> ## Documentation Index
> Fetch the complete documentation index at: https://docs.introw.io/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Approvals

> Enable AI validation on a form, write or generate review rules, set a certainty threshold, and let the AI accept, decline, or return submissions.

## Where it lives

AI Approvals sits under **Portal**, at [Forms](https://app.introw.io/forms).

<Frame>
  <img src="https://assets.introw.io/docs/features/ai/approvals/shots/gate.png" alt="The Approval Gate on a form's automations, which holds a submission for a human decision before any automation runs on it." />
</Frame>

## Before you start

| You need                     | Why                            | Fix it                                                                            |
| ---------------------------- | ------------------------------ | --------------------------------------------------------------------------------- |
| The AI Agent module          | On by default on your plan     | **Request access**                                                                |
| A form with an approval step | The AI recommends on that step | [Build a form](/features/forms/form-builder/guides/build-and-publish-a-form)      |
| Write access to the form     | To turn the recommendation on  | [Internal roles](/features/access/team-management/guides/create-an-internal-role) |

## How it works

AI Approvals is configured per form, on the form's **automation** under manual acceptance. Toggle **Enable Introw AI validation** and the AI reviewer joins that form's [approval workflow](/features/forms/submissions-approvals). You give it **instructions** - the rules a submission has to meet, written in plain language - and Introw can generate a first draft of those instructions from the form's own fields, which you then edit.

When a partner submits, the AI reads the submission and its attachments, checks it against your instructions, and produces a verdict with reasoning: accept, decline, or return for more information. By default it **guides** - the verdict and reasoning show up for your human approvers, who make the call. Turn on **auto-execution** and set a **certainty threshold** (defaults to 90) and the AI acts on its own when its confidence is at or above that bar, and escalates anything below it to your [acceptance steps](/features/forms/submissions-approvals). Because it runs on the form layer, the same reviewer covers deal registrations, [MDF](/features/mdf) requests, claims, ROI, and invoices, and shared leads.

### Guide vs. auto-execute

* **Guide (default)** - the AI assesses every submission and posts its verdict and reasoning; a human still accepts, declines, or returns.
* **Auto-execute** - above the certainty threshold the AI accepts, declines, or returns on its own; below it, the submission escalates to a human. This is how a stack of routine, well-documented submissions clears without sitting in a queue.

### Where it fits the approval flow

AI validation runs first, then your sequential human acceptance steps handle whatever escalates. [Channel conflict](/features/ai/channel-conflict) screening runs alongside it as the conflict-specific check on deal registrations.

```mermaid theme={"theme":{"light":"github-light","dark":"github-dark"}}
flowchart LR
  submit["Submission arrives"] --> ai["Introw AI validation"]
  ai -->|above threshold| auto["Auto accept / decline / return"]
  ai -->|below threshold| human["Human acceptance steps"]
```

## Settings & configuration

Configured on the form editor's automation tab, under manual acceptance.

### Enable Introw AI validation

Turn on validation for the form so the AI reviews its submissions.

### Instructions

Write the rules the AI checks each submission against, or generate a draft from the form's fields and edit it.

### Auto-execution

Let the AI act on its own above the certainty threshold; leave it off to keep the AI advisory while humans decide.

### Certainty threshold

Set how confident the AI must be to act without a human (defaults to 90). Lower it to automate more, raise it to escalate more.

## How-to guides

*None yet.*

## Troubleshooting

<Warning>
  Auto-execution acts on real submissions - accepting an MDF claim or a deal registration writes to your CRM and can draw down budget. Start advisory, review the AI's verdicts, and only raise auto-execution once its instructions are proven. The AI reviews against your instructions only, so vague rules produce vague verdicts.
</Warning>

<AccordionGroup>
  <Accordion title="The AI isn't reviewing submissions">
    Validation isn't enabled on that form, or the form has no approval workflow.
  </Accordion>

  <Accordion title="It escalates everything">
    The certainty threshold is too high for the instructions, or the instructions are ambiguous.
  </Accordion>

  <Accordion title="It auto-accepted something it shouldn't">
    Tighten the instructions and raise the threshold; auto-execution follows the rules you gave it.
  </Accordion>
</AccordionGroup>