# AI Lead Scoring



Every lead and every response carries an AI lead score, so routing, alerting and
follow-up can be automatic. It is a number you can filter on: a segment
definition like `Score is 75-100` is the everyday use of it.

## What Goes into It [#what-goes-into-it]

| Half       | What it asks                                             | Inputs                                                                                                                                   |
| ---------- | -------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
| **Fit**    | Does this account match your [ICP](/platform/leads/icp)? | Industry, company size, revenue, geography, job title, seniority, technologies, each with the weight you gave it                         |
| **Intent** | Did they behave like someone with a problem?             | Website behaviour, a weighted ICP dimension of its own: sessions, pages, forms viewed, started and completed, and the answers themselves |

A high-fit account with no behaviour and a low-fit one with intense behaviour
can score the same for opposite reasons, so read the score next to the record
rather than instead of it.

## Reading the Distribution [#reading-the-distribution]

Surface gives you the number and the statistics behind it. Reading the shape of
the distribution is judgement rather than a product feature, so treat the rest of
this section as how teams do it.

`get_ai_score_stats` returns average, min, max and distribution buckets in one
call, and `get_top_ai_scores` with `order: "asc"` returns the bottom, usually
the more informative end when you are debugging.

What people look for in that distribution is spread:

| Shape                              | What it usually means                                                                                |
| ---------------------------------- | ---------------------------------------------------------------------------------------------------- |
| Broad spread, thin high tail       | Healthy. The score is discriminating.                                                                |
| Everything clustered in the middle | The ICP is too vague, or its weights are too even, to separate anyone.                               |
| A large spike at the very bottom   | Often free-provider emails or bot traffic. Check [spam protection](/platform/forms/spam-protection). |
| A large spike at the top           | The ICP may be too narrow, or too much of your traffic is one account.                               |

## Setting a Threshold [#setting-a-threshold]

Wherever you draw the line, it decides who gets routed and who triggers an
alert, so pick it from your own data rather than from a round number. The method
most teams use:

<Steps>
  <Step title="Score your closed-won accounts">
    Most should land above the line you are considering.
  </Step>

  <Step title="Score your closed-lost-as-unqualified accounts">
    Most of those should fall below it.
  </Step>

  <Step title="Take the crossover, then check the volume it implies">
    A threshold that produces four leads a week for eight reps is set too high.
  </Step>
</Steps>

Eligibility to book is that threshold plus a territory check. See
[Routing](/platform/scheduling/routing) for how the two fit together, and talk
to sales before you move the line.

## Checking That It Works [#checking-that-it-works]

The real test is outcome: over a quarter, do high-scoring leads book and close
more often than low-scoring ones?

* The **Meetings** screen lists the score beside every booked meeting, which is
  the quickest look at what scores actually book.
* `get_leads_with_ai_scores` gives you the ranking to compare that against, and
  `get_leads_count` with `byEventType: true` gives the meeting-booked totals.
* `compare_forms` shows whether one form produces systematically worse-scoring
  leads than the rest. That is a targeting problem upstream, not a scoring one.

No correlation usually means an ICP that describes an aspiration rather than
your book of business. Fix [the ICP](/platform/leads/icp) first.

<Tip>
  [SAM](/platform/sam/overview&#x29; can rank on outcome rather than volume from Slack:
  &#x2A;"rank our blog posts by meetings booked, not pageviews"*. See the
  [prompt library](/platform/sam/prompts).
</Tip>