# Training your AI agents

Your agents will get things wrong. They will quote an old price, promise a
response time you no longer offer, or say something in a tone that is not how
your business talks. **Training** is how you fix that — not by editing a
config screen somewhere, but by correcting the agent **on the exact words it
said**, in your own language.

You do not need to know how the knowledge base is organised. You do not need
to write anything technical. You highlight what is wrong, say what is
actually true, and the AI works out where the fix belongs.

## Start here, by what you're trying to do

| You are… | Start with | Because |
|---|---|---|
| **Anyone on the team, wanting to fix something an agent said** | **[Your first correction, step by step](/help/knowledge/training-first-correction)** | One complete worked example, every click shown. You need no special permission to do this. |
| **A workspace admin, switching training on for the first time** | **[Turning training on and managing access](/help/knowledge/training-admin-setup)** | Training is off in a new workspace until an admin enables it — and there is a second, separate step to let your team see each other's corrections. |
| **A manager or admin, handling what colleagues submit** | **[Reviewing what your team teaches](/help/knowledge/training-reviewing)** | The reviewer's side: publishing, declining, undoing, and setting how much agents may publish on their own. |
| **Wondering what happened to something you taught** | **[What happened to your feedback](/help/knowledge/training-outcomes)** | Every receipt explained, and where to find your history. |

Everything else on this page is reference you can come back to.

## The idea in one picture

```mermaid
flowchart TD
    A["An agent says something wrong<br/>(call, chat, SMS, email, task run)"] --> B["You highlight the exact words<br/>and say what's actually true"]
    B --> C["The AI reads your correction<br/>and decides where it belongs"]
    C --> D["Durable business fact<br/>→ Tenant Brain knowledge"]
    C --> E["Tone, wording, behaviour<br/>→ a prompt change for an admin"]
    C --> F["One-off detail<br/>→ noted, not saved as knowledge"]
    D --> G["Every agent reads the corrected<br/>knowledge from now on"]
    E --> H["An admin actions it on the<br/>agent's instructions"]
    F --> I["Nothing changes — and the<br/>thread tells you so"]
    G --> J["You get a receipt saying<br/>exactly what happened"]
    H --> J
    I --> J
```

## The one thing worth understanding

Not every correction becomes knowledge, and **that is correct behaviour, not a
failure**. The AI sorts your feedback into one of three places:

| If your correction is about… | It goes to | Example |
|---|---|---|
| **A fact about your business** that will be true tomorrow too | **Knowledge** — the Tenant Brain, which every agent reads | "Our standard response time is one business day, not 24 hours." |
| **How the agent speaks or behaves** — tone, pacing, length, manner | **A prompt change** — filed for an admin to action on the agent's instructions | "Stop repeating the greeting." · "This reads as too pushy." |
| **Something true only in that one conversation** | **Nowhere** — it is noted and closed | "The caller's name was spelled wrong in this call." |

The thread always tells you which of the three happened. If you only remember
one thing from this page, remember that **a receipt that says "one-off" is a
successful outcome**, not a lost correction.

Full detail: **[What happened to your feedback](/help/knowledge/training-outcomes)**.

## Who can do what

Anyone signed in to your workspace can train — training is deliberately not
locked behind an admin permission, because the people who notice mistakes are
the people reading the conversations.

| You are a… | You can | Your corrections | You can see |
|---|---|---|---|
| **Member** (the default role) | Rate messages, leave corrections, reply in threads, and **resolve or reopen your own threads** | Always land as a **draft for a reviewer** — you contribute, you never publish | Your own training history. The team's, too, if your role carries `quality:events:read` |
| **Manager** | Everything a member can, plus resolving **other people's** threads | May **publish immediately** if your workspace's autonomy settings allow it — the thread tells you which happened | The full quality stream and the Curator Inbox |
| **Admin** | Everything, plus governance and the workspace switch itself | Same as manager | Everything, plus version history and undo |

This is why two people can leave a similar correction and see different
receipts. It is not a bug — a member's correction is deliberately held for a
second pair of eyes.

**Two different things control what you see**, and it is worth keeping them
apart when something is missing:

| Missing thing | Cause | Fix |
|---|---|---|
| No 👍/👎 or comment chips **anywhere**, for **anyone** | The workspace switch is off | An admin turns it on — see **[Turning training on and managing access](/help/knowledge/training-admin-setup)** |
| You can train, but **Quality Events** says *"You don't have access to this page"* | Your role is missing `quality:events:read` | An admin ticks **View Quality Events** on your role |
| You can see the team's threads but have **no Mark resolved buttons** | Your role has read but not `quality:events:manage` | Expected. Close your **own** threads from Train the AI instead |

> **Your own page always works.** **Agents → Workflows → Train the AI** carries
> no permission requirement at all — anyone who can train can always see what
> they taught and what became of it. If that page is empty, you have not
> submitted anything yet; if it is missing, training is off workspace-wide.

> **Custom roles.** If your workspace has custom roles, what you see depends
> on the permission keys that role carries, not on its name. If a page a guide
> mentions is missing from your menus, that is almost always a permission — ask
> a workspace admin rather than assuming the feature is off.

## Does a correction go live immediately?

Sometimes — and the honest answer is **it depends on your workspace, and the
thread will always tell you**.

Your workspace has autonomy settings that decide how much your agents can
publish on their own (**Agents → Workflows → Tenant Brain → Governance**).
Depending on those settings and your role, a correction either:

- **publishes straight away**, and the thread says so — usually with wording
  naming the autonomy level, and a note that it can still be reverted during a
  review window; or
- **lands as a draft**, and the thread says it is waiting for review.

**Never assume which one happened — read the receipt.** Any guide, article or
agent that tells you corrections *always* wait for a human is wrong about your
workspace unless your settings say so.

Admins: see **[Reviewing what your team teaches](/help/knowledge/training-reviewing)**
for how to set and audit this.

## The five-minute version

1. Open any completed conversation — a call, chat, SMS, sent email, or task run.
2. Hover a message **the agent wrote**. A thumbs-up and thumbs-down appear.
3. For a quick signal, click **👍** or **👎**. That is logged instantly.
4. For a real correction, **drag-select the exact words that are wrong**.
5. Click the amber **Comment on "…"** chip that appears.
6. Type what is actually true, in plain language. Send.
7. Wait — the AI takes up to about a minute on knowledge corrections.
8. Read the receipt under the quoted text. It names exactly what happened.
9. Find everything you have taught under **Agents → Workflows → Train the AI**.

## The rest of this guide

| Article | Read it when |
|---|---|
| **[Your first correction, step by step](/help/knowledge/training-first-correction)** | You want one complete worked example with every click shown |
| **[Turning training on and managing access](/help/knowledge/training-admin-setup)** | You are an admin switching training on, or a colleague cannot open Quality Events |
| **[Where you can train](/help/knowledge/training-everywhere)** | You want to train on email, chat, SMS or task-agent output — and to know what deliberately cannot be trained |
| **[What happened to your feedback](/help/knowledge/training-outcomes)** | You left a correction and want to understand the receipt, or find something you taught weeks ago |
| **[Reviewing what your team teaches](/help/knowledge/training-reviewing)** | You are a manager or admin handling what your colleagues submit |
| **[Quality Events and the Curator queue](/help/knowledge/training-curator)** | You are working the whole quality stream, muting a recurring topic, or approving and applying a prompt change |

## Related

- [Agents that learn from your team](/help/knowledge/training-overview) — the flywheel, the governance model, and how this compares to a thumbs-up dataset.
- [Brain Browser walkthrough](/help/knowledge/overview) — the knowledge base your corrections feed.