameters it was given, and what came back.

* **Not triggered:** make the **Trigger Scenario** and description match how customers ask; check it is linked to the right sub agent.
* **Wrong parameters:** improve the parameter names and descriptions; give formats, examples and allowed values.
* **Hard-to-use result:** refine **Response Processing**, and return only the fields the AI needs.

Then test the whole conversation in [Try](https://doc.aissist.io/tutorial/try.md). The debugger runs the action, so test actions that change data with test records.

### With Ask Engine

Click **Add Action** and describe what it should do:

```
Look up an order's shipping status by order number when a customer asks where their parcel is
```

```
Create a Zendesk ticket of type Refund, with the order number, when a customer asks for a refund
```

Ask Engine finds the connector that can do it, picks a ready-made action from its catalog or writes a custom one, and suggests the sub agents that should use it. You review the name, trigger, parameters and sub agents, then create it. In **Ask** you can also check a look-up live: "Call the order lookup for order 1001 and show me what it returns." Ask Engine runs actions that only read. It does not run actions that would change something.

### Best practice

* **Start with look-ups.** They are the safest and resolve the most "where is my order" questions.
* **Write the trigger as the customer's need,** not the API's name.
* **Link change actions to one sub agent** whose instruction collects and confirms the details first.
* **Test each action with several realistic messages** in the debugger, then in Try.

### Related

* [Add Actions](https://doc.aissist.io/tutorial/add-actions.md)
* [Connectors](https://doc.aissist.io/capabilities/connectors.md)
* [Sub Agents](https://doc.aissist.io/capabilities/sub-agents.md)
* [How do I allowlist Aissist's IP addresses?](https://doc.aissist.io/faq.md) - when a custom action calls an API behind an IP firewall
