# AI Filtering (/docs/sift-ai/ai-filtering)

Sift's AI filtering automatically evaluates every inbound message for relevance and actionability, so your queue only surfaces conversations that actually need a response.

## At a glance [#at-a-glance]

| | |
| --- | --- |
| Auto-evaluates every inbound message for relevance and actionability | ✅ |
| Closes noise: generic praise, off-topic chatter, emoji-only, spam, duplicates | ✅ |
| **Disabled** strategy | ❌ no auto-close, all manual review |
| **Conservative** strategy | ✅ closes only at very high confidence |
| **Standard** strategy (recommended) | ✅ closes low-relevance content at high confidence |
| **Aggressive** strategy | ✅ closes broadly at lower confidence |
| Exempt conversations with the `SiftIgnoreActionability` tag | ✅ |
| Affects historical items | ❌ new inbound only |

{/* screenshot */}
<figure>
  <img src="/assets/screenshots/sift-ai/ai-filtering.png" alt="The AI Filtering settings showing the four filtering strategies for auto-evaluating inbound messages" />
  <figcaption>AI Filtering: choose a strategy to auto-close noise and keep your queue actionable.</figcaption>
</figure>

## How it works [#how-it-works]

Messages that are clearly noise - generic praise, off-topic chatter, emoji-only reactions, or low-signal likes - are closed automatically. Anything that looks like a genuine question, complaint, or request is left open for your team.

## What gets filtered [#what-gets-filtered]

- Generic positive reactions with no actionable content
- Off-topic mentions that reference your brand but need no reply
- Duplicate or near-duplicate messages in a thread
- Spam and bot-generated content

## How sensitivity is controlled [#how-sensitivity-is-controlled]

Sift offers four filtering strategies, configurable per org:

- **Disabled**: no auto-close; every item requires manual review
- **Conservative**: only closes when the AI has very high confidence the message is irrelevant
- **Standard** (recommended) - closes low-relevance and low-actionability content at high confidence
- **Aggressive**: closes more broadly with lower confidence requirements

You can also tag specific conversations with the `SiftIgnoreActionability` tag group to permanently exempt certain message patterns from filtering - useful for VIP accounts or escalation queues where you always want human eyes.

## Auto-close patterns [#auto-close-patterns]

Above the strategy selector, Sift surfaces **learned auto-close patterns** - recurring message shapes it's noticed getting closed the same way over time - so you can review and approve them as an additional, targeted auto-close rule alongside the strategy-driven one.

There's also a separate **Auto-close spam and scam content** toggle under Additional options, distinct from the four strategies above: it targets spam/scam specifically rather than general noise.

## Setting it up [#setting-it-up]

Go to **Settings → AI → Triage**. Choose the strategy that fits your team's volume and risk tolerance. Changes take effect immediately for new inbound messages.

## Beyond auto-close: user rules and signal flags [#beyond-auto-close-user-rules-and-signal-flags]

Triage also holds two related signal-detection tools that don't affect auto-close:

- **User rules** tell Sift what to watch for on people. When a rule matches, Sift tags the user during its nightly analysis so you can filter on it. A built-in impersonator-detection rule ships on by default; add your own to watch for anything else.
- **Signal flags** tell Sift what to watch for in content. A match applies a signal tag during synthesis for filtering and reporting - flags never affect scoring or auto-close.


## Related [#related]

<Cards>
  <Card href="/docs/sift-ai/ai-responses" title="AI Responses">Draft replies for the conversations filtering keeps open.</Card>
  <Card href="/docs/sift-ai/chat" title="Chat">Ask the AI about your filtered queue.</Card>
</Cards>