Predictive SQL conversion score

A conversion score exists to order the inbound pile so a person calls the most likely buyers first. It does not create a sales-qualified opportunity, and it does not book a meeting. When Monday's queue is larger than the SDR team's day, the useful output is a work order: these leads get a human now, these wait, these do not get a live call until someone has a reason.

Treat a high score as a short SLA, not as a pipeline stage. The SDR still has to reach the person, test pain and timing, and decide whether an opportunity belongs in the CRM. Teams that auto-write SQL when a score crosses a threshold are using the model as a rubber stamp. That fills the forecast with records nobody has qualified.

The score sits downstream of capture. A richer inbound record, from an AI inbound chat qualifier or from a form that asks buying questions, gives the model something real to rank. A name, an email, and a content download do not. Website visitor de-anonymization can put an account on the list before a form is submitted. That still is not qualification. It is only a better-informed place in the queue.

If the score does not change who works what, and when, you have a dashboard, not a process.

Score each lead against whether similar leads became closed-won customers. That is the label. If you train on who submitted a form, who became an MQL, or who booked a first call, the model will get excellent at predicting marketing activity. SDRs will then spend the morning on people who fill forms, not people who buy.

The page title says SQL. Be careful with that word. If SQL is a human disposition, a model trained to predict "becomes SQL" is predicting your current SDR habits, not buyer conversion. Prefer closed-won, or opportunity created and later won, as the event the score is trying to estimate.

Fit and behavior both belong in the features. Firmographics, tech stack, and similarity to accounts you already win tell you whether this is your buyer. Recency, page path, and buying-signal trigger monitoring tell you whether they are in motion. A perfect-fit account that never returned is a research project. A hot visitor from an account you cannot sell is a distraction.

Watch the training set for a quieter failure. The model only sees outcomes for leads someone actually worked. Untouched leads look like non-converters, so the model learns your old routing habits and calls them truth. Hold out a small random sample that SDRs work regardless of score. Without that sample you cannot tell model quality from a self-fulfilling queue.

Retrain on a schedule. Product mix, ICP, and inbound sources shift. A score that was honest last quarter can quietly start ranking last year's buyer first.

Ask for calibration, not only ranking. If a band is labeled 70, then over a long window roughly seven in ten of those leads should reach the conversion event you trained on. A model that orders the queue well but is wildly overconfident will make you staff the phones for a conversion rate you do not have.

Routing is the implementation. High scores go to a live SDR first, with a short first-touch SLA. Mid scores go onto a sequenced cadence. Low scores go to nurture or to a human disqualification recommender review, not to an automatic closed-lost.

Write the rule in the CRM, not in a slide. Score bands need owners, SLAs, and a fallback when the primary SDR is out. Round-robin inside a band is fine. Round-robin across the whole pile, after you paid for a score, is how the ranking gets ignored.

Do not hide the reasons. Reps trust a queue when they can see why a lead sat at the top: similar to closed-won accounts, pricing page twice this week, title matches the buying role. If the score is a black box, they will cherry-pick the names they recognize and you are back to the old pile.

Picture a Monday inbound mix from a demo form, a webinar, and a content gate. The score puts a director of operations at a mid-market logistics company, demo request plus two pricing visits, at the top. It puts a student who gated an ebook from a personal email at the bottom. The SDR starts with the director. That is the job. If ops then stamps SQL because the score crossed a threshold, the director is still uncalled and already sitting in a pipeline stage they have not earned.

Staff first-touch against how many high-score leads you can actually call well, not against total form volume. When inbound spikes, the score keeps that first call from being spread thin across every form fill.

Every lead the score or the SDR declines needs a recorded reason. Without that log, two expensive things happen.

The queue starves long-tail ICP. New segments, smaller logos you have started to win, and unusual titles never accumulate enough worked-and-won history, so they score low. If low scores never get a human look, those accounts never get a chance to convert, and the model never learns they can. You lock in last year's ideal customer while the company is trying to expand it. ICP lookalike account discovery can widen the fit features. It cannot fix a routing policy that only feeds the model the accounts it already likes.

Humans also start silently skipping high-score junk: competitors, students, existing customers, bad data, agencies posing as buyers. If those skips are not logged, the model keeps sending the same junk to the top, and reps learn to ignore the ranking.

Run a thin, deliberate sample of the reject pile every week. A few low-score leads, worked as if they were high, tell you whether the floor is honest or whether you are burying a new ICP. Record the outcome the same way you record the rest of the funnel. That sample is how you keep the score from freezing your market.

When a high score is wrong, write the reason from a closed set: wrong persona, out of territory, already a customer, competitor, insufficient data, timing, other. Use free text only as a last resort. Those reasons are the next training set, and they are the coaching list for marketing's forms.

MadKudu, Salesforce Einstein, 6sense, and HubSpot all sit in the same job. They produce a number you can sort a queue by. Some of that number lives inside the CRM you already run. Some of it comes from a specialist scoring or intent product. None of that choice, by itself, makes a lead sales-qualified.

Do not pick among them as a ranking exercise. Do not assume native CRM scoring is automatically enough, or that a specialist tool is automatically smarter. Inspect three things instead: what event the score predicts (form fill, MQL, SQL, opportunity, or closed-won); whether it arrives in time to change first-touch routing; and whether a person can see the contributing factors.

If the product predicts likely MQL, you will optimize for marketing's current definition. If it predicts closed-won, the SDR queue starts to look like your revenue, which is the point.

After go-live, review two lists in the same meeting: high-score leads that did not convert, and low-score leads that did. The first list is where routing, data, or the model is overconfident. The second is where you are starving a buyer you actually win. Adjust features, labels, and SLAs from those lists, not from a vendor dashboard that reports score adoption.