Safeguarding detection monitors call recordings for concerning topics like self-harm or inappropriate content. You can customize which keywords trigger alerts and which to ignore, reducing false positives for your organization's specific needs.

## Prerequisites

- Your account must have the Conversational Analysis feature enabled
- You must have the Safeguarding Config role assigned to your user account
- Call recordings and safeguarding analysis must be enabled in Account Settings

## Steps

### Step 1: Navigate to Account Settings

Click on Settings in the main navigation menu, then select Account Settings from the sidebar. Scroll down to the 'Enable conversational analysis' section. Make sure the conversational analysis toggle is turned on, and that the 'Safeguarding analysis' checkbox is checked. If these options are disabled, you'll need to enable them first before customizing detection keywords.

![Navigate to Account Settings](https://help.together.ly/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6OTgsInB1ciI6ImJsb2JfaWQifX0=--7c62a438405eb1526d909ae858e25dbf06c523e3/step_0.png)

### Step 2: Locate the safeguarding customization section

Below the 'Safeguarding analysis' checkbox, you'll see a section titled 'Customise Safeguarding Alerts' with two text areas: 'Include alerts for' and 'Exclude alerts for'. This section will only appear if your account has the safeguarding configuration role enabled. The explanatory text tells you that leaving the Include field blank will use the default detection categories (self harm, PII sharing, incidents, and inappropriate content).

### Step 3: Add custom keywords to include

In the 'Include alerts for' text area, enter specific topics or keywords you want the AI to monitor. When you fill this field, it replaces the default safeguarding categories entirely. Be specific and separate multiple items with commas. For example, you might enter: 'mentions of drug use, financial exploitation, signs of hoarding, discussions of abuse or neglect'. You can use up to 500 characters. This field tells the AI exactly what to flag as concerning.

### Step 4: Add keywords to exclude

In the 'Exclude alerts for' text area, list topics that should never trigger safeguarding alerts. This helps reduce false positives by telling the AI to ignore certain conversations that might otherwise be flagged. For example: 'exchange of phone numbers between befrienders, general loneliness, historical health discussions, mentions of past resolved issues'. The AI will treat these topics as normal conversation even if they might match other detection patterns. You can use up to 500 characters for exclusions.

### Step 5: Save your changes

After entering your custom keywords, scroll to the bottom of the settings page and click the 'Save Settings' button. Your safeguarding detection rules will be applied to all future call recordings that are processed by the AI. Existing recordings won't be re-analyzed, only new recordings will use the updated configuration. If you need to test your changes, conduct a test call and check the recording analysis to verify the keywords work as expected.

## Tips

- Leave the 'Include' field empty to use the default safeguarding categories (self harm, PII sharing, incidents, inappropriate content) - only fill it if you need different detection rules.
- Be specific with exclusions to avoid over-blocking legitimate safeguarding concerns. For example, exclude 'historical health discussions' rather than just 'health'.
- Review your first few analyzed calls after changing keywords to ensure you're catching the right issues and not getting too many false positives.

## Summary

By customizing safeguarding keywords, you can tailor the AI detection to your organization's specific concerns while filtering out common false positives. Remember that any changes only affect future recordings, so test your configuration with a practice call to ensure it works as intended.