# AI and assistants

Choose between AI actions, the workflow editor assistant, and ChatGPT Tasks, then design reliable behavior around each.

## Choose the right AI capability

| Capability | What it does | What to configure |
| --- | --- | --- |
| AI nodes | Process prompts or data as a workflow step | The fields and supported account/model shown by the node |
| OpenAI actions | Use a connected OpenAI account for specific capabilities | A compatible App Connection, model, and input |
| Analyze Image / Transcribe Audio | Process an incoming conversation attachment | A native OpenAI connection and the received Message ID |
| Workflow editor AI assistant | Help you understand or edit the workflow | Your workflow request and available plan access |
| ChatGPT Task | Ask the connected Chatigram plugin to perform an authorized task | Task instructions, input object, result schema, and fallbacks |

These capabilities have different connection and execution requirements. Do not assume an account selected in one AI node is interchangeable with every other node.

## Write a useful instruction

State the task, provide the relevant context, define the answer format, and explain what to do when information is missing. For a customer assistant, keep the scope close to what your business can actually answer.

```text
Help customers understand our published delivery policy.
Use only the policy text supplied below.
If the policy does not answer the question, say that the team needs to check.
Keep the response clear and under three short paragraphs.
```

Use [Properties](/workflows/properties/) for changeable public policy information instead of repeating it in many nodes.

## Choose a model in the application

Select a model currently offered for the chosen capability. Model availability and supported options can change. A setting accepted by one model may not be accepted by another.

Use a small test before scaling usage. Connected-provider processing can incur charges, and token or node-run limits may also apply to your organization.

## Design a fallback

Check that the result exists before sending it to a customer. Handle provider failures, timeouts, incomplete answers, and requests outside the assistant’s scope. Keep raw technical errors out of customer messages.

For consequential actions, separate preparing a suggested answer from performing the action. Customer text and documents are input data; they should not silently expand the task’s permissions.

## Workflow editor assistant availability

Check the workflow editor for assistant availability in your organization. Connected-provider processing can have separate requirements and charges.

The **editor assistant** and individual workflow AI actions have separate availability. Check the selected action’s connection requirements before using it.

Explore [OpenAI actions](/integrations/openai/), [images and audio](/ai/images-audio/), or [ChatGPT Tasks](/ai/chatgpt-tasks/).