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

# QA & learning

> Grading calls, capturing corrections, and feeding them back into the agent.

Reading call logs tells you what went wrong. QA and learning is the loop that makes sure it stops going wrong.

## Grading calls

The QA view lets you review calls against consistent criteria rather than by gut feel. Grade a sample regularly — twenty calls a week beats a hundred once a quarter, because you catch problems while they're still small.

What to look for:

<Columns cols={2}>
  <Card title="Accuracy" icon="circle-check">
    Did the agent state prices, hours, and policies correctly? Any factual error is a knowledge-base defect, and it's being repeated on every similar call.
  </Card>

  <Card title="Completeness" icon="clipboard-check">
    Did it collect everything needed to dispatch the job? A booking missing an access detail costs a phone call later.
  </Card>

  <Card title="Outcome" icon="flag-checkered">
    Did the call end somewhere — booked, transferred, or a message taken? Calls that just end are the ones to study.
  </Card>

  <Card title="Judgment" icon="scale-balanced">
    Did it escalate what should have been escalated, and handle what it should have handled?
  </Card>
</Columns>

## Learnings

When review surfaces a correction — a wrong price, a missing answer, a question that should have been asked — capture it as a learning rather than fixing it in your head and moving on.

Learnings are tracked with their source and status, so you can see what came from a real call, what's been applied, and what's still outstanding. That record is what stops the same issue resurfacing three months later.

<Steps>
  <Step title="Capture the correction during review">
    Note what the agent said and what it should have said. Specifics, not "be better on pricing".
  </Step>

  <Step title="Apply it to the right input">
    Wrong fact → [knowledge base](/receptionist/knowledge-base). Missing question → [caller types](/receptionist/caller-types). Wrong escalation → [transfer rules](/receptionist/transfer-rules).
  </Step>

  <Step title="Verify on a live call">
    Call in and reproduce the scenario. Don't mark it done because you edited a field.
  </Step>
</Steps>

<Warning>
  Corrections applied nowhere but a spreadsheet don't change agent behavior. A learning is only closed when the underlying configuration changed and you tested it.
</Warning>

## A workable cadence

| When            | What                                                     |
| --------------- | -------------------------------------------------------- |
| First week live | Read every call. Fix as you go.                          |
| Weeks 2–4       | Read all failed and transferred calls, sample the rest.  |
| Ongoing         | Grade a weekly sample, plus every call that ended badly. |
| On any change   | Re-test the scenarios your change touched.               |

<Tip>
  Have the person who used to answer your phone do the reviewing. They spot a wrong answer instantly, where someone reading fresh will accept it as plausible.
</Tip>