Mercer Health — Propensity models that re-prioritised the pipeline
A custom account-propensity model surfaced 137 forgotten accounts with high buying intent — 22 became opportunities and 9 closed inside the quarter.
Mercer Health's CRM held 18,000 healthcare-provider accounts, 11,400 of which hadn't been touched by sales in over 12 months. Sales leadership assumed the cold accounts were genuinely cold. The data told a different story: a subset of them was quietly showing every signal of an active buying cycle — hiring patterns, regulatory filings, technology-stack changes, RFP language — but nothing in the CRM scored them, so no AE saw them. The brief: prove there is gold in the database and surface it in a form sales will act on.
Why dormant pipeline is the largest untapped asset in B2B
Most enterprise CRMs contain 4–7x more inactive accounts than active ones. The standard response — quarterly 'reactivation' email campaigns — produces single-digit reply rates because every dormant account is treated identically. The accounts that are genuinely re-entering a buying cycle are buried under thousands that aren't.
Propensity modelling solves this by ranking dormant accounts on the probability of opening a new buying cycle in the next 90 days, using features the CRM already holds plus external signals (hiring, filings, web behaviour, partner moves). Done well, it routinely surfaces 1–3% of the dormant pool as genuinely high-propensity — small enough for sales to action, large enough to move the number.
Healthcare adds two complications: longer buying cycles (12–18 months) and regulatory volatility. A change in CMS reimbursement guidance, a new HIPAA enforcement memo, or a state-level licensing shift can unlock a wave of latent demand inside a week. Models that don't ingest regulatory signal miss the moment.
Dormant accounts vs active accounts in a typical enterprise CRM
Forrester CRM Audit 2025
Of dormant accounts re-enter active buying any given quarter
Bain B2B Benchmark
Typical buying cycle in regulated healthcare SaaS
KLAS Research
"Mercer was paying to maintain 18,000 records. Less than a third had ever been worked. Inside the unworked two-thirds was a quarter's worth of pipeline, hiding."
How we actually ran it.
Target definition and label engineering
Define exactly what the model is predicting.
Joint workshop with sales and RevOps. Settled on a binary label: 'opened a qualified opportunity within 90 days of model scoring'. Pulled three years of historical opportunities to build positive labels; sampled negatives from the dormant pool.
A clean, sales-ratified label and a training set of 31,000 account-month rows.
Feature engineering
Build features that combine internal CRM with external signal.
Internal: tenure, last-touch recency, prior opp count, prior loss reason. External: hiring on Clinical Informatics roles, CMS regulatory filings naming the provider, partner-network moves, tech-stack changes (Datanyze, BuiltWith), RFP/RFI language detected via news monitoring.
147 candidate features reduced to 38 via SHAP-based importance and stability testing.
Model build and validation
Pick the model class that sales will trust.
Compared logistic regression, gradient-boosted trees and a small neural baseline. Selected gradient-boosted trees for the precision/explainability balance. Validated on a forward-time holdout — the only valid test for a temporal prediction.
Forward-time AUC of 0.84, precision-at-top-200 of 0.46. Every score shipped with the top-3 contributing features.
Sales activation
Deliver scores in a form an AE will use.
Top-200 list refreshed weekly in Salesforce, with a 'why now' panel for each account naming the top features and a recommended opener. Built a 30-minute enablement session with three AEs before broad rollout.
First-week meeting acceptance: 33%. Compared to legacy outbound: 6%.
Closed-loop and governance
Keep the model honest as the market shifts.
Weekly retrain. Quarterly review of feature drift, fairness across provider segments, and outcome reporting to the CRO. Documented retirement criteria up front.
By week 13, the model had been retrained 9 times with no precision drop.
The chronology of the work.
- Day 0
Brief: prove there is pipeline hiding in the dormant 11,400.
- Day 21
Label and training set ratified by sales leadership.
- Day 49
First scored list in Salesforce. 137 accounts surfaced from the dormant pool.
- Day 60
Of those 137: 89 contacted, 38 meetings booked, 22 advanced to discovery.
- Day 75
First two closed-won deals from the surfaced cohort.
- Day 90
9 closed-won, $11.4M ACV.
- Day 110
Model expanded to second product line. Same architecture, retrained on its own labels.
The three places most programmes die.
Predicting 'will close' instead of 'will open an opportunity'
Close is too rare and too noisy. Predict the next stage transition you can actually validate inside the quarter — usually 'opportunity created'.
Beautiful AUC, useless ranking
Optimise for precision-at-top-k, not AUC. Sales can action 200 accounts a week, not 2,000. The metric must match the workflow.
Black-box scores
Ship every score with the top contributing features in plain English. An AE will not call an account on faith.
A weekly-refreshed propensity model surfaced 137 high-intent accounts from a dormant pool of 11,400 inside one quarter, producing 22 opportunities and 9 closed-won deals worth $11.4M. The CRO retired the quarterly 'reactivation' email campaign in favour of model-driven sales outreach, and Mercer began rolling the same architecture across its other product lines.
Dormant accounts re-prioritised in 90 days
Closed-won deals from the surfaced cohort
Net-new ACV in one quarter
The model didn't tell us anything we couldn't have found ourselves — if we'd had a hundred extra people. What it did was tell us where to look first.
Lessons we'll carry into the next programme.
The dormant pool is usually the largest under-priced asset on the GTM balance sheet.
Precision-at-top-k is the only metric that matters when the consumer is a sales team.
Regulatory signal is the strongest single feature in healthcare propensity models — and the one most teams omit.
Forward-time validation is non-negotiable. Random splits flatter every temporal model.
An explanation panel is half the model. Without it, adoption collapses by the third week.
The next opportunity in your pipeline is almost always already in your CRM. The question is whether your scoring can find it before your competitor does.
Want a programme like this one?
Tell us where you want pipeline to come from next quarter — we'll show you how the next 90 days could look.