Do not begin by asking AI for the customers most likely to buy more. Define an expansion decision your operating team can execute: which account, for which product or capacity, within what window, through which owner, and with what evidence of value. A health score can help prioritize. The business case begins only when a signal becomes a useful action and the company measures incremental gross margin or revenue against a credible comparator.
The common mistake is to equate healthy usage, high engagement, or frequent contact with expansion intent. Those signals may instead reflect satisfaction, complexity, unresolved support, or a contract that is already saturated. A propensity model finds accounts that resemble prior expanders; it does not prove that marketing, sales, or customer-success outreach changed the result. Implementation must separate prediction, eligibility, treatment, and effect.
Define the Account Growth Decision Contract
- Unit — account, parent, subsidiary, contract, or business unit; do not aggregate incompatible levels.
- Outcome — incremental ARR, MRR, revenue, gross margin, or volume net of discount and delivery cost.
- Horizon — when the outcome can occur and how long each signal remains valid.
- Eligibility — current product, capacity, risk, pricing, channel, territory, and contract constraints.
- Evidence — usage, adoption, service, relationship, intent, organizational change, and unmet need.
- Action — educate, diagnose, recommend, route, wait, suppress, or require human review.
- Capacity — the team that receives the account, its SLA, portfolio limit, and return reason.
- Comparator — holdout, stepped rollout, or a documented prior policy.
- Owner — who repairs the data, approves the action, serves the account, and signs off on economics.
The contract prevents marketing from sending one expansion campaign to every healthy customer and sales from receiving a context-free ranking. It also makes no action a valid decision. If no eligible offer, delivery capacity, or verified reason for contact exists, the system should wait.
Health, propensity, and a decision are not the same
HubSpot documents company- and contact-level health scores built from properties, events, and capped groups, including support, relationship, and NPS signals. Salesforce documents an account-ranking capability based on the likelihood of opportunity creation. These are useful capabilities, but they answer different questions: what conditions surround this account, and how similar is it to accounts that created an opportunity.
- Health score — summarizes current conditions and should remain explainable by component.
- Propensity — estimates a future event from historical patterns.
- Eligibility — verifies that an action is commercially, operationally, and contractually valid.
- Priority — combines potential value, urgency, confidence, and available capacity.
- Uplift — estimates who changes because of the intervention, not who would buy anyway.
- Decision — selects an action or no action and records the reason.
Do not compress all six concepts into one number. An account can have high potential and poor health; another can be healthy but economically saturated; a third can show a clear need while a critical support issue remains open. Expose components and keep hard rules—payment status, permission, territory conflict, renewal risk, or product incompatibility—outside probabilistic optimization.
Example: seat and product expansion in a midmarket SaaS portfolio
Consider a software company with 600 active accounts that wants to expand an automation module. Login frequency alone is not enough. The loop combines current feature coverage, remaining manual work, integration maturity, open tickets, relationship sentiment, eligible users, renewal timing, account economics, conversation history, and communication permissions. AI summarizes the evidence and suggests a next conversation; rules confirm eligibility; the account owner accepts or rejects the recommendation.
Actions compete: educational content, a customer-success diagnostic, a sales demonstration, a technical review, waiting for a support resolution, or no outreach. Marketing does not claim success from an MQL. The operating scorecard tracks valid interventions, owner acceptance, sourced and influenced opportunities, incremental gross margin, time from signal to action, and harms such as repetitive or poorly timed outreach.
Run an eight-week implementation
- Week 1 — select one offer, one account portfolio, and one economic outcome.
- Week 2 — map account identity, signals, permissions, owners, and execution capacity.
- Week 3 — write the contract, eligibility policy, and exclusions.
- Week 4 — reconstruct the baseline and manually review 30 to 50 accounts.
- Week 5 — run rules or a model in shadow mode with reasons and confidence.
- Week 6 — release two low-risk actions to part of the eligible portfolio.
- Week 7 — monitor execution, data quality, experience, and results with a comparator.
- Week 8 — decide whether to fix, expand, automate, or stop.
Start with transparent rules when account volume is modest. Use native CRM capabilities when they expose score components and support workflow, feedback, and audit. Consider a custom model when volume, reliable labels, several actions with different costs, and a real optimization need justify it. Cost grows with account reconciliation, product and support integrations, near-real-time updates, offer-level personalization, experimentation, explainability, and regional operating differences.
Do not confuse likely response with incremental growth
Revenue-uplift research distinguishes predicting who will respond from estimating who will respond because of a treatment. That distinction matters: targeting only the highest-propensity accounts can spend capacity on customers who would have expanded without intervention. Begin with an approved, limited holdout or stepped rollout, preserve the previous policy, and compare incremental gross margin rather than opportunity rate alone.
When randomization is not feasible, disclose the limitation. Before-and-after, owner, and territory comparisons are exposed to selection, seasonality, and portfolio differences. Last-touch attribution cannot isolate the combined effects of product, customer success, sales, and marketing. The initial goal is a better decision and a better learning system, not causal claims the design cannot support.
Measure the whole system
- Data — reconciled accounts, valid events, latency, and missing signals.
- Decision — eligibility, coverage, confidence, reason, and no-action rate.
- Execution — SLA, capacity, owner acceptance, returns, and completed follow-up.
- Revenue — accepted opportunities, expansion, gross margin, cycle time, and incremental value.
- Experience — complaints, opt-outs, support conflicts, and repeated outreach.
- Learning — policy changes, false positives, removed signals, and retired actions.
Assign responsibility before building another dashboard
Marketing owns the thesis, message, and treatment; customer success explains health, adoption, and timing; sales defines qualification and ownership; product provides telemetry and usage limits; data maintains contracts and account identity; analytics designs evaluation; technology integrates and monitors; finance validates margin; legal and privacy review use and basis; the executive sponsor resolves capacity and incentives. NIST's AI RMF reinforces continuous, context-aware governance—especially when a score can influence customer pressure or employee compensation.
Bring one real account to the discussion
Use the minimum data loop at https://makinai.co/insights/en/cdp-ai-marketing-minimum-customer-data-loop-revenue, next-best action at https://makinai.co/insights/en/ai-next-best-action-crm-personalization, customer-conversation signals at https://makinai.co/insights/en/ai-sales-support-conversations-marketing-decisions, and the experiment system at https://makinai.co/insights/en/ai-marketing-capacity-valid-experiment-operating-system. MAKINAI connects CRM, data, and commerce at https://makinai.co/services/en/crm-ecommerce-commerce-transformation. Bring one account, the offer, current signals, and intended action; we can bound the smallest expansion loop worth testing.
Scores learn from history and can reproduce uneven coverage, owner behavior, and old commercial policy. A green health score does not prove need, timing, or incremental effect. Native features still require configuration and monitoring. Eight weeks is a scope frame, not a revenue promise, and requirements for privacy and automated decisioning vary by jurisdiction and industry.