To implement AI next-best-action, do not begin by asking a model to write a unique message for every customer. Choose one recurring CRM decision and make five alternatives compete: send useful content, wait, suppress outreach, route to service, or present an approved offer. Keep permission, frequency, eligibility, inventory, margin, and sensitive cases in auditable rules. Scale only if the workflow improves incremental revenue or contribution without increasing opt-outs, complaints, unnecessary discounts, or operating load.
The useful problem is rarely a shortage of content. Email, paid media, mobile, and service teams often make separate decisions about the same person. A customer can receive an acquisition offer while waiting for support or get a coupon despite being likely to buy at full price. Next-best-action is a decision layer between signals and channels, not a copy generator inside every campaign.
Replace the campaign calendar with one observable decision
Choose a moment with enough volume, economic consequence, and real alternatives: after a first purchase, before likely replenishment, after a meaningful abandonment, or when usage declines. Frame the decision as: for this eligible customer at this moment, which permitted action has the highest expected value after cost and risk? If the answer is always send, the project is campaign automation, not decisioning.
The decision table your system must complete
- Context — event, journey stage, products, permitted channel, and information that is still current.
- Alternatives — send, wait, suppress, serve, or offer; each needs an owner and entry conditions.
- Value — expected revenue or contribution, channel cost, incentive, service load, and cannibalization risk.
- Constraints — permission, frequency, inventory, open service case, recent purchase, vulnerability, and applicable rules.
- Evidence — observed signal, model inference, approved content, decision record, and subsequent outcome.
This is an operating contract, not a maturity score. Rules can remove incompatible actions; a model can rank the remaining options; generative AI can adapt approved content for the selected action. Preserve the option to do nothing. Without it, the system learns only which message to send, not when silence protects customer value.
Use only data that can change the decision
A first pilot usually needs a stable-enough identity, purchase and return events, service status, product availability, contact history, permissions, and a reconcilable business outcome. Do not replicate the entire warehouse into a vendor environment. Identify which fields affect eligibility, ranking, or measurement and exclude the rest. Specify latency: yesterday's inventory might support a newsletter but not a recommendation that implies availability now.
Adobe documents centrally managed audiences built from profiles and rules, while its Privacy Service guidance stresses understanding data types, labels, and the identifiers used in privacy requests. These illustrate implementation concerns, not a requirement to buy a CDP. If the existing CRM can support the pilot, a selective integration may be faster and less expensive than full profile unification.
A practical U.S. example
Consider a subscription business after a failed renewal. The workflow checks payment state, open support issues, product usage, contact permission, and previous attempts. It may wait, send an approved service explanation, route an account issue to a person, or present a permitted recovery option. An unresolved service defect should outrank promotional messaging. Sensitive attributes should not be inferred for persuasion, and legal review should determine applicable notice, consent, and automated-decision obligations.
An eight-week implementation scope
- Weeks 1–2: choose one customer moment, map current decisions, establish baselines, and name owners.
- Weeks 3–4: build the decision table, exclusion rules, minimum data, approved actions, and test cases.
- Weeks 5–6: integrate in shadow mode, compare recommendations with current decisions, and review errors.
- Weeks 7–8: release to a small cohort, preserve a comparator, measure outcomes, and decide to keep, expand, or stop.
Eight weeks is a scope boundary, not a performance promise. Deliverables should include the data inventory, rules, ranking logic, action library, templates, integrations, logs, test cases, dashboard, runbook, team access, and rollback plan. A reviewer should be able to explain why each action was available, blocked, or selected.
Measure the decision, not personalization volume
Compare the new workflow with the current process and, when volume permits, retain an eligible holdout without the new intervention. Braze documents a global control group for comparing messaged and unmessaged users, along with practical limits; that feature is not automatically the right experiment for a bounded pilot. Assign groups before launch, keep them stable, and analyze everyone assigned rather than only openers or clickers.
- Primary outcome — incremental revenue or contribution per eligible customer, matched to the decision.
- Guardrails — opt-out, complaint, discount, return, service contact, and fatigue.
- Decision quality — correct suppression, stale data, unexplained choice, and integration failure.
- Operations — time to change a rule, cost per decision, exception queue, and human review hours.
Lifecycle marketing owns the journey and action library; data owns variables and evaluation; engineering owns integration and monitoring; service owns escalations; finance validates contribution; legal and privacy teams assess notice, permission, sharing, retention, and other applicable duties. NIST's generative-AI profile supports lifecycle measurement and monitoring but does not certify the system. Product documentation should be treated as capability evidence, not outcome evidence.
What drives cost and what you may not need to buy
Cost grows with customer moments, channels, markets, identities, real-time sources, action count, models, integrations, and support expectations. Compare three designs: better rules in the current CRM, a native decisioning feature, and a custom decision layer. Rules are transparent and inexpensive; native features reduce integration but deepen platform dependency; a custom layer increases control and operating cost. Do not make a CDP or CRM replacement a prerequisite unless the pilot proves the decision cannot be supported otherwise.
Bring one real decision to the first conversation
Start with the 90-day workflow plan at https://makinai.co/insights/en/where-start-ai-marketing-90-day-first-workflow, review data readiness at https://makinai.co/insights/en/assess-data-readiness-before-hiring-ai-company, and compare the win-back pilot at https://makinai.co/insights/en/ai-customer-reactivation-crm-without-discounts. MAKINAI's CRM consulting connects journey design, data, and implementation: https://makinai.co/services/en/crm-ecommerce-commerce-transformation. Bring one customer moment, the current actions, and anonymized exceptions. We can discuss whether the next investment should be a rule, integration, experiment, or model.
The sources support specific capabilities and cautions. They do not show that next-best-action will improve sales for a particular company. The decision table, pilot, and metrics are MAKINAI editorial recommendations that require operational and legal validation.