← Research library

Lead scoring with agents: from static rules to reasoned judgments

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Agentic marketing is moving into the parts of the funnel where judgment used to be encoded as rules. Lead scoring is one of the first places that shift shows up, because it sits between marketing and sales and both sides feel it when it goes wrong.

Traditional scoring is simple and transparent, but it struggles with nuance. Agents can read the same signals and make a reasoned judgment with an explanation. Done well, this improves prioritization. Done poorly, it produces an opaque score that nobody trusts.

Where point-based scoring breaks

A point-based model assigns values to attributes and behaviors. A hypothetical model might give 10 points for a director title, five for a pricing page visit, and minus 20 for a student email address. Each rule is easy to read, and the total is easy to audit.

The trouble is context. A pricing page visit from a current customer near renewal means something different from the same visit by a new prospect. A director title at a 30-person company carries different weight than one at a large enterprise. Each exception gets its own rule, and the rules start to interact.

After a few years, the model becomes a ledger of patches. Nobody remembers why webinar attendance is worth seven points, and nobody wants to remove the rule in case something breaks. The model is still transparent in theory, but no one can explain it in practice.

What agents add

Agents change the unit of scoring from a sum of points to a judgment about an account or a person. That judgment can draw on three things rules handle poorly.

  • Context: an agent can weigh combinations of signals, such as recent hiring in a relevant team, engagement with technical content, and an existing relationship with a partner.
  • Explanations: instead of "score: 78," an agent can say "likely in active evaluation: three people from the account viewed integration documentation this week, and the company recently posted roles for this function."
  • Research: before scoring, an agent can gather public information about an account that might take a person 15 minutes to find.

The explanation is the most valuable of the three. It turns the score into something a sales rep can check against their own knowledge, and it gives reviewers a way to catch errors. A wrong number looks the same as a right one. A wrong reason is usually visible.

Design principles

The design choices below decide whether an agent's judgment becomes useful or becomes another black box. We treat them as requirements rather than options.

  • Keep a rules baseline: run the agent alongside your existing model and compare. If its judgments do not lead to better outcomes, the added complexity is not justified.
  • Require explanations: every judgment should name the specific signals that drove it. Sales teams will not act on scores they cannot understand.
  • Use categories, not false precision: a judgment of high, medium, or low priority, with reasons, is often more honest than a score from 1 to 100.
  • Guard against bias: review which signals drive judgments, and explicitly exclude ones that act as proxies for characteristics you should not use.
  • Close the loop: track what happens to scored leads, and feed that evidence back into the agent's instructions.

Bias deserves more than a line. An agent can learn that leads from certain regions, schools, or company types convert better, and some of those patterns may track protected characteristics or create unfair outcomes. Write the excluded signals into the agent's instructions, check explanations for them during review, and confirm with counsel what you may and may not use.

The same care applies to consent. An agent should score only on data the person agreed to share, which we cover in consent-aware personalization.

A worked example

Consider a hypothetical B2B software team that receives 400 new leads a month. Its rules model sends the top 100 by score to sales development. The team wants to know whether an agent would choose a better 100.

The agent reads each lead's CRM record, recent web activity, and public company information. It assigns a priority category and writes a two-sentence rationale. It does not change routing. Sales still receives the rules-based list, while the agent's choices are logged next to it.

After a full sales cycle, the team compares the two lists. Suppose 70 leads appear on both. The interesting cases are the 30 the agent picked that the rules missed, and the 30 the rules picked that the agent ranked low.

Reading those 60 rationales, along with what happened to each lead, tells the team more than any aggregate number. It shows where the agent sees something real and where it is confidently wrong.

A practical rollout

The rollout we recommend moves the agent from observer to contributor in four steps, with evidence required at each one.

  1. Run the agent in shadow mode, scoring leads without affecting routing.
  2. Have sales reviewers rate a sample of explanations for usefulness and accuracy.
  3. Compare outcomes against the rules-based model over a full sales cycle.
  4. If results are better, use the agent's judgment for prioritization while humans keep control of routing for top accounts.

The connection to your CRM matters throughout. An agent that writes scores into shared fields can overwrite data other systems depend on. Give it its own fields, read-only access where possible, and a clear record of every write, as described in connecting agents to your CRM.

What to measure

Lead scoring has an outcome that arrives late, so you need early signals while you wait for conversions. We suggest tracking four.

  • Explanation accuracy: the share of sampled rationales where every cited signal is true.
  • Rep usefulness: how often sales reviewers rate a rationale as helpful for their first conversation.
  • Disagreement rate: how often the agent and the rules model disagree, and in which direction.
  • Downstream conversion: meetings, opportunities, and closed deals by priority category, compared with the baseline.

Explanation accuracy should come first. If the agent cites a webinar the lead never attended, nothing else it says will be trusted, and the doubt spreads to marketing as a whole. We discuss how that trust is built and lost in building trust with sales.

What we are studying next

The open question for us is how judgment-based scoring holds up as conditions shift. A rules model fails visibly when the market changes. An agent may keep producing confident rationales built on signals that no longer mean what they used to.

We are studying how often teams should rerun the baseline comparison. We also want to know which signs suggest that an agent's reasoning has drifted away from what actually predicts a sale.

How to apply this: start by asking an agent to explain, not score. Have it write a two-sentence rationale for each new high-intent lead, and ask your sales team whether those rationales help. Their answer will tell you whether scoring is the next step.

Related research

Frameworks

Writing briefs agents can actually follow

Briefs written for people rely on context agents lack. A few structural changes, starting with a strict proof section, make briefs work for both.

5 min read

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.