title: How AI lead scoring actually works (and why most score-based routing fails) slug: how-ai-lead-scoring-actually-works date: "2026-04-22" category: AI author: founding-attorney dek: A practical, field-tested look at why generic LLM-based lead scoring rarely outperforms a well-instrumented intake checklist — and what to do instead. tags:
- AI
- Intake
- Operations
Most lead-scoring marketing demos look great in a slide deck and fall apart in production. The pitch is consistent: a model reads the intake notes, returns a score from 0 to 100, and the higher numbers route to senior attorneys for follow-up. After three quarters of deploying this pattern across personal-injury firms, the answer we keep arriving at is the same: scoring on a single number is the wrong abstraction.
What a good signal actually looks like
When a senior intake manager evaluates a new lead, they are not computing a single score. They are checking five or six dimensions in parallel:
- Incident type and venue match the firm's case profile
- Liability is plausible — there is at least one defendant with assets or insurance
- Damages are large enough to justify the firm's case-acquisition cost
- The claimant is reachable and has not already retained
- The statute of limitations clock is workable
- The lead source is one the firm trusts (existing client referral, paid channel with known conversion, etc.)
A good model represents these as a vector, not a scalar. An honest intake form captures them explicitly. Once you have the vector, routing becomes a deterministic decision: matters with strong damages but weak liability route to the senior attorney's queue; matters with strong liability but weak damages route to a paralegal for further qualification; matters with weak both auto-decline.
The single biggest mistake we see is letting a model score the lead and then treating its rationale as a black box. The model should always return its reasoning per dimension, and the routing logic should be inspectable as a table — not as a number with vibes.
Where scalar scoring breaks
Scoring breaks the moment the firm's case mix changes. A 0-100 model trained on Q4 2024 leads gets stale fast: medical-malpractice referrals have different signal than slip-and-fall. We have watched firms quietly under-route their best leads for two months because the model was anchored on stale weights.
You also lose the ability to debug. When the scalar says 42, the partner asks "why?" and the answer ("the model said so") is not actionable. When the vector says "damages: high; liability: medium; reachability: low; SOL: 11 months", the partner has a clear next step — call the lead, lock down liability witness statements, file before the clock closes.
What we ship in Intake AI
We use the LLM to extract dimensions from the unstructured note (the receptionist's call summary or the chat transcript). We do not use it to compute a final score. The routing rules sit in the firm's settings — written by the operations lead, reviewable by the partner, swappable per practice area.
This is the same pattern the Paralegal Agent follows for matter management: the model extracts structured fields; deterministic logic decides what happens next. The audit trail records both — every state change in the matter has a model rationale and a deterministic routing decision attached.
If you are evaluating a CRM that promises AI lead scoring, ask three questions:
- Does the model output reasoning per dimension, or just a number?
- Can a non-engineer read and edit the routing rules?
- Is every routing decision recorded in an audit log with the model's input visible?
If the answer to any of those is no, the system will work great until it stops working — and you will not be able to tell the difference until you lose a case.
The next step
If you want to see how this looks in practice, the fastest path is a 30-minute demo on your firm's intake data. Bring a CSV of last quarter's leads and we will run the extraction live; you will see the dimension vector for each lead and the routing decision the engine would have made.