AI lead scoring: a practical guide

Open most CRMs and you'll find a lead score that nobody trusts. It was a point table someone built in an afternoon — +10 for a demo request, +5 for an email open, +1 for a webinar from eight months ago — and it has been quietly drifting from reality ever since. Sales ignores it. Marketing defends it. And the warmest lead on the list is buried under a researcher who happened to open four emails. This guide is about fixing that: how AI lead scoring actually works, what signals matter, and how to wire a score so it changes what your team does next instead of decorating a column.

What AI lead scoring actually is

At its core, a lead score answers one question: of everyone in the pipeline, who is most worth a human's next hour? Traditional lead qualification answers it with a fixed point table — manual weights, set once, rarely revisited. AI lead scoring answers it with a model that learns from your real outcomes, weighs many signals at once, and re-ranks continuously as new behaviour arrives. The output isn't a vibe; it's an ordered queue.

The shift matters because buying behaviour is bursty and time-sensitive. A prospect who visited your pricing page twice this morning is a different person from the one who downloaded a PDF last spring — even if a static table gives them the same number. A real lead scoring model treats recency as first-class, which is exactly where point tables fall apart. That recency-aware, outcome-driven ranking is what people mean when they say predictive lead scoring: not a tally of past activity, but a live estimate of who is buying now.

The anatomy of a good score

A score that survives contact with reality is a composite. It blends three ingredients, and getting the balance right matters more than the model you pick. In dolv these map directly to the signals already flowing through your AI sales stack and CRM, so you're not standing up a separate scoring silo — you're scoring on data the command center already holds.

Fit — who the lead is

Firmographics and role: company size, industry, seniority, region. Fit is what stops you over-scoring an intern and under-scoring a perfect-fit buyer who is simply early. It's the stable backbone of lead qualification — slow-moving, but it sets the ceiling on how interesting a lead can ever be. A junior researcher at a target account is worth knowing about; they are not worth interrupting a rep's morning for.

Intent — what they just did

The behavioural layer: pricing-page visits, repeat sessions, demo or contact requests, replies to outreach, high-intent page views. These intent signals are the fastest-moving, highest-value inputs in any lead scoring model, because they reflect a decision forming right now rather than a profile that fit last quarter. Intent-based lead scoring lives or dies on how well you capture this layer — and on whether you can see it the moment it happens.

Recency — how fresh the intent is

Time-decay is the quiet ingredient that separates predictive from decorative. A live pricing visit should outrank a three-month-old download, automatically. Decay keeps the queue honest so your team always works the warmest lead, not the one with the longest history. Without it, every old action accumulates forever and your highest "scores" become a hall of fame for tire-kickers.

Why a composite beats any single number

Fit without intent gives you a perfect-profile lead who isn't ready. Intent without fit gives you a curious tire-kicker. Recency without either gives you whoever clicked most recently. The art of predictive lead scoring is combining all three so a high score means fit buyer, acting now. That's the lead a human should reply to before lunch — and it's a very different list from "highest point total this quarter."

How scoring works inside the dolv command center

Scoring on a spreadsheet is a dead end — the number lives somewhere your reps don't, and nothing acts on it. The point of intent-based lead scoring is that the score and the action live in the same place. dolv runs intent-signal lead scoring inside the same grounded command center as your full CRM and 27 read+write integrations, so a score immediately drives real work: rank the queue, update a CRM field, create a task, prepare a follow-up. The score stops being a column you read and becomes the thing that decides which lead becomes a prepared email next.

Because the whole thing is grounded, the follow-up it prepares already sounds like your company — product, positioning, voice — instead of generic outreach. That grounding is a topic in its own right; if your scored leads deserve on-brand replies, see how to keep AI on-brand. And the difference between a score that merely fires a rule and one that reasons about context is the same line we draw in AI agents vs marketing automation: rules trigger, agents reason — and a scoring model that drives action beats a static table that just counts clicks.

Score automatically, route with a human

Here's the trap teams fall into: they let the model score and auto-send. A scoring miss then becomes a cold, badly-timed email in a real buyer's inbox. dolv splits the two. Scoring, ranking, CRM updates, and task creation are internal and reversible, so they run immediately. Anything that reaches a prospect is prepared by a grounded agent and queued for your sign-off in a clear prepare → approve → executing → done flow. You move at machine speed on the bookkeeping and stay human on the message — which is the heart of human-in-the-loop routing.

That gate is also how the model gets better. When you adjust a prepared message — or skip a lead the score ranked highly — you're labelling reality, and the system learns which signals deserve more weight. Multi-agent campaigns can run the volume behind that gate, each agent scoped to a role with a $250/mo budget cap and a full run history, so speed never means losing the audit trail. More on where that line sits — and what AI can and can't run on its own — is in can AI run marketing campaigns on its own?

Feed the score: connect your stack

A scoring model is only as good as the signals it can see. Fit needs CRM data; intent needs analytics and engagement; recency needs all of it streaming in continuously. dolv's 27 read+write integrations cover both halves — CRM and email through Gmail and Outlook, behaviour through GA4 and Search Console, plus Sheets, Docs, LinkedIn and more — so neither side of the score is guessing.

• Connect the CRM so fit signals — role, company, stage — feed the score with real firmographics.
• Wire analytics (GA4, Search Console) so pricing visits and high-intent pages become live intent signals.
• Crawl your site into the knowledge base so the outreach a score triggers is grounded in your real copy.

If you're weighing a dedicated signal tool against an all-in-one command center, the practical trade-offs are laid out in our dolv vs Clay comparison — the short version is that scoring inside the system that also acts on the score beats scoring in a tool that hands you a number and walks away.

Prove the score predicts revenue

A lead score is a hypothesis: "these leads convert better." You should test it like one. Because scores in dolv flow through the same funnel intelligence as everything else, you can watch whether high-scored leads actually convert against a rolling 30-day baseline. The unified TOFU/MOFU/BOFU model rolls into one weighted composite health number — TOFU 0.25, MOFU 0.40, BOFU 0.35 — and the cross-metric correlation engine tells you which inputs earn their weight. If "repeat pricing visit" correlates with BOFU conversion and "email open" doesn't, you have a reason to re-weight — backed by a real z-test, not a hunch.

Frame each change as an ICE experiment against a North Star metric, let it run behind the approval inbox, and read the result. Pair it with multi-touch attribution across five models and you can finally answer the only question that matters about a score: does ranking leads this way put more revenue in the pipe? That feedback loop — score, act, measure, re-weight — is what turns lead scoring from a static config into a system that improves every month.

A six-step AI lead scoring checklist

1. List the signals you can actually see. Separate fit (CRM) from intent (behaviour) from recency. You can't score what you don't collect.
2. Connect the sources. Wire CRM and analytics so both halves of the score are grounded in live data, not last quarter's export.
3. Build a composite, not a tally. Combine fit and intent, and apply time-decay so fresh behaviour always outranks stale activity.
4. Drive an action from the score. Rank the queue and prepare the top lead's follow-up — a number nobody acts on is just decoration.
5. Keep a human on the send. Let internal updates run; route anything reaching a prospect through prepare → approve → executing → done.
6. Validate against the funnel. Confirm high scores predict BOFU conversion, prune the inputs that don't correlate, and re-weight.

Do these six and your lead score stops being the column everyone ignores. It becomes the order your team works in — and a number you can defend with your funnel, not your gut. That's the whole promise of AI lead scoring done right: not a smarter sorting trick, but an operator that ranks the pipeline, prepares the next move, and proves it worked. dolv it.