---
title: AI filtering
description: How Hearl's AI judges each post, the thresholds it uses, and what happens when it isn't sure.
---

Hearl's AI answers a set of questions about each post, each with a probability. It reads each post together with your product's name, description, **Not to be confused with** text and official accounts. When the post is a reply, the post it replied to goes along with it.

## What the AI answers [#questions]

| Question | Answer |
| --- | --- |
| Is the post about this product? | A probability from 0 to 1. |
| Would a reply from the official team be valuable? | A probability from 0 to 1. |
| What is the post mainly about? | One of 8 categories. |
| How urgently should the team handle it? | Low, Medium or High. |

### Categories [#categories]

| Category | Meaning |
| --- | --- |
| Bug | Something is broken, crashes, errors or behaves wrongly. |
| Question | The person asks how to do something or needs help. |
| Feature request | The person requests or suggests a feature or a change. |
| Pricing | Pricing, subscription, billing or refunds. |
| Criticism | General dissatisfaction without a specific bug. |
| Praise | Praise, thanks or excitement. |
| Buying | Considering buying, or comparing before purchase. |
| Other | None of the above. |

### Priority [#priority]

- **Low.** Praise, casual chat or an aside. Nothing goes wrong if nobody answers.
- **Medium.** A question, feature request or buying question that deserves an answer within a few days.
- **High.** A bug, a data or billing problem, strong negative sentiment, or a post many people are seeing.

## Thresholds [#thresholds]

Hearl turns each probability into a decision:

| Probability | Decision |
| --- | --- |
| 0.7 or higher | Yes |
| 0.3 or lower | No |
| Between 0.3 and 0.7 | Not sure |

- A post that is clearly not about your product gets the status **Not about the product** and leaves the inbox.
- A post that is about your product but clearly needs no reply, such as pure praise or a joke, gets **No reply needed**.
- Everything else stays in the inbox.

## Needs a human check [#human-check]

Hearl doesn't drop a post because of a judgment it isn't sure about. When a post would be filtered out but one of the answers falls between 0.3 and 0.7, it gets the status **Needs a human check** and appears in the **To review** view. A person decides.

Posts with an uncertain judgment are not sent in [notifications](/docs/notifications).

## See why [#see-why]

Each conversation shows the AI's probabilities, for example "relevant 92%" and "needs reply 81%", so you can see why something was kept or set aside. If a post was set aside by mistake, you can still find it under **All**.

## Replies under your own posts [#official-posts]

A popular post from your team can collect hundreds of replies. Hearl judges each new reply and flags only those that are about your product, need a reply and are Medium or High priority. They appear under **Replies that need attention**.