Chatbot Lead Scoring: A Model Sales Will Use

What the score is for

A lead score has exactly one job: to order a queue. It is not a measure of how good a customer someone will be, or how interested they seemed, or how pleasant the conversation was. It answers one question — who should be called first — and any complexity that does not improve that ordering is decoration.

This matters because scoring models grow. Somebody adds a signal, then a sub-score, then a decay curve, and within six months there is a number nobody can explain and therefore nobody trusts. The version that survives is the one a salesperson can reason about out loud: 'this is an eighty because they asked about integrations, said this quarter, and match our size.'

So the design constraint is not accuracy. It is explicability. A slightly worse model that sales acts on beats a better one they ignore, every time.

The signals that actually predict

Conversation data contains far more signal than a form, but only a handful of it predicts anything. These are the ones worth scoring, roughly in order of usefulness.

1. 1Stated timingThe strongest single predictor available and the cheapest to collect, because people answer it honestly. 'This quarter' and 'just researching' should sit at opposite ends of the scale and should dominate the score.
2. 2Specificity of the questionA visitor asking whether you support a named system they already run is further along than one asking what the product does. Specific, technical and integration questions are strong signals; general capability questions are weak ones.
3. 3Fit against your actual customersIndustry, size and the channels they use, weighted by what your real closed deals look like rather than by your ideal-customer slide. This is the signal that most often disagrees with sales intuition, and it is usually right.
4. 4Depth of engagementNumber of exchanges, whether they returned, whether they read specific pages before opening the conversation. Useful, but weaker than the three above — and easy to over-weight because it is the easiest to measure.
5. 5Explicit buying languageAsking about contracts, onboarding time, references, or who else uses this. These questions come from someone who has already decided the product might work.
6. 6Negative signals, which matter as muchA student, a job applicant, a competitor, a country you do not serve, a budget an order of magnitude off. These should subtract decisively rather than merely fail to add.

Weighting without pretending to be precise

Do not start with a hundred-point scale and fractional weights. Start with three bands, because three bands is what a queue can actually use.

• Call now — timing is near, fit is good, and there is a specific question. These should be a small minority of conversations; if they are not, your criteria are too loose.
• Call this week — two of the three above, or strong signals with something missing.
• Nurture — genuine interest, distant timing. These should be the largest group, and calling them is the most common way sales time gets wasted.
• Do not call — negative signal present. Explicitly separating this from 'nurture' is what keeps the nurture list usable.

Only once those bands are stable and sales agrees with them is there any point in a numeric score, and the number should then be derived from the bands rather than the other way round. Most organisations never need to go further than four bands.

Validating the model against reality

A scoring model that is never checked against outcomes is a set of opinions with a number attached. Validation is not complicated, but it requires sales to record something.

1. 1Record a verdict per leadOne click after the first contact: worth calling, or not. Without this single field, nothing below is possible, and this is where most scoring projects quietly die.
2. 2Compare score bands against verdicts monthlyYou are looking for one thing: do high-scoring leads get accepted more often than low-scoring ones? If the bands do not separate, the model is decorative.
3. 3Read the disagreements, not the averagesThe high-scoring leads sales rejected, and the low-scoring ones that closed. Ten of each, read as transcripts, will tell you more about which signals to change than any aggregate.
4. 4Adjust one thing at a timeChange a single weight or threshold per cycle. Changing three at once means you learn nothing from the result.

Why scores decay, and what to do

Every scoring model degrades, and predictably. Knowing the three mechanisms lets you schedule the maintenance instead of discovering the problem when sales stops using it.

• The market moves. The questions that indicated a serious buyer last year are being asked by everyone this year, because the category matured.
• Your product changes. A new capability creates new high-intent questions the model has never seen, and they score as general enquiries.
• Traffic mix changes. A new campaign or channel brings a different population, and a model tuned on the old mix mis-ranks the new one.
• The team games it. If a score affects who gets which lead, people will learn to influence it. This is not dishonesty; it is a predictable response to an incentive, and the model has to be designed for it.

A quarterly review of twenty transcripts against their scores catches all four, and it is the only maintenance most models need.

Where scoring should not be used

• To decide whether to answer someone. Score the queue, never the service — a low-scoring person still deserves a good answer.
• To replace a conversation. If a score is uncertain, one clarifying question resolves it better than any model.
• For existing customers, who should be routed by account, not scored as new business.
• As a performance metric for salespeople, which converts the score into something to be manipulated rather than used.
• In place of speed. A perfectly ranked queue contacted four hours late loses to a roughly ranked one contacted in ten minutes.

That last point deserves weight. On the MIT Sloan Lead Response Management Study led by Dr. James Oldroyd, leads contacted within five minutes were far more likely to be qualified than those contacted after thirty — roughly twenty-one times more likely. Note the precise claim: the study measured the odds of qualification, not of a sale. Ranking helps, but it never buys back the delay.

How Vexvon supports scoring

Vexvon does not ship an opaque score. It produces the inputs a score is built from, which is the part that is actually hard. Each conversation writes extracted fields, a one-sentence summary and the transcript to the customer record, so timing, specificity and fit are available as data rather than as an impression.

A buyer-intent filter runs before a lead is created, which is the negative-signal step described above applied automatically: contacts extracted from conversations that are not genuine buying signals do not become leads, and the company's own numbers are excluded so internal tests never enter the queue. Duplicates are linked rather than counted separately, so one interested buyer is one record rather than three.

The queue itself is built in the CRM: leads carry a priority marker, a category, a stage in a nine-stage pipeline, a close reason and a contact-attempt count, with every change written to an activity log covering ten action types. Assignment runs in one of four modes — manual, round robin, by channel or by load — and reminders are checked every minute, which is what keeps the ordering from being undone by delay.

For validation, reporting exports conversations and leads to a file, and answer transparency records which knowledge produced which reply — so when you read the ten disagreements described above, you can see what the conversation actually contained rather than reconstructing it.

Frequently asked questions

1. 1What is chatbot lead scoring?It is the practice of ranking leads produced by conversations so sales knows who to contact first. Its only job is to order a queue, and any complexity that does not improve that ordering should be removed.
2. 2Which signals predict best?Stated timing first, then how specific the questions are, then fit against your actual closed customers. Engagement depth is real but weaker, and it is the one most often over-weighted because it is easiest to count.
3. 3How many points should the scale have?Start with three or four bands rather than a numeric scale. Most organisations never need more, and a band a salesperson can explain out loud beats a number nobody trusts.
4. 4What data and integrations are required?Extracted conversation fields, a customer record matched on phone or email, a sales verdict recorded after first contact, and a queue that can be ordered. The verdict is the one people skip and the one that makes validation possible.
5. 5How do you know the model works?High-scoring leads should be accepted by sales more often than low-scoring ones. If the bands do not separate on that measure, the model is decorative.
6. 6When should a human override the score?Whenever they have information the conversation did not contain — a known account, a referral, a market event. The score orders a queue; it does not have authority over it.

Start with the exclusions

If you are building this for the first time, do not begin with weights. Take last month's leads, mark the ones nobody should have called, and find what they had in common. That list is your first version of the model, it costs an afternoon, and it usually improves the queue more than the ranking work that follows it.