Lead scoring

Sales teams get far more leads than they can call, and most of them never buy. Lead scoring learns from past won and lost leads which properties and behaviours predict conversion. It gives every new lead a probability, a grade and the reasons behind the score, so reps start with the best leads and marketing nurtures the rest. Rule-based point systems have existed for 20 years. What has changed is that the main CRMs now train the model for you once you have a few dozen outcomes. LLMs now also read the free text and fill in missing firmographics, and agents act on the score straight away.

Problem & value

• Today: Marketing qualifies leads (MQL) with hand-maintained point rules: +10 for a pricing-page visit, +5 for an email open, −20 for a student address. The rules drift, nobody checks them against outcomes, and sales stops trusting them. Reps then pick leads by gut feeling or simply take them in order of arrival.
• AI value:
• Scores are trained on real outcomes and recalibrate themselves.
• Fit and intent are measured separately.
• Reason codes make the score easy to act on.
• Good leads are routed faster, low-fit leads are disqualified early, and reps spend less time on dead leads.
• Low entry barrier: Dynamics 365 Sales needs only 40 qualified and 40 disqualified leads, closed within a training window of 3 months to 2 years, to train a predictive model. It can retrain automatically every 15 days [1]. HubSpot's AI-built fit/engagement scores need Marketing Hub Enterprise and at least 50 contacts, 25 of them converted and 25 not converted [2]. Odoo CRM's predictive probability is always on and is computed with naive Bayes from won/lost history [3].
• Measured results: published figures are almost all vendor-reported. For example, 6sense reports a 13 % higher win rate and 18 % higher average opportunity value for its "6QA" accounts than for other accounts in its 2025 customer data. That is a correlation without a control group, not a causal effect [4]. Measure your own lift with a holdout (see KPIs).

Variants

• SMB B2B (Mittelstand): Rule-based fit scoring plus engagement scoring in the CRM or marketing automation tool. Predictive scoring is switched on once enough outcomes have been closed. Website visitor identification at company level is common (e.g. Dealfront, see Path A).
• Enterprise B2B / account-based marketing: Account-level scoring and buying stages, using first-party and third-party intent data (6sense, Demandbase, Bombora Company Surge). The scores feed the buying-group view and advertising audiences [4][5]. Trigger-event extraction from news, job postings and funding rounds is covered in LANG-017. Acting on the score with outreach is AGENT-007.
• Product-led growth (SaaS): Product-qualified leads (PQL) are scored on in-product usage (seats, feature activation, usage limits reached) as well as firmographics.
• B2C high-ticket: Real estate, car dealers, solar/heat-pump installers, education and private clinics score enquiries from web forms and portals. This is personal data about consumers, so see the regulation section.
• Financial services: Prospect propensity scoring for loans and insurance. It must stay separate from creditworthiness or risk assessment (PRED-009, PRED-010), which is high-risk under the AI Act.
• Non-profit: Donor propensity scoring follows the same pattern. Private use is not relevant.

How it works

1. CRM + marketing automation + web/product events (+ intent data, enrichment)
2. identity resolution (lead ↔ contact ↔ account, dedup)
3. feature building: fit (firmographics, role), engagement (recency/frequency, decay), text signals
4. label definition (lead
5. qualified / opportunity / won within N days)
6. model: gradient boosting or tabular foundation model, time-based split, probability calibration
7. explanation: top factors per lead (SHAP / reason codes)
8. score bands + routing rules
9. CRM fields, rep queue, SLA, nurturing journey
10. feedback: outcomes flow back
11. monitoring (AUC, lift, drift)
12. scheduled retraining

• Core model: A binary classifier on tabular data. Common choices are LightGBM, XGBoost, CatBoost or logistic regression for small datasets. Tabular foundation models such as TabPFN reach strong accuracy on small and medium datasets without tuning [6].
• Fit and engagement as separate scores: HubSpot's A–C × 1–3 grid shows the pattern. The letter is fit, the number is engagement, and engagement points decay over 1, 3, 6 or 12 months [2].
• LLM role (supporting, not core):
• Classify free-text form fields and emails (budget mentioned? student? competitor?).
• Map company descriptions to an ideal customer profile (ICP), using embeddings for lookalikes.
• Enrich missing firmographics from the company website.
• Write a one-line "why this lead" summary for the rep.
• The LLM should not produce the probability itself. Calibrated tabular models are more accurate and easier to audit.
• Human-in-the-loop:
• Reps can override grades and give disqualification reasons. These become training labels.
• Marketing and sales jointly define the threshold and SLA (when an MQL becomes an SQL).
• Outreach to high scores is reviewed by a person or follows an approved playbook.

Data preparation

The model learns from historical leads with a known outcome (qualified, won, lost), so the CRM is the core source (HubSpot, Pipedrive, Salesforce, often plus Excel lists, the marketing automation tool, web analytics and the mailbox). Typical SMB gaps: outcomes are not maintained ("lost" reasons empty), duplicate accounts and contacts, free-text industry fields, leads created by hand without source, and call notes that live in personal mail or notebooks. Consent and opt-in status are often scattered across forms and newsletter tools.

Preparation steps: (1) deduplicate and merge accounts and contacts (built-in CRM tools, OpenRefine), (2) enforce mandatory fields for source, status and loss reason and backfill at least 12-24 months of history, (3) normalise industry (NACE) and company size, (4) join web and email engagement events to contacts, (5) separate personal data from the features actually needed and pseudonymise exports, (6) hold back a recent time window as test set. Useful PREP cases: PREP-005, PREP-011, PREP-020, PREP-022, PREP-025. See also DATA-003 for deduplication.

Rough share of effort: about 40-50 percent of the project. Many SMBs have too few closed deals (fewer than a few hundred) for a trained model; then a rule-based score built jointly with sales is the honest start. See GUIDE-003.

Implementation paths

Which path for whom

• SMB — Path A: CRM-integrated scoring (HubSpot, Pipedrive, Zoho, Salesforce) with EU hosting is mature, cheap and the process is standard.
• Enterprise — Path B: Scores should use proprietary CRM, product-usage and intent data and feed own routing logic, which a generic product rarely covers.
• Developer — Path C: Gradient-boosted models (LightGBM, scikit-learn) run on HW1 and give full control over features and reason codes.

Path A — Buy (recommended for most SMBs)

• Built into the CRM:
• Microsoft Dynamics 365 Sales predictive lead and opportunity scoring, with up to 10 models per org and an AUC gate before publishing [1].
• HubSpot lead scoring: rule-based in Professional, AI-built in Marketing Hub Enterprise [2].
• Salesforce Sales Cloud Einstein Lead Scoring, now part of the Agentforce/Einstein bundles.
• Odoo CRM predictive lead scoring (Belgian vendor, EU) [3], Zoho CRM (Zia), Pipedrive.
• Mautic points for rule-based scoring [7].
• Intent and ABM platforms: 6sense [4], Demandbase, Bombora (intent data co-op) [5], and Dealfront (EU; Dealfront Group GmbH, which merged Leadfeeder and Echobot, "built & hosted in EU") for company-level visitor identification and intent [8].
• Choose A when:
• The CRM already offers predictive scoring for your licence tier.
• Leads are mostly inbound.
• Fewer than a few thousand outcomes per year are closed.
• Check where data is hosted, whether you can use custom fields, whether reason codes are shown, and whether you can export the score for reporting.

Path B — Build on cloud APIs

• AutoML / tabular ML in EU regions:
• Azure Machine Learning AutoML (classification).
• Google BigQuery ML or Vertex AI tabular (EU multi-region).
• Amazon SageMaker Canvas/Autopilot (Frankfurt).
• Databricks or Snowflake ML, where the warehouse already holds the CRM data.
• LLM enrichment and text features: An EU-hosted model API (Mistral La Plateforme, Azure OpenAI EU Data Zone, Claude or Gemini via EU regions on Bedrock/Vertex) with structured outputs. It classifies form text, maps companies to the ICP and writes reason summaries.
• Integration:
• Reverse-ETL or a CRM API writes the score back to a custom field.
• n8n, Make or Power Automate handles routing.
• Keep the training pipeline in the warehouse (dbt), so that sales forecasting (PRED-002) and churn (PRED-003) can reuse the same feature tables.
• Choose B when:
• You have several data sources, such as product usage, ERP order history or offline events.
• The CRM model is a black box you cannot tune.
• You want one scoring layer across several CRMs.

Path C — Self-host (open source)

• Models:
• LightGBM, XGBoost or CatBoost with scikit-learn pipelines.
• Optuna for tuning and SHAP for reason codes.
• TabPFN for small datasets. Its code is Apache 2.0, but recent weights (2.5–3.5) are licensed for non-commercial use only, so check commercial terms with Prior Labs. Only the TabPFN-2 weights carry an Apache-based licence [6].
• MLOps: MLflow (registry), Evidently (drift reports), Feast (optional feature store), and Airflow or Dagster for scheduling.
• CRM/MA stack: Odoo Community CRM [3], Mautic marketing automation [7], SuiteCRM or Twenty. Website analytics without third-party cookies via Matomo.
• LLM enrichment on-prem: A small open-weight model (e.g. Mistral Small, Qwen3 8B–32B) via vLLM or Ollama for text classification. A sentence-embedding model (e.g. multilingual-e5 or bge-m3) for ICP similarity.
• Hardware:
• HW1 is enough. Gradient boosting trains on a CPU in minutes, even with hundreds of thousands of leads.
• TabPFN runs on a CPU only for small datasets. A GPU with about 8–16 GB helps for larger ones [6].
• Add HW2 only if a local LLM also handles enrichment.
• Ops effort: Low to moderate: a weekly retraining job, monitoring and CRM-extract maintenance.

Path D — Custom training

Training your own model is Paths B and C. Lead scoring is always trained on your own outcomes. "Custom" here means going beyond that: - Fine-tuning a small LLM or encoder (BERT-class) to classify your form and email text in your industry's language. - Training account-level sequence models on event streams. - Uplift models that predict who responds because of sales contact, not just who converts anyway (see REC-003, causal-inference).

These are worth it only with thousands of outcomes per year and an analytics team.

State of the art (2026)

• Accuracy: Gradient-boosted trees remain the workhorse. Tabular foundation models have reached the point where they compete without tuning: TabPFN-3.5 handles up to 1 M rows and 20 k features [6]. For lead scoring, though, label quality and leakage matter more than the choice of algorithm.
• Built-in, low-data CRM models: The minimum data needed has dropped to dozens of outcomes [1][2], which makes predictive scoring realistic for SMBs.
• Signals:
• The trend is towards account-level intent and buying-group engagement (6sense, Demandbase) and away from individual contact scores [4].
• Intent data comes from publisher co-ops (Bombora reports 5 M company domains and 25,300+ topics) [5].
• Ad-tech consent rules limit how this data can be used in the EU.
• LLM-native workflows: LLMs now extract features from unstructured text, write the explanation and hand the lead to AI SDR agents (AGENT-007) or qualification chatbots (CONV-005). These then write new engagement events back, which is a feedback loop to watch.
• Limits and failure modes:
• Label leakage: Fields that are filled only after qualification inflate offline AUC.
• Self-fulfilling prophecy: Only high-scored leads get called, so only they convert. Keep a random holdout.
• Cold start: New markets or products have no history.
• Drift: Campaigns, pricing and ICP changes shift the data.
• Bot traffic: Bots inflate engagement scores.
• Consent gaps: Cookie refusal leaves gaps in behavioural data.

Shared resources & combinations

• Same tabular-ml stack and feature store (CRM + events):
• PRED-002 (sales forecasting and win prediction, the next funnel stage).
• PRED-003 (churn) and PRED-004 (CLV): one warehouse feature layer, one MLflow registry and one retraining scheduler on HW1 serve all of them.
• REC-004 (segmentation) and REC-010 (B2B cross-/upsell).
• Upstream data:
• AGENT-011 (CRM enrichment and hygiene). Lead scoring is only as good as the deduplicated, enriched records.
• LANG-017 (trigger events and market intelligence) as features.
• web-scraping for firmographics.
• Downstream action:
• AGENT-007 (AI SDR outreach to top-scored accounts).
• CONV-005 (chatbot qualification adds engagement events).
• REC-006 (nurture emails and send-time for low scores).
• REC-003 (next-best-action).
• Shared LLM: The EU API or self-hosted small LLM and embedding model of other use cases can handle text features.

Effort, cost & KPIs

• Path A: Days to a few weeks. The cost is mainly licence tiers: AI scoring is usually only in the top CRM/MA editions [1][2]. Intent platforms are enterprise-priced annual contracts.
• Path B: 1–3 months, including data pipeline and CRM write-back. Compute cost is negligible. The main cost is data engineering.
• Path C: 2–3 months plus around 0.1–0.2 FTE for ongoing operations (estimate, not sourced). The hardware is an existing server.
• KPIs:
• Model: AUC/PR-AUC, lift in the top decile, calibration.
• Funnel: MQL→SQL and lead→opportunity conversion by score band, against a random or unscored holdout.
• Speed and efficiency: speed-to-lead for A-grade leads, rep time per closed deal.
• Data quality and adoption: share of leads scored with complete data, rep override rate.

People & change

Start with management and the head of sales (process owner) to agree on the goal: prioritise, not replace, salespeople. Involve end users (sales reps) from the beginning, because their CRM discipline decides the data quality and their trust decides adoption. IT handles CRM integration and accounts; the data owner of the CRM clears field definitions. Involve the DPO early: scoring is profiling of natural persons (sole traders, contact persons), so check legal basis, transparency in the privacy notice, an AVV with the vendor and whether a DPIA is needed (GUIDE-005).

Works council: depends. A pure lead model does not monitor employees, but CRM activity data and "rep performance" dashboards are suited to monitor behaviour, so §87(1) no. 6 BetrVG often applies; §90(1) no. 3 requires early information. Agree explicitly that scores are not used for individual performance evaluation (GUIDE-004).

Support AI literacy (Art. 4 AI Act) with a short session on what a score means and where it fails (GUIDE-006). Pilot with one volunteer team, show reason codes, and compare with their own gut ranking. Typical concerns: "the machine ignores my experience" (reps can override, and overrides are logged as feedback) and bias against small customers (review score bands regularly). Humans decide on every contact. See GUIDE-008.

Risks, regulation & compliance

• EU AI Act: Minimal risk in the typical B2B and marketing deployment. Lead prioritisation is not an Annex III use case, but AI literacy (Art. 4) applies. Art. 50 transparency applies only if a chatbot or AI-generated outreach talks to the lead (CONV-005, AGENT-007).
• Boundary: If a "lead score" decides whether a natural person can access credit or life/health insurance, it becomes creditworthiness or risk assessment. That is high-risk under Annex III point 5(b)/(c) [9] and belongs in PRED-009/PRED-010.
• Lead scoring in HR recruiting (scoring candidates) is high-risk under Annex III point 4 and is not covered here.
• GDPR:
• depends because pure account-level scoring of companies (legal entities) is not personal data. Scoring identifiable contacts, which is the common case, is profiling. The usual legal basis is legitimate interest (Art. 6(1)(f)), with a documented balancing test.
• The data subject can object at any time to profiling for direct marketing, and processing must then stop (Art. 21(2)–(3)) [10]. Build a suppression flag into the pipeline.
• Art. 22 restricts decisions "based solely on automated processing" that have legal or similarly significant effects [11]. B2B prioritisation usually stays below that threshold. Automatically rejecting consumer applications may not.
• Follow the WP29/EDPB profiling guidelines [12]. Do a DPIA for large-scale B2C profiling, use data minimisation, and do not use special categories of data.
• ePrivacy / Germany:
• Behavioural tracking that reads or stores data on the user's device needs consent under §25 TDDDG [13], so engagement scores will have gaps for visitors who refuse consent.
• Outbound email or phone contact triggered by the score must comply with §7 UWG (opt-in for marketing emails to consumers, and in practice to most B2B contacts too).
• Third-party intent and visitor-identification data: check the vendor's legal basis, whether the data is company-level only, and where it is hosted.
• Fairness and bias: Exclude proxies for protected attributes (name-based gender, nationality) from B2C models.
• Mitigations: Random holdout, reason codes, human override, objection/suppression handling, and an audit log of model versions.

Getting started

1. Define the outcome ("converted to opportunity within 90 days") and the stages from MQL to SQL together with sales. Clean up lead status and disqualification reasons in the CRM.
2. Check the data: at least 40 + 40 or 25 + 25 closed outcomes for built-in models [1][2]. Check dedup rate, field completeness and consent status.
3. Switch on the built-in predictive scoring in the CRM you already have, or build a LightGBM baseline on a CRM export. Compare it with your current point rules on a time-based holdout.
4. Publish scores with reason codes. Set routing and SLAs per grade, and keep 5–10 % of leads unscored or randomly routed to measure lift.
5. Review lift and drift monthly. Add product-usage, intent or enrichment features only once the baseline is trusted.

Sources

1. Microsoft Learn — Configure predictive lead scoring (Dynamics 365 Sales) — https://learn.microsoft.com/en-us/dynamics365/sales/configure-predictive-lead-scoring
2. HubSpot Knowledge Base — Understand the lead scoring tool — https://knowledge.hubspot.com/scoring/understand-the-lead-scoring-tool
3. Odoo documentation — Predictive lead scoring — https://www.odoo.com/documentation/18.0/applications/sales/crm/track_leads/lead_scoring.html
4. 6sense — Predictive analytics (vendor-reported 6QA comparison) — https://6sense.com/platform/predictive-analytics/
5. Bombora — Company Surge intent data and Data Co-op — https://www.bombora.com/
6. Prior Labs — TabPFN repository (versions, limits, licences) — https://github.com/PriorLabs/TabPFN
7. Mautic documentation — Points — https://docs.mautic.org/en/5.x/points/points.html
8. Dealfront (Leadfeeder/Echobot) — https://www.leadfeeder.com/dealfront/
9. European Commission AI Act Service Desk — Annex III — https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3
10. GDPR Art. 21 — Right to object — https://gdpr-info.eu/art-21-gdpr/
11. GDPR Art. 22 — Automated individual decision-making, including profiling — https://gdpr-info.eu/art-22-gdpr/
12. EDPB — Guidelines on automated decision-making and profiling (WP251, endorsed 2018) — https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/automated-decision-making-and-profiling_en
13. §25 TDDDG (German Telecommunications Digital Services Data Protection Act) — https://www.gesetze-im-internet.de/ttdsg/__25.html
14. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj