Lead qualification with AI: rules first, models second
Plain rules sort most leads; AI handles the messy remainder. How to qualify fast, explainably and without losing good leads.
The short answer
Most leads that arrive through a website can be sorted in seconds by rules anyone in sales could write on a whiteboard: which country, what company size, whether a budget was mentioned, whether they asked for something you actually sell. Rules are instant, explainable and never hallucinate. Where they stop is free text: the paragraph a prospect wrote describing their situation, which says more about intent and urgency than any form field. That is where AI earns its place. Build the rules first, add the model for the reading, route uncertainty to a person, and measure against customers won rather than leads scored.
Rules and models, side by side
| Question | Best answered by | Why |
|---|---|---|
| Are they in a market we serve? | Rules | A fact from the form or the email domain |
| Company size, sector, role | Rules, with enrichment | Structured data, lookups |
| Did they state a budget or timeline? | Rules | Present or absent |
| What do they actually need, and how urgently? | AI reading the message | Free text; nuance; intent |
| Is this spam, a vendor pitch or a student project? | AI, with a rule for the obvious | Patterns in language |
| What should the first reply say? | AI drafting, person sending | Context-dependent writing |
| Which salesperson should take it? | Rules | Territory and load |
Building it
- Write the rules down with the sales team: the facts that make a lead a yes, a no or a maybe. Implement them as plain logic.
- Add the reading step: the model summarises the free text into intent, urgency, fit and any red flags, with a confidence level and the sentence it based each judgement on.
- Route by confidence and stakes. Clear yes with high confidence: assigned and a draft reply prepared. Clear no: logged, not deleted. Everything else: a person, same day.
- Draft, never send. The first reply is prepared with context; a salesperson reads, edits and sends.
- Log every decision, rule or model, with the reason, so any lead’s path can be explained.
- Review weekly: sample the rejects, read the maybes, compare qualified to won. Adjust rules and prompts.
What good looks like after three months
The sales team spends its mornings on conversations instead of an inbox. Hot leads get a personal, informed reply within the hour because the draft was ready. Obvious non-leads never reach a person, and the weekly sample confirms they were indeed non-leads. The maybes, which are where the interesting customers often hide, get a human look the same day. And the number that matters, customers won from website leads, is up, because response is faster and attention is better placed.
What this means for you
Lead qualification is a sorting problem with a reading problem inside it. Solve the sorting with rules you can explain, the reading with a model that shows its reasons, and the uncertainty with a person. Never discard silently, always draft rather than send, and measure customers, not scores. Done that way, the system makes the sales team faster and better informed without ever making a decision they could not defend.
Frequently asked questions
Why not let AI score every lead from the start?
Because a model's score is hard to explain and easy to trust too much, and because most of the signal is in structured facts that rules handle better: where they are, how big, what they asked for. Rules first give you a transparent baseline; AI on top adds the reading of free text that rules cannot do. The combination is faster and more defensible than either alone.
What if the AI misreads a message and we lose a good lead?
Design so it cannot lose one silently. Anything the model is unsure about, and anything it would reject, goes to a person or to a daily review list. Sample the rejects weekly. The cost of a person spending ten minutes a day on edge cases is far below the cost of one lost customer.
How do we know the qualification is working?
Track qualified leads through to won deals and compare with the period before. If the system qualifies more but wins fewer, it is qualifying the wrong things. The only metric that matters is customers, and the review loop should adjust rules and prompts against that.