What is a lead scoring model?

A lead scoring model is the framework you use to rank prospects by how likely they are to become customers. It assigns values to two things: fit (who the lead is — company size, industry, role, ICP match) and intent or engagement (what they've done). The score tells sales who to work first. The shift in 2026 is that the grading is moving from manual point tables to AI that reads the actual company and person.

What are the components of a good lead scoring model?

Four parts: a fit score for the company (does the account match your ICP), a fit score for the contact (is this the right role and seniority), an intent or signal layer (recent activity, hiring, funding, trigger events), and negative scoring (disqualifiers that push the score down — wrong industry, competitor, free-mail student, stale role). Most broken models only have the engagement layer and skip company fit entirely.

What is a two-level lead scoring model?

A two-level model scores the company and the decision-maker separately, then combines them. A perfect-fit account with the wrong contact is not a hot lead; a great contact at an out-of-ICP company isn't either. Scoring both levels stops the classic failure where a model fires an MQL because someone junior downloaded an ebook. Lead Scorer is built around this: it rates the company on ICP fit and the individual 0-10 against your product, and rejects off-target leads with a written reason.

Rules-based vs predictive vs AI lead scoring — what's the difference?

Rules-based models use point tables you set by hand (title = +10, opened email = +2). Predictive models train machine learning on your closed-won history to weight factors automatically — powerful, but they need a large, clean dataset. AI-agent scoring is the newest: an LLM reads the company and person like an analyst would and grades them against your ICP in plain language, so you get a rationale instead of an opaque number, and it works on day one without years of CRM data.

Why do most lead scoring models fail?

Because they score engagement, not fit. Email opens, page views, and form fills are easy to track and easy to game — bots and tire-kickers rack up points while a perfect-fit buyer who didn't click anything yet scores zero. The result is the pattern teams complain about in 2026: thousands of MQLs, a handful of opportunities. A model that leads with company-and-contact fit, not clicks, fixes the root cause.

What should the threshold be for a sales-ready lead?

Set the threshold by working backwards from conversion, not by picking a round number. Look at what scores your last 20-30 closed-won deals actually had at handoff and set the bar just below that band. Then enforce negative scoring so disqualifiers cap the score. A threshold is only as honest as the model under it — a high bar on an engagement-only model just passes well-behaved tire-kickers.

How often should I update my lead scoring model?

Review it every quarter against real outcomes: pull the leads that scored high but never converted and the low scorers that closed, and adjust. People and companies also drift — a contact who was the right role in January gets promoted out of the buying seat by April — so any model that scores fit needs fresh data, not a one-time enrichment that decays.

Can I build an AI lead scoring model without a data team?

Yes. The predictive route needs a data scientist and a big closed-won dataset, but AI-agent scoring doesn't. You describe your product and ICP in plain language and the agent grades each company and contact against it with a rationale. Lead Scorer does this as part of its Outbound SDR run, so a two-person team gets a working two-level model without training anything.

Lead Scoring Model in 2026: Build a Two-Level Model That Survives the Death of the MQL

A B2B growth leader's post went around LinkedIn in mid-June with a confession in the first line: "I deleted our lead scoring model this morning." The CEO's reply, in the retelling: "You deleted... wtf we paid a consultant $30K to build that!" The punchline was the math — that model had produced "1,400 MQLs last year" and, out of them, 23 opportunities (LinkedIn, 13 Jun 2026).

That ratio is the whole problem with how most teams score leads, and it's why the conversation in 2026 keeps circling back to the same blunt take. As one demand-gen founder put it: "The MQL is dead. AI fills fake forms faster than scoring models catch them. The real qualification signal lives in the sales conversation." (X, 16 Jun 2026). The classic lead scoring model — points for an email open, points for a pricing-page visit, an MQL when the total crosses a line — has stopped correlating with revenue. This is a guide to building one that does.

We'll define what a lead scoring model actually is, show why the engagement-first version breaks, and walk through the model that's replacing it: a two-level model that scores the company and the decision-maker on real data, with AI doing the grading.

What a lead scoring model is

A lead scoring model is the framework you use to rank prospects by how likely they are to convert, so your team works the best ones first instead of treating every lead as equal. Every model, however it's built, is weighing two kinds of information:

Fit (explicit data) — who the lead is. Does the prospect match your ideal customer profile? Company size, industry, revenue, location, job title, seniority. This is the question "should we sell to them at all?"
Intent (implicit data) — what the lead does or signals. Recent behavior and trigger events: a funding round, a relevant hire, repeat visits, a demo request. This is the question "are they in-market right now?"

A good model combines both and subtracts for disqualifiers. The trouble is that one half — intent measured as clicks — is far easier to instrument than the other, so most models quietly become all engagement and no fit. That imbalance is exactly what produces 1,400 MQLs and 23 opportunities.

Why the engagement-first model broke

For a decade the default lead scoring model was a point table bolted onto marketing automation: +5 for an email open, +10 for a content download, +15 for a pricing-page visit, MQL at 50. It was attractive because every input was trackable. It's now failing for three reasons that have only gotten worse.

Engagement is trivially gamed. Bots open emails and trip link-tracking. Form fills can be fake or automated — that's the precise point behind "AI fills fake forms faster than scoring models catch them." A model that rewards clicks rewards the wrong behavior and the wrong leads.

The best-fit buyer often does nothing measurable. The economic buyer who matches your ICP perfectly may read one email on their phone and never click. Under an engagement model they score zero while a curious student who binge-downloads your ebooks becomes a hot lead. The model inverts priority.

Fit is the part that actually predicts revenue, and it's the part that gets skipped. As one widely shared take framed it in late May: "If your lead scoring model still runs on email opens and page views, it is outdated" — modern scoring "combines who someone is (fit) with what they've done (engagement)" rather than leaning on activity alone (LinkedIn, 28 May 2026). Fit is harder to compute because it requires real knowledge of the company and the person — which is exactly where AI changes the economics.

The two-level lead scoring model

The fix isn't a better point table. It's scoring two things that the classic model collapses into one: the company and the decision-maker. A lead is only hot when both clear the bar.

Level 1 — score the company (ICP fit)

First, grade the account independently of any single contact. Is this the kind of company you win? Industry and sub-sector, headcount, maturity, geography, and — for the French market — NAF code or OPCO eligibility. This is where verified, real data matters most: a model that scores a company on hallucinated or stale firmographics is worse than no model, because it's confidently wrong. Anchoring company fit on official sources (in France, the State registry — recherche-entreprises, SIRENE, the INPI RNE — with a real SIREN and the actual director) removes that failure mode.

Level 2 — score the decision-maker

A perfect-fit account is not a lead until you've found the right person inside it. Score the individual on role, seniority, and whether their responsibilities map to the pain your product solves. A great company with the wrong contact — too junior, wrong function, just left the role — is not sales-ready, and a two-level model says so instead of firing an MQL because someone with a company email opened a newsletter.

Combine, then subtract

The final score is a function of both levels, with negative scoring on top: competitors, out-of-ICP industries, free-mail addresses, roles that can't buy, and stale data all cap or kill the score. The output you want isn't a naked number — it's a ranked lead with a one-line reason a rep can read in two seconds. The same logic underpins good predictive lead scoring: weight fit heavily, treat engagement as a tiebreaker, and never let activity alone manufacture a qualified lead.

Rules vs predictive vs AI-agent scoring

There are three ways to actually compute the score, and they sit on a spectrum of effort versus intelligence.

Rules-based. You hand-write the point table. Transparent and fast to set up, but brittle and subjective — it encodes your assumptions, not reality, and it's the version that produces the 1,400-to-23 outcomes. Fine as a v1, dangerous as a permanent answer.
Predictive (machine learning). You train a model on your closed-won history so factor weights are learned, not guessed. Genuinely better — but it needs a large, clean dataset and someone to maintain it, which most teams under a few hundred deals a year simply don't have.
AI-agent scoring. An LLM reads the company and the person the way a sharp SDR would and grades them against your ICP in plain language. No years of training data required, and crucially it returns a rationale, not an opaque probability. This is the route that makes a fit-first, two-level model practical for a small team on day one.

The teams getting real lift are pairing AI grading with outcome data. One operator described connecting lead scoring "to actual customer LTV" and watching qualification improve by 35% once the model learned which "expensive leads" were actually the best customers (X, 16 Jun 2026). The model isn't the point — the loop between scoring and revenue is.

How Lead Scorer runs a two-level model for you

Lead Scorer was rebuilt around this exact model. Instead of a point table you tune in a marketing automation tool, scoring is one stage of an Outbound SDR agent run that goes from a plain-language brief to ready-to-launch outreach — and you approve each step.

You brief it like a human SDR: who you target, what qualifies a lead, and what disqualifies one. The agent discovers companies from real, official data — the web plus the French State registry, so firmographics are verified, not invented. Then it scores at both levels: the company on ICP fit and the decision-maker 0-10 against your product description, each with a written rationale, and it rejects off-target leads with the reason attached rather than passing them through. Two supporting agents feed the model: Find Key People in a List of Companies turns a target-account list into the right scored contacts, and Find People by Context builds a fresh segment from a description like "heads of RevOps at Series-B SaaS." A second LLM (Mistral) then reviews the drafted messages before you ever see them. You get a transparent, replayable run — discovery, scoring, review, ready-to-launch — instead of a black-box number in a CRM field.

Common lead scoring model mistakes

Scoring engagement only. The original sin. If your model can produce a hot lead from clicks alone, it will, and the clicks won't be from buyers.
No negative scoring. Without disqualifiers, every model drifts upward and the threshold becomes meaningless. Competitors and students will out-engage your real ICP.
One level, not two. Scoring the person without the company (or vice versa) is how you get great contacts at companies you'd never sell to, marked "hot."
Scoring on stale or invented data. A fit score is only as good as the firmographics under it. Hallucinated company data makes a model confidently wrong; refresh from real sources.
Setting the threshold by gut. Anchor it to the scores your actual closed-won deals had at handoff — see the MQL-to-SQL handoff — not to a round number that feels right.

The takeaway

The lead scoring model that worked in 2018 — points on opens and clicks, MQL at 50 — is the one teams are publicly deleting in 2026, because it manufactures volume that sales ignores. The model replacing it scores fit first and at two levels: the company against your ICP on verified data, and the decision-maker against your product, with negative scoring to kill the noise and a rationale instead of a bare number. Whether you build that as a rules-based v1, train a predictive model, or let an AI agent grade it, the principle is the same — qualify on who they are, use engagement as a tiebreaker, and close the loop with real outcomes.

Want a two-level model running without building a point table? Try Lead Scorer free → or see pricing. Further reading: the AI lead scoring guide · define your ICP with AI.