# Your first correction, step by step

This walks a single correction all the way through. Follow it once on a real
conversation and the gesture will stick — most people need two or three goes
before it feels natural, so do not worry if the first one feels slow.

**Time:** about five minutes, including roughly a minute of waiting while the
AI works.

## What you'll need

- Any account signed in to your workspace. Training is open to everyone —
  there is no permission to request first.
- One completed conversation where an agent said something you can correct. A
  voice call with a transcript is the easiest place to start.

---

## Step 1 — Open a conversation the agent handled

Go to **Agents → Agent Types → Voice Agents** (the Agents list, filtered to
voice), open any agent with **View**, and open a completed **call**.

> **Where is that menu?** Navigation is the **bar across the top** of the
> window, not a sidebar. Each menu opens a wide panel split into sections with
> grey headings — so the full path reads *menu → section → item*.

You will see the conversation as a transcript: the agent's messages and the
other person's, in order.

## Step 2 — Find a message the agent wrote

**You can only train what the agent said.** The other person's messages have
no training controls — they are a record of what happened, not something to
correct.

Look at any agent message without touching it. The controls are already there,
just quiet:

![An AI message bubble with faint grey thumbs-up and thumbs-down icons below its right corner.](/help/assets/knowledge/train-thumbs-at-rest.png)

Move your pointer over the message and they darken:

![The same AI message bubble, with the thumbs-up and thumbs-down icons now in dark grey.](/help/assets/knowledge/train-thumbs-on-hover.png)

## Step 3 — (Optional) Leave a quick rating

If the message was simply good or simply bad and you have nothing to add,
click **👍** or **👎**. The buttons are replaced by a confirmation:

![The same AI message with the thumb buttons replaced by a green check mark and the words "Training recorded".](/help/assets/knowledge/train-training-recorded.png)

That is the whole interaction — no form, no follow-up. Ratings are counted as
signal; they do not create a correction on their own.

**To actually fix something, keep going.**

## Step 4 — Select the exact words that are wrong

Click and **drag across the whole phrase** you want to correct. An amber chip
appears showing what you selected:

![An AI message with a phrase highlighted in blue, and a yellow chip below reading "Comment on …".](/help/assets/knowledge/train-comment-chip.png)

> **Read the chip before you click it.** It shows the exact words that will be
> quoted in your correction. Double-clicking selects only a single word, and a
> one-word quote gives the AI very little to work with — it will usually come
> back with a clarifying question, costing you an extra round trip. Drag across
> the full phrase instead.

## Step 5 — Open the thread

Click the chip. A thread opens, pinned under that message. The AI quotes your
selection back and asks you a question — you are never given a blank mandatory
box:

![A feedback thread card quoting a phrase with the question "You highlighted this part — what's wrong or missing here?" above a reply box and Send button.](/help/assets/knowledge/train-thread-opens.png)

## Step 6 — Say what is actually true

Type your correction in plain language. Write it the way you would explain it
to a new colleague:

> The agent said support requests get a response within 24 hours. Our
> commitment is one business day for standard requests — this needs correcting.

Two things that make a correction land well:

- **Name the wrong value and the right one.** "This is wrong" gives the AI
  nothing to act on; "it says 24 hours, it should be one business day" does.
- **Say if it should not become a rule.** If it is specific to this one
  conversation, say so — the AI will note it and close, rather than changing
  what every agent knows.

Click **Send**.

## Step 7 — Wait, and watch it work

The AI now reads your knowledge base, finds the relevant material, and works
out what to change. **This can take up to about a minute** on a knowledge
correction — it is doing real work, not just replying.

You will see a line telling you it is working. On the reviewer side, the same
process shows its steps:

![Three completed steps listed in order — Searching the brain, Reading knowledge, Drafting a correction — each with a green check.](/help/assets/knowledge/train-agent-working-steps.png)

If it comes back with a **clarifying question** instead of an outcome, answer
it in the same thread — that is normal, especially if your highlighted phrase
was short or ambiguous.

## Step 8 — Read your receipt

This is the most important step, and the one people skip.

Underneath the quoted text you will find a **receipt** — a short status line
saying exactly what happened to your correction. There are several possible
receipts and **they all count as success**. Here is a draft awaiting review:

![An amber confirmation panel reading "Draft created from this — waiting on an admin to review it", with an "Open the draft" link.](/help/assets/knowledge/train-receipt-draft.png)

And here is one that became knowledge outright:

![A green notice reading "This became knowledge — policies/service-levels.md. Thank you for teaching it."](/help/assets/knowledge/review-after.png)

Every possible receipt, and what each one means, is in
**[What happened to your feedback](/help/knowledge/training-outcomes)**.

> **A thread with no receipt at all is the only real failure.** If you see one,
> that is worth reporting.

## Step 9 — Find it again later

You never have to remember which conversation you corrected. Open the **Agents**
menu in the top bar — **Train the AI** sits under the **Workflows** heading:

![The Karmaflow.ai top navigation with the Agents menu open, showing "Train the AI" under the Workflows column.](/help/assets/knowledge/train-nav-menu.png)

That page lists everything you have taught, with counters across the top and
an outcome badge on every entry:

![The Train the AI page showing counters for feedback threads, became knowledge, pending review and prompt changes, with feedback entries below.](/help/assets/knowledge/train-page.png)

Your entry shows the phrase you quoted, your comment, and what became of it:

![A Train the AI entry showing a quoted phrase, the member's correction, and an outcome badge.](/help/assets/knowledge/train-entry-detail.png)

## Step 10 — When it becomes knowledge

If your correction became knowledge, the entry turns green and offers a link
to the knowledge itself:

![A Train the AI entry with a green "Became knowledge" badge and an "Open the concept" link.](/help/assets/knowledge/train-became-knowledge-entry.png)

Back on the original conversation, the thread shows the same outcome and
closes itself:

![A feedback thread on a call transcript showing a green notice that the correction became knowledge, naming the concept, with the thread marked closed.](/help/assets/knowledge/train-thread-closed.png)

The thread closing is deliberate — the question it was opened to answer has
been answered. **To raise something else about the same message, start a new
comment on the highlight.**

> **If the concept name is plain text rather than a link**, your role cannot
> open the knowledge base directly. The correction still landed — hover the
> text and it will tell you who can give you access.

## Troubleshooting

**I highlighted a phrase but no chip appeared.**
Check you selected text inside a message the *agent* wrote. The other person's
messages are not trainable. If you selected inside an input field or a
non-agent block, the chip will not offer itself. And if there are no thumbs
anywhere either, training is switched off for the workspace — a workspace
admin turns it on under **Tenant Settings → Agents & AI → Agent Feedback &
Training**.

**The chip quotes fewer words than I highlighted.**
Drag-select rather than double-clicking, and check the chip text before
clicking it.

**The AI asked me a question instead of fixing it.**
Answer in the same thread. This usually means the quoted phrase was too short
to identify the fact, or your correction did not state the correct value.

**Nothing happened for a minute and I thought it had frozen.**
Knowledge corrections routinely take 40–60 seconds. A progress line tells you
it is still working.

**My correction became a "prompt change" and I expected knowledge.**
That means the AI judged it to be about *how* the agent speaks rather than
*what* is true. If you disagree, reply in the thread and say the correction is
a business fact — see
[What happened to your feedback](/help/knowledge/training-outcomes).

## Next steps

- **[Where you can train](/help/knowledge/training-everywhere)** — the same
  gesture on email, chat, SMS and task-agent output.
- **[What happened to your feedback](/help/knowledge/training-outcomes)** —
  every receipt explained.

## Related

- [Training your AI agents](/help/knowledge/training) — the overview and role model.
- [Brain Browser walkthrough](/help/knowledge/overview) — the knowledge base your corrections feed.