What lead scoring is and when it pays off
Lead scoring is an order, not a prophecy. It helps you pick, out of 50 enquiries, the five to call first where a conversation is most likely to pay off. No more, but that’s a lot.
| Situation | Do you need scoring? |
|---|---|
| A few enquiries a week, you talk to everyone anyway | No, a quick personal look is enough |
| Dozens of enquiries, sales can’t keep up | Yes, a simple point model |
| Hundreds of leads a month, many closed deals in the history | Yes, predictive scoring can pay off |
| Many leads but few closed deals in the history | Point model first; predictive models have too little to learn from |
The point model to copy: fit plus behaviour
A simple model separates two questions: does the contact fit your offer (fit)? And are they showing interest right now (behaviour)? Together they give up to 100 points.
Fit (up to 50 points)
| Attribute | Points | Example B2B service provider |
|---|---|---|
| Industry matches your best customers | 15 | Software, agencies, retail |
| Company size in the target range | 15 | 10–200 employees |
| Role with decision authority | 10 | Managing director, head of marketing |
| Region you serve | 10 | Germany, Austria, Switzerland |
Behaviour (up to 50 points)
| Signal | Points | Note |
|---|---|---|
| Enquiry via contact form or meeting booking | 25 | Strongest signal |
| Visited pricing or offer page | 10 | Only measurable with consent |
| Newsletter click on an offer topic | 10 | From the newsletter tool |
| Downloaded a guide | 5 | Interest, but early in the decision |
Deductions and thresholds
- Minus 20: competitors, students, job seekers (recognisable from free text).
- Minus 10: no activity for 90 days.
- 60 points and up: personal contact within one working day.
- 30–59 points: newsletter and relevant content, re-score after four weeks.
- Below 30: no active sales, but every enquiry gets a reply.
The numbers are a starting point, not a law of nature. After three months, compare: did high-scoring leads actually buy more often? If not, adjust the weights.
Example: two enquiries, two scores
Two made-up but typical enquiries to a B2B service provider show how the model works.
| Attribute | Enquiry A | Points | Enquiry B | Points |
|---|---|---|---|---|
| Industry | Software company | 15 | Retail outside the target industries | 0 |
| Size | 40 employees | 15 | 3 employees | 0 |
| Role | Head of marketing | 10 | Intern | 0 |
| Region | Munich | 10 | Vienna | 10 |
| Behaviour | Meeting booking + pricing page | 35 | Guide download | 5 |
| Total | 85 | 15 |
Enquiry A gets a call the same day. Enquiry B gets a friendly reply with relevant content and joins the newsletter if they consented. Both are taken seriously, but at different speeds and levels of personal attention.
Buying signals that need no tracking
Many of the strongest signals aren’t in an analytics tool but in the inbox and the conversation. They’re often more reliable than page views.
- Concrete questions: about price, process, start date or contract term.
- Timing: “We need this by …” or an occasion such as a launch.
- Several people: two contacts from the same company get in touch.
- Fast replies: responses within hours rather than weeks.
- Own documents: the person sends numbers, a brief or their current solution.
AI can pick these signals out of emails and call notes and suggest points. Whether they apply should be decided by the person who knows the contact.
Where scoring models go wrong
- Old data: a model learning from customers two years ago favours audiences that may no longer fit.
- Channel bias: if only one channel was used so far, the model rates leads from new channels lower, even if they’re good.
- Self-fulfilling prophecy: high-scoring leads get more attention and therefore buy more often; the model looks better than it is.
- Missing data: people who decline tracking have fewer behaviour points but aren’t less interested.
Countermeasures: review weights regularly, still contact a small sample of low-scoring leads personally, and don’t overweight behaviour points.
Where AI helps in your own model
- Classify free text: AI reads enquiries and suggests industry, role, urgency and possible exclusions, with a one-sentence reason.
- Summarise: two sentences for sales: what does the person want, why now?
- Check weights: with an (anonymised) CRM export, have AI compare which attributes were more common among won and lost leads, with a visible calculation.
- Draft a reply: a first-response draft that a person checks and sends.
Important: have the AI choose from fixed categories and allow “unclear”. Free-form ratings vary from run to run. How to build this into workflows is shown in marketing automation with AI.
Marketing and sales: shared definitions
Scoring often fails not because of the model but because marketing and sales mean different things by “good lead”. Agree together:
| Term | Meaning | Who decides |
|---|---|---|
| Lead | Any contact with consent or an enquiry | Automatic |
| Marketing-qualified | Score above the first threshold, fit given | Model, marketing checks |
| Sales-qualified | Sales has talked and confirmed the need | Sales |
| Disqualified | No need, competitor or no contact possible, with a reason | Sales |
The reason for disqualified leads is gold for the next model adjustment. Have AI summarise the reasons once a month: which patterns appear?
Scoring in shops and B2C
B2C rarely has sales conversations, but the logic still helps, e.g. for newsletters and offers. A proven pattern is grouping by recency, frequency and value of purchases.
- Recency: when was the last purchase?
- Frequency: how often did they buy?
- Value: how much did they spend in total?
That gives groups like “loyal regulars”, “haven’t bought in a while” or “bought once, expensive”, each getting different emails. AI helps calculate with code and draft the emails. Prices or discounts shouldn’t automatically differ by score without a legal check.
Predictive scoring in CRM systems
Predictive scoring means a model learns from your past leads and deals which attributes point to a purchase and assigns scores itself. It only works with enough history.
| Vendor | Feature per vendor | Requirement | Source |
|---|---|---|---|
| HubSpot | Predictive lead scoring: likelihood a contact becomes a customer within 90 days, plus priority in four equal tiers | Marketing Hub or Sales Hub Enterprise; HubSpot itself calls the model a black box | HubSpot |
| Salesforce | Einstein Lead Scoring rates leads by similarity to previously converted leads | Per the implementation guide, at least 1,000 leads in the last 200 days per segment; Enterprise (extra cost), Performance, Unlimited | Salesforce |
| Pipedrive | Pulse: AI feature to prioritise leads by buying intent | Announced in 2024 as a closed beta with a waitlist; check current status in your own account | Pipedrive |
As of Nov 15, 2026, per vendor pages. Please check current pricing pages for prices.
The upside of predictive models: they find relationships you wouldn’t have guessed. The downside: you often can’t see why a lead scores high. For transparency towards the people affected and for your sales team’s trust, that’s a real problem.
GDPR: Art. 22, profiling and transparency
| Topic | What applies | Source |
|---|---|---|
| Automated decisions | Art. 22: people have the right not to be subject to a decision based solely on automated processing with legal or similarly significant effects | GDPR |
| Notice | Art. 13 and 14: information on purposes, legal basis and, for Art. 22 decisions, the logic involved | GDPR |
| Objection | Art. 21(2): objection to processing for direct marketing, including profiling, at any time | GDPR |
| Legal basis | Art. 6: consent, contract or legitimate interest, checked for the specific purpose | GDPR |
For typical marketing scoring: as long as the score only sets the order and a person decides whom to approach and how, there’s usually no decision based solely on automated processing. It becomes critical when a score automatically denies offers, prices or access to services.
- Explain in the privacy policy that and how leads are scored.
- Use only the data needed for scoring; no sensitive data.
- Website behaviour data only with consent.
- Allow and honour objection to profiling for marketing.
- For marketing emails to scored leads, German law (§ 7 UWG) also requires consent.
Rolling it out in four weeks
| Week | What you do | Result |
|---|---|---|
| 1 | Look at the last 30 to 50 wins and losses: what did the winners have in common? | First list of fit attributes |
| 2 | Write down the point model, set thresholds, agree it with sales | Model on one page |
| 3 | Implement in the CRM or a spreadsheet, test AI for free-text classification | Scores for all new leads |
| 4 | Review: were high-scoring leads contacted faster? Did they buy more often? | Adjusted weights |
After three months, compare the close rate per score tier. If it rises with the score, the model works. If it’s the same everywhere, it’s measuring the wrong thing.
From practice: enquiries at our end
At i4u.business, enquiries come through the contact form, the marketing office hours and our AI tools. Each enquiry is first saved on our server and then delivered; the form asks for the topic so the classification is clear from the start. Every enquiry is assessed and answered by a person.
At our volume, a scoring model would be overkill. The lesson still holds: the most important information for classification comes from a well-posed form field, not an algorithm.
Checklist for AI lead scoring
- Workflow clear: who contacts high-scoring leads, how fast?
- Winners and losers of recent months analysed
- Point model from fit and behaviour with thresholds written down
- AI only with fixed categories and an “unclear” case
- Every enquiry gets a reply, including low scores
- Behaviour data only with consent
- Privacy policy explains the scoring
- Objection to profiling for marketing possible
- Review after three months: close rate per score tier
12 prompts for lead scoring
Here’s anonymised data on won and lost leads: [table]. Which attributes are clearly more common among winners? Calculate with code and show the numbers.
My offer: [description]. My best customers: [attributes]. Draft a point model with fit (up to 50) and behaviour (up to 50), deductions and thresholds.
Here’s my point model: [model]. Where might it be systematically wrong? Which leads would be scored too high or too low?
Classify this enquiry: industry, company size, role, urgency (high/medium/low/unclear), exclusion reason yes/no. One sentence of reasoning each: [enquiry].
Summarise this enquiry in two sentences: need and why now. Don’t invent anything: [enquiry].
Write a reply draft to this enquiry, max 100 words, with one concrete next step: [enquiry].
With [number] leads a week and [hours] of sales time: from which score should we follow up personally so the time suffices?
Here are close rates per score tier after three months: [data]. Does the model work? Which weights should I change?
Draft a clear privacy policy section explaining that enquiries are prioritised by [criteria] and a person decides. For review by a professional.
Which one or two questions in the contact form would most improve enquiry classification without putting prospects off?
We have [number] leads a month and [number] closed deals in our history. Is predictive scoring worth it, or is a point model enough? Explain.
Which patterns in enquiries suggest competitors, students or job seekers? List signals to check without automatically rejecting anyone.
The most common mistakes
- Scoring without a workflow. A high score is useless if nobody calls.
- Too many attributes. Four fit and four behaviour signals are enough to start.
- Predictive without history. Without enough closed deals, the model learns noise.
- Ignoring low scores. Every enquiry deserves a reply.
- Black box without explanation. Sales and the people affected should understand why.
- Never adjusting. Without a three-month review, the model may measure the wrong thing.
Frequently asked questions
How many attributes should my model have?
Four for fit and four for behaviour to start. More attributes don’t make the model more accurate, just harder to understand and maintain.
Can AI reject leads automatically?
Better not. AI can flag competitors or unsuitable enquiries; the decision and a friendly reply belong with a person. Automatic rejections with significant effects fall under Art. 22 GDPR.
What about leads without a score?
Every enquiry gets a reply first. If data is missing for a score, a quick personal look or a follow-up question beats an estimate.
Which signals show real buying interest?
Concrete questions about price, process or start, a time reference, several contacts from the same company, fast replies and documents they send along. These signals are often in the inbox, not the analytics tool.
Can a scoring model be unfair?
Yes, e.g. when it learns from old data, disadvantages new channels or rates people without tracking consent lower. Regular reviews and personal spot checks on low scores help.
What’s the difference between an MQL and an SQL?
A marketing-qualified lead (MQL) has reached the scoring threshold and fits the offer. A sales-qualified lead (SQL) is one where sales has talked and confirmed the need. Marketing and sales should define both terms together.
Does lead scoring work for online shops?
Yes, usually as grouping by recency, frequency and value of purchases. That gives customer groups who get different emails and content.
What is lead scoring?
Rating contacts with points so sales approaches the most promising first. Usually a combination of fit (does the contact match the offer?) and behaviour (how active are they?).
Do I need AI for it?
Not necessarily. A simple point model in a CRM or spreadsheet is enough for most small teams. AI helps classify free text and check weights; predictive scoring only pays off with lots of history.
Which HubSpot edition has predictive lead scoring?
Per the HubSpot Knowledge Base, Marketing Hub Enterprise and Sales Hub Enterprise. The model estimates the likelihood that a contact becomes a customer within 90 days.
What does Salesforce Einstein Lead Scoring need?
Per the Salesforce implementation guide, at least 1,000 leads in the last 200 days per segment. Available in Enterprise at extra cost, and in Performance and Unlimited.
Is lead scoring allowed under the GDPR?
Usually yes, with a legal basis, transparent information and a person making the decision. Art. 22 GDPR protects against decisions with significant effects made solely by automated means. Not legal advice.
Do I have to tell leads about scoring?
Your privacy policy should explain that and by which criteria enquiries are prioritised. Under Art. 21(2) GDPR, people can object to profiling for direct marketing at any time.
How do I check whether my model works?
After about three months, compare the close rate per score tier. If it rises with the score, the model works; if it’s the same everywhere, adjust the weights.
Sources
- HubSpot: Predictive lead scoring – Enterprise, 90 days, updated Jan 11, 2026
- Salesforce: Sales Cloud Einstein Implementation Guide – Einstein Lead Scoring requirements
- Pipedrive: Pipedrive Pulse – Sep 16, 2024, closed beta
- GDPR (Regulation (EU) 2016/679) – Art. 6, 13, 14, 21, 22
- § 7 UWG (Germany) – Marketing email






