How to Score and Prioritize Leads with AI

Verdict

Start with explicit qualification rules, then use AI to classify the evidence those rules need. Choose Apollo when reps already search and work leads there and need a shared, explainable ranking. Choose Clay when priority depends on custom classifications, several data providers, or your own scoring formula. Keep exclusions, fit, timing, and missing evidence separate. A high score should earn a place in a reviewed queue; it does not prove purchase intent or authorize outreach.

This guide ends with an ordered queue, reason codes, owner, and review date. It differs from target-account research, which builds the evidence brief, and CRM enrichment, which maintains the underlying fields. For platform selection, use the AI sales prospecting tools guide, Clay vs Apollo, and the Apollo and Clay profiles.

Decide what the queue is for

Pick one team, territory, product, and next action. Inbound demo requests, outbound account selection, and expansion opportunities should not share one undifferentiated leaderboard. An existing customer's support request may be urgent without being a new-sales lead.

The output is an operational ordering. If you want a probability of conversion, define the outcome, time horizon, sample, and calibration method separately. Neither a vendor label nor a points total establishes that probability.

Choose the scoring surface

Apollo's scoring documentation distinguishes AI auto-scores from custom criteria for people and companies. It groups matches into Excellent, Good, Fair, and Not a fit. Reps can inspect the contributing criteria. AI scoring is available on paid Basic, Professional, and Organization plans; Free lacks AI scoring. Free and Basic permit up to two custom scores, while Professional and Organization permit unlimited custom scores. Separate account fit from person fit before choosing which model should be primary.

Apollo custom scores can be drafted with its assistant or built manually. AI-assisted setup requires business context; operators still review filters and weights. Creating models requires the relevant permission, and setting an organization's primary score requires an admin. Numeric criterion impact runs from 0 to 20. Watch overlapping criteria: a title matching both “Product” and “Manager” can accumulate both weights. Changing a primary score changes the default your team uses, so review the distribution before applying it.

Clay's formula approach suits custom fields and explicit rules. A formula can return a numerical score, a grade, or a yes/no fit label. You supply the data, criteria, and destination CRM fields. Have AI propose the formula if useful, then inspect the generated logic against examples. A formula generated by AI is different from paying a model to reassess every row.

Build the workflow

1. Freeze the input and remove exclusions

Take a dated snapshot keyed by CRM record ID and company domain. Deduplicate before scoring; attach contacts to the right account. Preserve customers, open opportunities, active sequences, opt-outs, and ownership as separate control fields. Do not let an attractive firmographic match override them.

Store scoring_version, evaluated_at, and the input revision. A rep needs to know whether the displayed score came from today's evidence or a previous company profile.

2. Calculate fit with rules; classify only ambiguous fields

Use deterministic comparisons for geography, size bands, approved industries, and required attributes. Ask AI only for a bounded classification that cannot be obtained reliably from structured data. For example: classify a company's published product description into your approved business-model categories, return the supporting URL, or return unknown.

Keep the extracted fact separate from the point allocation. A model can propose “sells enterprise infrastructure software”; your reviewed rule decides how that category affects fit. Never ask a model to invent a score by reading a company name and guessing its budget.

3. Add timing with an expiry

Separate strong first-party events, such as an explicit demo request, from weaker external signals, such as hiring or a content interaction. Record event time, source, and an expiry policy. Avoid counting the same event repeatedly because it appears in a provider feed, CRM note, and research summary.

Make the decay rule visible. An illustrative policy might expire hiring signals after 30 days, while keeping an unanswered direct buyer request in an overdue queue. Those are example operating rules to adapt, not vendor defaults or proven optimal windows.

4. Route with a rule the rep can explain

The following example is an illustrative policy, not Apollo's native scale, a Clay template, or a tested conversion model:

if excluded: remove from new-outreach queue; preserve reason
else if required_evidence_missing: review
else if hard_fit_failed: reject
else if explicit_buyer_request: priority_1
else if fit_points >= 40 and fresh_relevant_signal: priority_2
else: nurture_or_research

Suppose two fictional accounts both earn 45 fit points. Account A has an unverified hiring rumor; Account B submitted a dated request for a supported use case. B belongs ahead of A. A remains in research until the signal is verified. Adding ten guessed points to A would conceal the missing evidence.

Record the winning reasons and exclusions, not just a total. Keep account and contact ranks distinguishable so ten contacts at one account do not consume the whole queue.

5. Inspect the distribution before assigning work

Check top-ranked, middle-ranked, rejected, and unknown records with a reviewer. Include known customers, stale employment, ambiguous domains, duplicated events, and companies just outside the ICP. A useful model must handle those boundary cases, not merely produce plausible top picks.

In Apollo search, apply the score filter to people or companies, sort by the score column, and open individual records to inspect contributing filters and signals. In Clay, keep explanatory fields beside the formula result and review the proposed export before updating another system. Our CRM enrichment guide covers field-level change review.

6. Assign ownership without automatically sending

Attach the next action, owner, reason, review deadline, and score version. Before sequence enrollment, recheck suppression and existing activity against current CRM state. A change in priority must not enroll the same person twice or steal an account from an active opportunity owner.

Use the automated prospecting workflow for the controlled handoff from a ranked queue into outreach.

Current costs and limits

Checked October 8, 2026. Pricing is a platform cost, not a promise of qualified leads.

Verify the Apollo selector and Clay selector for your billing term. Clay's Enterprise Action figures conflict across public pricing and documentation; obtain a written allowance instead of extrapolating these entry tiers. Apollo's fourth-plan naming also differs between its cards and FAQ. The comparison records those unresolved boundaries.

For an illustrative Apollo research budget, 100 leads with two AI Research output columns consume 200 research credits. That arithmetic does not say scoring itself costs 200 credits or that all 100 leads are usable. For Clay, calculate each enrichment, AI run, and export stage separately; do not multiply a formula-only row count by an assumed AI price.

What practitioner evidence changes

In an HN discussion from November 2025, Poomba described rising Clay costs while finding decision-maker emails, technology stacks, and hiring information for agencies. They later clarified they were new to outbound and did not use a CRM. AznHisoka argued that stitching alternatives together could consume the apparent savings, while also promoting their own alternatives blog. These are individual reports with narrow context and a disclosed promotional interest, not comparative accuracy evidence.

The useful question for a scoring pilot is therefore: does this additional field change which leads the team accepts enough to justify its acquisition and maintenance cost? Remove expensive fields that do not change the reviewed queue. Neither anecdote establishes current credit prices or proves that a ranking predicts sales.

Validate priority before scaling

Compare the proposed queue with a simple rules-only baseline on the same records. Review precision among the first rep-capacity-sized group, false exclusions, missing-evidence rate, duplicate accounts, reviewer disagreement, and total cost per accepted lead. Track those measures by territory and segment; an acceptable average can hide a poorly covered market.

When evaluating later outcomes, use only information available at scoring time. Do not leak future replies, opportunity creation, or closed-won status into historical features. Keep human overrides with reasons so a future model does not learn every rep preference as objective truth.

Roll out one queue at a time, keep the previous scoring version available, and suspend automated assignment if hard exclusions or ownership controls fail. Expand only after reviewers can explain why the top records deserve attention. For the wider stack beyond prioritization, see AI sales tools by workflow.

Sources & further reading

• Apollo Scores Overview and plan limits
• Apollo custom scoring and primary score controls
• Apollo score filtering and sorting
• Apollo pricing
• Apollo AI Research billing
• Clay lead scoring formulas
• Clay Actions and Data Credits
• Clay pricing
• Practitioner discussion of enrichment cost and integration tradeoffs