AI lead scoring: prioritise leads automatically

Most sales teams work through inbound leads in the order they arrive, not in the order of their value. If you get 150 leads a week and maybe 15 of them can actually buy, first-come-first-served means most of your sales hours go to the wrong people while the good leads cool off. AI lead scoring flips the sequence: score first, call second.

AI lead scoring rates every inbound lead automatically on fit (who they are) and intent signals (what they do), and assigns a priority before anyone on your team spends time on it. Predictive models learn from the won and lost leads in your CRM, while a language model can classify free-text form answers on top. The score only pays off once it drives routing, follow-up and the bidding signals you send back to your ad platforms.

Rule-based vs. predictive lead scoring: what actually changes?

Rule-based scoring is a points system you define yourself. A "Managing Director" job title earns 15 points, a pricing page visit earns 10, a free-mail address costs 10. It is transparent and quick to set up, but the weights are guesswork. Nobody checks whether a pricing page visit really correlates with closed deals in your market.

Predictive scoring works the other way round. A model looks at which attributes historically showed up in leads that became customers, and which showed up in leads that went nowhere, and turns that into a probability or a score. Salesforce, for example, describes Einstein Lead Scoring as rating leads from 1 to 99 based on how well they fit your past conversion patterns. HubSpot takes a middle path: its AI analyses past contacts and suggests criteria and point values that you can still edit before switching the score on.

The third, newer option is classification with a large language model. It does not replace a statistical model, but it handles something points systems never did well: free text. Whether someone describes a concrete project with a timeline in the "How can we help?" field, or is really sending a job application, is something an LLM can tell reliably enough to pre-sort your inbox.

Fit and engagement signals: what belongs in the score?

A useful score separates two questions. Does this contact fit us at all? And are they showing buying intent right now? HubSpot's scoring tool mirrors this with separate fit and engagement scores plus a combined score shown as a matrix from A1 to C3. The letter reflects fit, the number reflects engagement. An A1 lead fits and is active; a C1 lead is active but outside your target market.

Typical fit signals in B2B are industry, company size, the contact's role, country and tech stack. If you sell to trade or wholesale customers online, order volume and customer type matter too. Engagement signals include visits to pricing and comparison pages, demo requests, repeat visits within a few days, email replies and the campaign the lead came from. Time matters: a whitepaper download eight months ago says little about today. HubSpot offers score decay for engagement groups, applied linearly over intervals of 1, 3, 6 or 12 months.

Negative signals matter as much as positive ones. Job seekers, vendors pitching you, students doing research and competitors will distort any score unless you actively subtract them.

Your CRM data is the model: no clean stages, no prediction

Every predictive model needs a clear outcome to learn from, and this is where most setups break. If your CRM does not reliably record which leads were qualified, won or lost, the model learns from noise.

The vendors are specific about minimum volumes. HubSpot's AI-built contact scores require Marketing Hub Enterprise and at least 50 contacts, 25 converted and 25 not converted. Salesforce's Einstein implementation guide asks for at least 1,000 leads created in the last 200 days, at least 120 of which were converted, before it builds a model on your own data. Below that, Einstein falls back to a global model trained on anonymised data from many Salesforce customers and switches to your own model once that performs better. According to the guide, scores are re-analysed roughly every ten days.

Pipedrive is a different story. Its help centre says the Scores feature rates deals only, is available on Premium plans and above, and is rule-based: criteria add 25 or 10 points or subtract 10, with one active score per pipeline. The documentation does not describe an AI component, so predictive scoring in Pipedrive means a third-party tool or your own automation.

How to roll out AI lead scoring in five steps

Step 1: Define the target event. Decide what a "good lead" is in measurable terms. For most B2B teams that is not the form fill but a sales-qualified lead or a deal that reaches a certain stage. This event is what the model trains on and what your ads will later optimise towards.

Step 2: Clean the data. For the last 6 to 12 months, check that every lead has a lifecycle stage and, where relevant, a loss reason. Fill gaps in company data and standardise industry and size fields. Skip this step and every number that follows is meaningless.

Step 3: Switch on the native model. Use what your CRM already offers: HubSpot's AI suggestions on Enterprise, Einstein Lead Scoring in Salesforce, rule-based scores in Pipedrive as a starting point. Write down your thresholds, for example: 70 and above goes to sales within the hour, 40 to 69 enters a nurture track, everything below goes to the newsletter.

Step 4: Add LLM classification for free text. n8n has a Text Classifier node for exactly this. You define categories with a short description, such as "project enquiry with budget", "general question", "job application" or "vendor pitch", and you can route unclear cases to a separate "Other" branch instead of discarding them. Make does the same through its OpenAI integration. Write the result to its own CRM field rather than straight into the score, so you can measure its accuracy separately. For what happens after routing, see our guide on automating lead follow-up.

Step 5: Feed the score back into your ad accounts. This is where the performance upside lives. For lead generation, Google recommends uploading offline conversions as often as possible, ideally daily, attaching values to lead stages and using value-based bidding when leads can be valued differently. Enhanced conversions for leads match hashed data such as email addresses. Note that, according to Google's help centre, these uploads move from the Google Ads API to the Data Manager API from 15 June 2026. On Meta, the Conversions API for CRM connects your lead stages with the Conversion Leads performance goal. Meta's prerequisites: at least 200 leads a month, at least one upload per day, the target stage reached within 28 days of the lead and a conversion rate between 1 and 40 percent. The goal currently works only with Lead Ads using Instant Forms.

A worked example: what prioritisation means in numbers

Take a simplified B2B software company: 400 leads a month, 5 percent become customers, so 20 deals. Without scoring, the team spreads its time evenly, about 20 minutes of first qualification per lead, roughly 133 hours in total.

Now split leads by score. Suppose the top quarter (100 leads) contains 14 of the 20 eventual customers. Conversion there is 14 percent, in the rest 2 percent. Give top leads 40 minutes each and put the other 300 into an automated track with a 5-minute check, and effort drops to about 92 hours, with the time going where 70 percent of revenue originates.

The formula is simple. Concentration = share of deals in the top segment ÷ share of leads in the top segment. In the example, 70% ÷ 25% = 2.8. A value around 1 means the score separates no better than chance. The further above 1 you get, the more it pays to treat the segments differently.

Which KPIs tell you the scoring works?

Don't measure whether the score looks plausible. Measure whether it improves decisions. Five metrics are enough.

Conversion rate by score band shows whether high scores really close more often. Concentration, as above, shows how sharply the score separates. Time to first contact for top leads shows whether routing works. Sales acceptance rate, the share of handed-over leads that sales confirms as qualified, shows whether marketing and sales agree on what a good lead is. And cost per qualified lead in your campaigns shows whether the feedback loop into your ads is doing its job.

Review these monthly against a baseline period before launch. If conversion in the top band drops, your audience, offer or campaign mix has usually shifted, and the model needs retraining or the rules need adjusting.

GDPR: what to watch when you score people

B2B contacts are still natural persons, so lead scoring will usually count as profiling under Article 4(4) GDPR. Article 22 gives people the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. In its SCHUFA ruling of 7 December 2023 (C-634/21), the Court of Justice of the EU held that generating a score can itself be such a decision when a third party's decision depends decisively on it.

In practice, a score that only sets the order in which leads are handled is not the same as a score that rejects enquiries automatically. Keep a person in the loop for edge cases, document the logic, and cover the processing in your privacy notice and records of processing. Article 13 requires information about automated decision-making, and Article 21 grants a right to object that explicitly includes profiling for direct marketing. This is not legal advice; agree the setup with your data protection officer before going live.

If you want scoring, routing and the ad feedback loop built as one system, take a look at our AI automation service. For clean offline conversion measurement, see tracking & analytics.

FAQ

How many leads do I need for predictive lead scoring?

The vendors set minimums: HubSpot needs at least 50 contacts, 25 converted and 25 not converted; Salesforce needs 1,000 leads in 200 days with 120 converted for its own Einstein model. Below that, a well-maintained rule-based model plus LLM classification for free text is usually the more honest option.

Can a language model handle the whole score on its own?

Not sensibly. An LLM is good at classifying free text, but it does not automatically learn from your win/loss data. Use it as one extra input alongside fit and engagement signals and track its accuracy separately.

How do I get the score into Google Ads and Meta?

Through offline conversions with values per lead stage. On Google via enhanced conversions for leads, using the Data Manager API from 15 June 2026. On Meta via the Conversions API for CRM combined with the Conversion Leads goal for Instant Forms campaigns.

Sources

Want more of the attention you get to turn into customers? Let’s spend 15 minutes on your setup, free and without a pitch deck. Book your free call

AI lead scoring: prioritise inbound leads before your sales team wastes time