A lead scoring model is supposed to tell your team who to contact first. In practice, many reps ignore the score and work the list their own way. Scoring itself is a sound idea. Reps give up on it because the score is a black box, it keeps changing for no visible reason, and the “hot” leads it flags often turn out to be students and competitors who clicked a lot of emails.
This guide covers how to build a lead scoring model that sales will actually use. It walks through the three components (fit, warmth and timing), a table of buying signals with suggested weights, a worked example you can copy, and an honest look at where AI lead scoring helps and where it falls short.
What a lead scoring model is
Lead scoring gives each lead a number that answers two questions: should we sell to this person at all? and should we contact them now? A lead scoring model is the set of rules that produces that number.
Most mature systems split the work into separate parts. HubSpot’s lead scoring tool, for example, distinguishes fit scores (based on properties like job title, company size and revenue) from engagement scores (based on actions like email opens and website visits), and lets you combine them while still viewing each part separately (HubSpot Knowledge Base, as of November 2026). Of all the design choices you’ll make, keeping those parts separate matters most, so we’ll start there.
For outbound teams, we recommend thinking in three parts:
| Component | Question it answers | Example inputs | Changes how often |
|---|---|---|---|
| Fit | Are they the kind of buyer we win with? | Industry, company size, title, seniority, location | Rarely (job or company changes) |
| Warmth | Do they know us, and are they engaging? | Accepted connection, replies, profile views, post engagement | Daily, and fades over time |
| Timing | Is something happening that makes now a good moment? | Funding, hiring, new leader, office move, relevant post | When events happen, and expires |
Fit should carry the most weight, because no amount of engagement turns a bad-fit lead into a customer. A junior marketer at a 3-person company who opens every email is warm, but they aren’t going to buy.
Why sales teams stop trusting lead scores
Almost every model that reps distrust has at least one of these six problems:
- Nobody can explain the number. A rep sees “82” and has no idea why, and when the 82 turns out to be a poor fit, they stop looking at scores.
- Fit and engagement are mashed together, so a bad-fit lead with lots of clicks outranks a perfect-fit lead who hasn’t engaged yet.
- Points never expire. A webinar attended 14 months ago counts as much as a reply yesterday.
- There are too many inputs. With forty rules, each worth 3 to 7 points, nobody (including the person who built it) can predict the result.
- It was never checked against outcomes. The weights were set once in a workshop and never compared with who actually replied, met or bought.
- The score drifts. If the model is a black-box AI that scores a lead differently today than it did yesterday with no new data, reps notice.
The steps below are built to avoid each of them.
Start with your ICP before you assign points
Fit scoring is your ideal customer profile turned into numbers. If you don’t have a written ICP with specific firmographics, triggers and disqualifiers, build your ICP first. Otherwise you’ll be inventing weights for criteria nobody has agreed on.
From the ICP, pull out:
- The must-haves, which become hard filters. A lead outside them is excluded outright rather than given a low score.
- The strong-fit criteria, which earn the most fit points.
- The disqualifiers, which either exclude the lead or apply a large negative.
- The trigger events, which feed the timing component.
Buying signals: what to score and how much
Buying signals are events or behaviors that suggest an account may be ready to buy. Some are about the person, some about the company. The table below is a starting point for B2B outbound on LinkedIn and email. Treat the weights as suggestions and adjust them once you have your own outcome data.
| Buying signal | Type | What it suggests | Suggested weight | Expires after |
|---|---|---|---|---|
| Replied positively to outreach | Warmth | Active interest | Very high | Stays until the conversation resolves |
| Accepted connection request | Warmth | Open to contact | Medium | ~30 days |
| Viewed your profile after you contacted them | Warmth | Curiosity | Low-medium | ~14 days |
| Liked or commented on your post | Warmth | Familiarity | Low-medium | ~21 days |
| Posts regularly on LinkedIn | Warmth / reachability | Likely to see and answer messages | Medium | Re-check monthly |
| New leader in the role you sell to (first 90 days) | Timing | New budget, new priorities | High | ~90 days |
| Funding round announced | Timing | Budget and growth pressure | High | ~120 days |
| Hiring for the function you help | Timing | The problem is live right now | High | While the job ad is open |
| Public post about the problem you solve | Timing | Pain in their own words | High | ~30 days |
| New office or expansion | Timing | Change creates new needs | Medium | ~90 days |
| Uses a competitor or adjacent tool | Fit | Category awareness | Medium (plus or minus) | Re-check quarterly |
| Already has an in-house solution | Fit | Less need | Negative | Re-check quarterly |
The table only works if warmth decays and timing expires. A reply last week should count far more than a like three months ago, so use a simple decay (for example, halve the value every two to four weeks) instead of letting points pile up forever. And a funding round from last year is just background by now.
How to build a lead scoring model in 7 steps
Step 1: Pick the outcome you are predicting
“Good lead” is not an outcome. “Replies positively within 30 days” or “books a meeting within 60 days” is. For outbound, a positive reply is usually the best early outcome because you collect enough data quickly. Closed-won is better in theory, but it takes months to build up.
Step 2: Choose three to six fit criteria
Use the ICP. Common criteria are industry match, company size band, title match, seniority and location. Give each one a 0-100 sub-score with clear steps. For example, here is company size against an ICP of 10 to 200 employees:
| Company size | Size sub-score |
|---|---|
| 10-200 (core ICP) | 100 |
| 201-500 or 5-9 | 50 |
| Outside those ranges | 0 |
Step 3: Define warmth from real interactions
List the interactions you can actually see: accepted connections, LinkedIn replies, profile views, post engagement, email replies. Assign points and a decay rate to each, and keep it to five or six inputs.
Step 4: Add two or three timing signals
Pick the triggers that came before most of your best deals (your ICP should already list them), and give each an expiry date.
Step 5: Weight the components
A reasonable starting split for outbound leans on fit, for example:
- Fit: 40-50%
- Warmth or activity: 25-30%
- Timing or authority: 25-30%
As one reference point, Wino AI’s priority score uses default weights of 40 / 30 / 30 across ICP fit, LinkedIn activity and seniority, which each workspace can tune. Your mix will depend on your market. Whatever split you pick, choose it on purpose and write it down.
Step 6: Set score bands and actions
A score is only useful if it maps to an action. Use three or four bands:
| Band | Score | Action |
|---|---|---|
| A | 80-100 | Contact this week with the most personalized outreach; a human writes or reviews every message |
| B | 60-79 | Standard sequence with personalization |
| C | 40-59 | Nurture: lighter touches, revisit when a timing signal appears |
| D | 0-39 | Do not contact now |
Step 7: Check the bands against outcomes
After a few hundred contacts, compare outcomes by band. You want a clear staircase where A beats B and B beats C. If you don’t see one, your weights are wrong.
| Band | Leads contacted | Positive replies | Positive reply rate |
|---|---|---|---|
| A | |||
| B | |||
| C |
Fill this in from your own data. If C converts as well as A, one of your fit criteria isn’t predictive, and the fix is usually to remove a criterion rather than add one.
A worked lead scoring example
Here are two prospects for a fictional recruiting software company whose ICP is “founders of London recruiting agencies with 10 to 200 staff who are growing their team”. Weights: fit 40%, LinkedIn activity 30%, seniority 30%.
Maya Chen, Founder, Northwind Talent (40 staff, London), posts weekly about hiring and recently wrote about scaling the team.
Jordan Ellis, Talent Coordinator at a 2,000-person bank (London), opens most emails and rarely posts.
| Component | Weight | Maya: sub-score | Maya: weighted | Jordan: sub-score | Jordan: weighted |
|---|---|---|---|---|---|
| ICP fit | 40% | 94 | 37.6 | 20 | 8.0 |
| LinkedIn activity | 30% | 88 | 26.4 | 35 | 10.5 |
| Seniority | 30% | 90 | 27.0 | 30 | 9.0 |
| Priority score | 91 | 28 |
Jordan’s email opens don’t rescue a poor fit, which is the behavior you want. A rep can also see why Maya is a 91. If she turns out to be a poor lead, you can point to the input that was wrong, and that traceability is how a model earns trust over time.
The score also leaves something out: Northwind may already have an in-house sourcer. That’s a signal against the match, and a human or an AI fit check should show it next to the score where the rep can see it.
Account scoring: when one contact is not enough
In B2B you sell to companies, so a lead-level score is only half the picture. An account score answers “how far along are we with this company?” A simple version combines:
- The warmth of your warmest contact at the account
- How many relevant people you have reached there, out of how many exist
Account scoring also helps you spot your champion, the person at the account who is engaging the most. If Maya is warm and two of her colleagues have accepted but not replied, Maya is the person to lean on, and the account deserves more attention than any single contact score suggests.
Rules vs AI lead scoring
AI lead scoring is now a standard feature in lead scoring software. It handles some parts of the job well and others badly:
| Rule-based scoring | AI lead scoring | Hybrid | |
|---|---|---|---|
| Consistency | Same inputs, same score | Can vary between runs | Score from rules, consistent |
| Explainability | Every point traceable | Often a single number with weak reasoning | Rules give the number, AI explains it in words |
| Handling messy data | Poor: needs clean fields | Good: can read bios, posts, company pages | Good |
| Needs historical data | No | Predictive models usually need a lot of past outcomes | No for the score, yes for tuning |
| Setup effort | Moderate | Low to configure, hard to audit | Moderate |
| Best for | Fit and priority | Reading context, spotting nuance | Most outbound teams |
We recommend the hybrid, where rules do the math and AI does the reading. Deterministic rules calculate the number so it never drifts with a model’s mood, and AI reads the profile and company to list the signals for and against the match in plain language. Reps get a stable number and a reason they can check.
Wino AI’s Find handles it this way: priority, warmth and account scores follow fixed rules you can see, and an AI fit check sits alongside them with the reasons behind the match. Whatever you use, ask every vendor the same question: “If this score is wrong, can I see which input was wrong?”
Choosing lead scoring software: a checklist
Whether you use your CRM’s built-in scoring or a separate tool, check for:
- Separate fit and engagement (or warmth) scores you can view on their own
- A visible breakdown of every score, not just the total
- Decay for engagement points, and expiry for timing signals
- Adjustable weights per team or workspace
- Negative scoring and hard disqualifiers
- Account-level scoring, not only contact-level
- Signals from the channels you actually use (for outbound teams, LinkedIn and email, not only website visits)
- A way to compare score bands against real outcomes
- If AI is involved: written reasons for each score, and a stable number between runs
Keep tuning: let outcomes correct the weights
Treat your lead scoring model as a hypothesis and use outcomes to test it. Every month while it’s new, then quarterly:
- Pull outcomes by band (the table in Step 7).
- Look at the A-band leads who never replied. What do they have in common?
- Look at the C- and D-band leads who did reply. What did the model miss?
- Change one weight or criterion at a time, and give it enough contacts to tell whether it helped.
Apply the same discipline to the outreach itself. A high score tells you who to contact, but it says nothing about what to say or how persistent to be. Leads in your B band often need two or three well-spaced touches before they reply, and a sensible cold email follow-up cadence covers that part. When you test changes to messaging, compare each variant against your current baseline instead of switching on a hunch. Wino AI’s Learn works this way, tracking each variant against the baseline with visible sample sizes so you promote what works.
If you’re weighing how much of this to hand to software, our guide to what an AI SDR is covers which parts of the SDR job, scoring included, AI handles well and which a human should still decide.