Next-best-action, often shortened to NBA, is the idea that a CRM should tell the person working with a customer what to do next, instead of leaving them to scan a record and work it out. It appears as a card in the sidebar: “Send the case study,” “Schedule a renewal call,” “Offer onboarding help.” Done well, it feels like a thoughtful colleague. Done badly, it is a nagging notification that people learn to close without reading.

How suggestions get chosen

Behind the card there are usually three layers, and understanding them helps you control the output.

Eligibility rules. The simplest layer decides which actions are allowed at all. A customer in an open complaint should not receive a promotional offer. A contact who opted out of email should not be suggested an email. These rules are written by people, and they protect customers and the business.

Scoring or ranking. Among eligible actions, something ranks them. It may be a handwritten priority list, a predictive model that estimates how likely each action is to lead to a good outcome, or a mix. This layer is where the “AI” is.

Constraints and fatigue control. Good systems limit how often a customer is contacted, avoid repeating the same suggestion, and respect business priorities such as inventory or capacity.

If your vendor demo skips the first and third layers and talks only about the model, ask about them. They prevent most of the embarrassing outcomes.

Where it works and where it does not

NBA does best where there is a clear menu of actions, repeated situations and a measurable result. Examples include renewals, onboarding, service follow-ups and upsell moments in a defined journey. It struggles with rare, complicated, high-stakes interactions, where context lives in a rep’s head and a generic suggestion can feel tone-deaf.

It also needs real signals. If the CRM has little record of what customers did, a suggestion based only on company size and industry is just a segment rule with a nicer interface.

Design the action menu first

Before turning on any model, list the actions you want the system to recommend. Keep it modest at first: perhaps eight to twelve. For each action, write down:

• The situation it fits
• Who performs it (rep, success manager, automated message)
• What counts as success
• When it must not be suggested

This exercise often reveals that the team’s real playbook is thinner than assumed, and that is useful on its own. A model cannot recommend a good action that nobody has defined.

Give reps a way to talk back

A suggestion that cannot be dismissed with a reason becomes noise. Provide a small set of response options: done, not relevant, wrong timing, customer asked not to. Each one is feedback that improves rules and ranking.

Review that feedback regularly. If “not relevant” dominates for one action, the eligibility rule is too loose. If “wrong timing” dominates, tighten the conditions.

Measure what matters

The easy metric is acceptance rate, how often reps click the suggestion. It is a weak measure on its own, since reps may click because it is convenient. Better measures follow the outcome: did customers who received the action progress, renew or respond more often than similar customers who did not? Whenever possible, compare against a holdout group that receives no suggestion, so you can tell the difference between a good recommendation and a customer who would have acted anyway.

Respect the customer side

Personalized suggestions often rely on personal data. Know what you are allowed to use and for what purpose. Rules differ by region and by whether your customers are consumers or businesses, and consent and transparency obligations may apply. Have your privacy lead or a qualified professional review which data feeds recommendations, particularly anything sensitive.

Also consider how a suggestion feels from outside. An offer that appears a day after a customer complained is not clever. Keep a rule that pauses promotional actions when there is an open service issue.

A guide to the design pieces

Piece	Purpose	Owner
Action menu	Defines what can be suggested	Sales or success leadership
Eligibility rules	Prevents inappropriate suggestions	Operations with compliance input
Ranking logic	Orders eligible actions	Analytics or vendor model
Frequency limits	Avoids fatigue	Operations
Rep feedback options	Improves the system	Operations
Outcome tracking	Proves value	Analytics

A hypothetical example

A subscription business wants renewal conversations to start earlier. They define six actions: onboarding check-in, usage review, training offer, executive sponsor call, renewal quote and win-back call. Eligibility rules prevent the renewal quote when a support escalation is open, and the win-back call only appears for lapsed accounts.

For the first month, ranking is a simple priority list by days to renewal and usage trend. The reps dismiss many “training offer” cards as irrelevant. The team learns that the offer only makes sense for accounts with a low usage of a particular feature, narrows the rule, and complaints drop. Only afterward do they consider adding a predictive model to rank among eligible actions.

Start with rules, then layer in models

Many teams do better by launching on rules and a priority list, then adding model-based ranking when they have outcome data. A model on top of a thin menu and messy eligibility rules will optimize the wrong thing very efficiently.

Frequently Asked Questions

Is next-best-action only for sales? No. It is used in service, onboarding and retention too. The principle is the same: pick the most useful eligible action for the situation.

How many suggestions should a rep see at once? Few. One to three is plenty for most workflows. More than that encourages scanning and ignoring.

Can customers see these recommendations? Some systems drive customer-facing messages too. In that case be especially careful with consent, frequency and tone.

What if the recommendation is wrong? Make it easy to dismiss with a reason, and log that. Wrong suggestions are expected; the real failure is a system that never learns from them.

Do we need a data science team? Not to begin. A well-built action menu with clear rules gets you far. Data science helps later with ranking and measurement.

Next Step

Write the action menu for one customer journey, such as onboarding, with eligibility rules for each action. Turn on a rule-based version for a small group and collect dismissal reasons for a month before considering any model.