> ## Documentation Index
> Fetch the complete documentation index at: https://docs.poly.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# AI Scores
> Score every conversation from 1 to 5 against criteria you write yourself.
**AI Scores** let you score conversations against your own criteria. You give the score a name, describe what it measures, and add up to five weighted criteria. After each conversation ends, the AI reads the transcript, grades every criterion, and gives the conversation a **1–5 score**.
AI Scores sit alongside [PolyScore](/analytics/polyscore). PolyScore keeps working exactly as before.
## AI Scores or PolyScore?
| | [PolyScore](/analytics/polyscore) | AI Scores |
| - | - | - |
| **Who defines it** | PolyAI | You |
| **How it is measured** | The AI grades the transcript against a fixed PolyAI rubric | The AI grades the transcript against criteria you write |
| **Value** | 1–5 | 1–5, one decimal place |
| **Best for** | A standard quality score you can compare across agents | Quality, outcome, or compliance checks that are specific to your business |
Use an AI Score when you need a check that PolyScore doesn't cover, such as *"Did the agent give the required disclosures?"*
## How an AI Score works
Each AI Score has:
* A **description** that gives the AI the overall context for the score.
* Up to **five criteria**. Each criterion has a title, a description that tells the AI what to look for, a score type, and a weight.
There are two score types:
| Score type | Answers |
| - | - |
| **3-point scale** | Good, Fair, or Poor |
| **Yes/No** | Yes or No |
Criterion weights are whole percentages, and they must add up to exactly 100%.
### How the score is calculated
The AI grades each criterion. Each answer is worth a value on the 1–5 scale:
| Score type | Answer | Value |
| - | - | - |
| **3-point scale** | Good | 5 |
| | Fair | 3 |
| | Poor | 1 |
| **Yes/No** | Yes | 5 |
| | No | 1 |
The score is the weighted average of these values, rounded to one decimal place. Criteria with a higher weight move the score more.
For example, a score has two criteria: **Request resolved** (Yes/No) and **Accurate information** (3-point).
| Request resolved (70% weight) | Accurate information (30% weight) | Score |
| - | - | - |
| Yes | Good | 5.0 |
| Yes | Fair | 4.4 |
| No | Good | 2.2 |
| No | Poor | 1.0 |
The score badge is color-coded:
| Score | Color |
| - | - |
| 4.0–5.0 | Green |
| 3.0–3.9 | Yellow |
| 1.0–2.9 | Red |
### Which conversations get scored
An active AI Score scores a conversation when all of these are true:
* The conversation is on a channel the score applies to: **Voice**, **Webchat**, or both.
* The conversation has at least **3 turns**. This is a fixed system default, and you can't change it per score.
* The conversation does not match the score's **exclusion criteria**, if it has any.
* The conversation ended after the score was turned on. AI Scores do not score past conversations.
If a conversation isn't scored, it shows **N/A**. Hover over **N/A** to see the reason. See [Scoring states](#scoring-states).
## Create an AI Score
Only **Admins** can create, edit, turn on, turn off, duplicate, and delete AI Scores.
Go to **Analyze > Analytics**, select **Edit**, then select **AI scores**.
Select **+ AI score**, then choose a [template](#templates) or select **Start from scratch**.
* **Score name** — a short, unique name for the score. It becomes the column header in Conversations. Up to 60 characters.
* **Description** — what the score measures. This gives the AI the overall context for the score. Up to 500 characters.
* **Apply to** — the channels to score: **Voice**, **Webchat**, or both.
* **Exclusion criteria** (optional) — describe the conversations the score should skip, for example *"The caller dialed the wrong number."* Matching conversations show **N/A** and are left out of totals. Up to 500 characters.
Select **Continue**.
For each criterion, enter:
* **Title** — what the criterion checks, for example *"Verified account holder"*. Up to 60 characters, unique within the score.
* **Description** — what the AI should look for when it grades this criterion. Up to 500 characters.
* **Score type** — **3-point scale** or **Yes/No**.
* **Weight (%)** — a whole number from 1 to 100.
Select **+ Criterion** to add another criterion, up to five. The running total above the criteria shows how close the weights are to 100%.
Select **Create score**, then confirm.
New scores start **off**. Use the toggle on the score's card to turn it on. The score starts with the next conversations that end.
Changes can take up to **5 minutes** to apply. Conversations that end before then are scored against the previous version of the score, if there was one.
### Write good criteria
* **Check one thing per criterion.** *"Verified the caller before discussing the account"* is easier to grade than *"Was secure and polite"*.
* **Describe each answer.** For a 3-point criterion, say what Good, Fair, and Poor look like. For a Yes/No criterion, say when the answer is Yes. The templates show this pattern.
* **Describe only what the transcript can show.** The AI reads the transcript only. It can't check your systems or your knowledge base.
* **Use exclusion criteria to skip conversations.** If the score doesn't apply to a type of conversation, such as wrong numbers or silent calls, exclude those conversations. Don't add the condition to every criterion.
* **Weight what matters most.** A criterion with 60% weight moves the score more than a criterion with 20% weight.
* **Check the first results.** After you turn a score on, open some scored conversations in [Conversation review](/analytics/conversations/review) and read the breakdown. Refine the descriptions where the grades don't match your judgment.
## Manage AI Scores
All management happens in the **AI scores** panel. Open it from **Analyze > Analytics** > **Edit** > **AI scores**.
| Limit | Value |
| - | - |
| AI Scores per project | 10 |
| Active AI Scores per project | 3 |
| Criteria per score | 5 |
The counters at the bottom of the panel show how many scores you have created and how many are active.
* **Turn on or off** — use the toggle on a score's card. Turning a score off stops it scoring new conversations. Past results are kept, and you can still chart them. Turning a score back on does not score the conversations that ended while it was off.
* **Edit** — select **⋮** > **Edit**. Conversations that were already scored keep the criteria and results they were scored with. Edits apply only to new conversations.
* **Duplicate** — select **⋮** > **Duplicate**. The copy is named *"\ (copy)"* and starts off.
* **Delete** — select **⋮** > **Delete**. The score stops scoring conversations and is removed from the list, from the Conversations columns, and from the filters.
To change what a score measures without mixing old and new results, duplicate it, edit the copy, then turn the original off and the copy on.
Users who are not Admins can open the **AI scores** panel to see each score and its criteria, but they can't change anything.
## Where AI Scores appear
### Conversations table
Each active AI Score has its own column in the [conversations table](/analytics/conversations/review#conversations-table). The columns are shown by default, after the standard columns. Hover over a score to see the breakdown: each criterion with its weight and its answer.
To show or hide AI Score columns, select **Column**. They are listed with your custom metrics.
### Filters
Select **Filter** and pick a score under **AI Scores created by you**. Compare the score with a value from 1 to 5 using *equals*, *less than*, *greater than*, *less than or equal to*, or *greater than or equal to*. Use *exists* to find the conversations that have a score, or the ones that don't.
Scores that are turned off stay in the filter list, so saved views that filter on them keep working.
Save a [view](/analytics/conversations/views) filtered to a low AI Score to build a QA queue for that check.
### Scores tab
Open a conversation and select the **Scores** tab. Each AI Score is listed below PolyScore with its badge. Expand a score to see each criterion with its weight, its answer, and its description. The same breakdown appears in the full-page conversation view.
### CSV export
[Exports](/analytics/conversations/review#export) include one column for each AI Score, named **AI Score: \**.
### Dashboards
AI Scores are available in the **Metric** field when you add a tile to a [self-serve dashboard](/analytics/dashboards/introduction). Look under **AI Scores created by you**. Scores that are turned off are still listed, so you can chart their past results.
### Wren
[Wren](/wren/analyze) knows your project's AI Scores and can query their values. For example, ask *"What's the average Compliance check score this week?"* or *"Which days had the lowest Task success score?"* Wren can see each score's name, description, and whether it is on. It can't see the criteria or the weights.
## Scoring states
While a conversation that has ended is being scored, the cell shows a loading badge. Scoring usually takes a few moments.
When a conversation has no value for an AI Score, the cell shows **N/A**. Hover over it to see why.
| Reason | What it means |
| - | - |
| **Too short to score** | The conversation had fewer than 3 turns. |
| **Matched the exclusion criteria** | The score's exclusion criteria matched the conversation, so it was skipped. |
| **Scoring failed** | The AI couldn't produce a valid score for this conversation. |
| **Not scored** | The conversation ended before the score was turned on, is on a channel the score doesn't cover, or is still being scored. |
Conversations with **N/A** are left out of averages and other totals.
## Limitations
AI Scores grade the transcript only. They can't access your knowledge base, flows, or external systems.
* An AI Score **can't confirm that an action happened** in an external system, such as a booking or a refund. It can only judge what the conversation says.
* **Past conversations are not scored.** Turning a score on or editing it applies to new conversations only.
* The breakdown shows each criterion's answer, but not the AI's reasoning for it.
* You can filter by the overall score, but not by a single criterion. AI Score columns can't be sorted.
## Related pages
The standard 1–5 quality score that PolyAI defines.
Filter by AI Score and read the criteria breakdown.
Chart AI Scores over time.