# Where you can train

Once you have made [your first correction](/help/knowledge/training-first-correction),
the useful thing to know is that **the gesture is identical everywhere**.
Select the words → click the chip → say what is true → read the receipt. You
never have to learn a second way of doing it.

What changes is only *where you find the agent's words*.

## The rule that decides everything

> **You can train what the agent wrote. You cannot train what the agent was
> given.**

Everything on this page follows from that one sentence. If an agent authored
the text, it is trainable. If the text is an input — a prompt, a data payload,
a customer's own words — it is not, because correcting it would teach the
agent something about material it never chose.

## Voice call transcripts

**Agents → Agent Types → Voice Agents →** an agent's row **→** a call (the row
expands to its recent calls; **Open all** shows every call).

The most common place to train. Every message the agent spoke carries the
thumbs and supports highlighting. The caller's messages do not.

## Chat sessions

**Agents → Agent Types → Chat Agents →** an agent **→** a session.

Identical to transcripts. Highlight anything the agent wrote:

![A chat session with a phrase in an AI message highlighted and an amber "Comment on …" chip beneath it.](/help/assets/knowledge/train-chat-chip.png)

## SMS conversations

**Agents → Agent Types → SMS Agents →** an agent **→** a conversation.

Same again. SMS activity also appears inside CRM contact records — see below.

## Sent emails — subject and body are separate

**Agents → Reports & Logs → Sent Emails →** open an email an agent sent.

Email is the one surface with a distinctive extra: **the subject line is
trainable on its own**, separately from the body.

That matters more than it sounds. Subject lines are where an email strategy
most often goes wrong — too pushy, misleading, or promising something the body
does not deliver — and being able to correct the subject without touching the
body means the agent learns the right lesson.

- Highlight words in the **Subject** → you get the chip.
- Highlight words in the **Body** → you get the chip there too.
- Both create separate threads, each anchored to what you actually flagged.

## Task-agent runs

**Agents → Agent Types → Task Agents →** an agent **→ Logs →** the **Train**
button on any row.

Task agents do not have transcripts — they have outcomes. The trainable part
is the **AI Response** block, which shows the answer the agent produced in
readable form, with the machine-readable version tucked behind a toggle:

![A task run's AI Response block showing readable answer text, with a "View raw output" control and thumbs-up/thumbs-down at the bottom right.](/help/assets/knowledge/train-ai-response-readable.png)

Click **View raw output** if you need to see the underlying structured
response. It is shown for reference and is deliberately **not** trainable —
it is machinery, not authorship.

### What is not trainable on a task run, and why

Scroll to the **Final Prompt** or **Input Data** blocks and try to highlight
something. No chip appears:

![A task run's Final Prompt and Input Data blocks, with text selected and no comment chip offered.](/help/assets/knowledge/train-inputs-not-trainable.png)

That is deliberate. Those blocks are what the agent *was given* — the
instructions it ran under and the data it was handed. Correcting them here
would be like correcting a question instead of the answer. If the instructions
themselves are wrong, that is a prompt change: leave your correction on the
**AI Response**, describe what the agent should have done differently, and the
AI will file it for an admin.

## CRM records

**CRM → Records → Contacts →** a contact with agent activity.

Conversations an agent had with a contact are reachable from the contact
record itself, so you can correct something without leaving the record you
were working in.

> **Note:** SMS activity opens in a side drawer on the record. Email and chat
> activity open their own full pages instead. The training gesture is the same
> in both cases — only the container differs.

## Rating a whole interaction

Some surfaces also offer a way to give feedback on an interaction **as a
whole**, rather than on one message — useful when the problem is the shape of
the whole conversation ("it never actually answered the question") rather than
any single sentence.

Use it when no individual message is wrong but the outcome was. Where a
specific sentence *is* wrong, always prefer highlighting it — a quoted span
gives the AI far more to work with.

## Quick reference

| Surface | Where | Trainable |
|---|---|---|
| Voice call | Agents → Agent Types → Voice Agents → a call | Agent messages |
| Chat | Agents → Agent Types → Chat Agents → a session | Agent messages |
| SMS | Agents → Agent Types → SMS Agents → a conversation | Agent messages |
| Sent email | Agents → Reports & Logs → Sent Emails | **Subject and body, separately** |
| Task run | Agents → Agent Types → Task Agents → Logs → Train | **AI Response only** |
| CRM activity | CRM → Records → a contact | Agent messages in the activity |
| — | — | **Never:** the other person's messages, Final Prompt, Input Data, raw output |

## Next steps

- **[What happened to your feedback](/help/knowledge/training-outcomes)** —
  read your receipts and find everything you have taught.

## Related

- [Training your AI agents](/help/knowledge/training) — the overview and role model.
- [Your first correction, step by step](/help/knowledge/training-first-correction) — the full worked example.