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

# 19. Observers

> Background compliance monitoring and the incident queue.

# Observers and incidents

> [!NOTE]
>
> **Module goal:** set up background monitoring of the rules and know where to read the violations.

An observer is a **detector of a single event** plus a **reaction**. It watches the conversation in parallel and does not slow the agent down. The detector is described in words in the **Event description (business rules)** field, and it has its own severity, evaluation context and observer scope — **agent only**, **user only** or **full dialogue**. Observers are configured in the agent settings: the **Quality & integrations** card → the **Observers** section ([details](https://docs.hubtalk.ai/v4/platform/agent-settings/observers.md)).

An example description for an English-speaking agent on the US market (`us_tcpa`): "The agent promises a discount that is not in the price list". For a Russian-speaking agent the description is written in Russian, for example «Агент обещает скидку, которой нет в прайсе». The description is free text: the auditor role and the detector's response format are built in and need not be described.

The reaction is configured separately and per agent. **Soft actions** do not interrupt the conversation: **Agent correction** (a hint during the conversation, such as "don't promise discounts"), **Incident record & log**, **Flag for review**, **Notify** and tags. **Hard actions (opt-in)** are switched on deliberately: **Transfer to human**, **End call** and **Pause agent**. Plus the thresholds: **Confidence threshold** — the detector confidence at which to react, and **N detections in a row** — how many consecutive detections it takes to react. A hard action cannot be saved without a limiter: it needs **Critical** severity, a confidence threshold, at least two detections in a row, or Human-in-the-loop.

![Observers window with the agent's detector](https://mintcdn.com/hubtalk/UuNkjJ26eB-xpPwv/images/tutorial/tut-37-settings-observers_en.webp?fit=max\&auto=format\&n=UuNkjJ26eB-xpPwv\&q=85\&s=f9972dbe00d1086aeac2560a9fd18068)

The agent's observers. A local detector, Asks not to be called, with Critical severity is attached to the agent; the THEN line shows the default soft actions. Below is the minimum interval between checks: a direct lever on observation spend.

> [accordion] Show the full panel
>
> ![Observers window with the agent's detector](https://mintcdn.com/hubtalk/UuNkjJ26eB-xpPwv/images/tutorial/tut-37-settings-observers_en.webp?fit=max\&auto=format\&n=UuNkjJ26eB-xpPwv\&q=85\&s=f9972dbe00d1086aeac2560a9fd18068)

![Observer incident queue](https://mintcdn.com/hubtalk/UuNkjJ26eB-xpPwv/images/tutorial/tut-38-incidents-queue_en.webp?fit=max\&auto=format\&n=UuNkjJ26eB-xpPwv\&q=85\&s=f4aeca3050635ac4b61c76a910eec2a3)

The Incidents page is the review queue: filters by agent, severity, review status and traffic; each row shows the observer, severity, hint, correction status and review status (open / reviewed / resolved / false positive). Open incidents are shown by default.

Every detection goes to the [**Incidents**](https://docs.hubtalk.ai/v4/platform/incidents.md) page. A row shows the name of the observer that fired, the severity, the hint and whether the correction reached the agent; an expanded row shows the actions taken, the detector's confidence and the trigger quote. Incidents are reviewed by an administrator and a developer. They also see incidents in the call card: on the **Transcript** tab with **Show system events** turned on, and in the **Background observer** panel on the **Overview** tab.

> [!NOTE]
>
> **Advice: where to start**
>
> Three detectors cover most real outbound-calling risks: **promising something that is not in the price list**; **the agent did not introduce itself** (in the USA, telemarketing rules require identifying yourself at the start of the call — check the wording with your lawyer); **the customer asked not to be called again**. Start with soft actions without a correction: a new observer has **Agent correction** and **Incident record & log** on by default — keep the record and add **Flag for review**. A week of observation will show whether to turn on the correction and the hard actions.
>
> The observer does not add the number to the Do not call registry by itself. Under the `us_tcpa` market profile, such requests are written to the registry by automatic capture from the conversation; an "asks not to be called" observer then helps check how the agent responded to the request. Under the `kz` profile the registry is filled manually: the observer flags the call, and the person in charge moves such numbers to the registry — [module 25, Do not call](https://docs.hubtalk.ai/v4/tutorial/dnc.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.