Next Best Action Marketing: There Is Nobody to Recommend To

Next best action marketing uses a decision engine to choose, for each customer and moment, the most valuable thing to do next. The term comes from banks and insurers, where the engine recommends an action to a person who performs it. A direct-to-consumer brand has nobody in that seat, and that changes what the engine must be.

Relvino is a decision engine for ecommerce that acts on its own decisions, so read this as informed but interested. Every vendor statement below is quoted from that vendor’s own page, or, where a page would not render, from a search-result snippet of it (noted in the evidence); the argument about the empty seat is ours, and it applies to any store whether or not it ever considers Relvino.

What next best action is, in the vendors’ words

The results for this term are almost entirely enterprise software. The vocabulary is consistent across them.

• Pega: the aim is to “Automate real-time decision-making with AI” so that “Recommended actions are delivered consistently, no matter where the customer interaction occurs.” The inputs are “real-time data, customer profiles, and historical data” and signals “like intent, preferences, and behavior” to assess customer needs. The customer examples on the page are Wells Fargo, Bupa and National Australia Bank.
• Salesforce: Einstein Next Best Action “is a solution that uses flows, strategies, and the Recommendation object to recommend actions to users.” “Recommendations are offers or actions to recommend to users using Einstein Next Best Action” and “Strategies determine which recommendations to display to users, based on your data and business processes.”
• Tiger Analytics on Databricks: the solution “evaluates hundreds of variables including affinity scores, recent activity, territory priorities, compliance rules, and channel preferences” and “recommends the single most effective action for each HCP at any moment,” where an HCP is a healthcare provider and the people receiving the recommendation are sales representatives, medical science liaisons and marketing teams.
• Braze, under the newer name AI decisioning: “an AI-powered decision system that uses reinforcement learning (RL) to personalize experiences to optimize a specific goal, such as revenue, retention, or engagement.” It chooses across “messages, channels, offers, and timing.”

Two of the four are explicit that the recommendation goes to an employee, not the customer: Salesforce says “users” and Tiger Analytics says sales representatives and medical science liaisons. Pega’s own page says it will “seamlessly execute intentional interactions across all customer touchpoints,” and Braze’s AI decisioning names no human recipient at all. Salesforce’s own worked example is the clearest. When a customer abandons their cart, “an AI-enabled recommendation pops up on the salesperson’s dashboard” with options like sending custom offers, cart reminders or product information. The engine decides; a person acts.

The seat the term was built around

Next best action grew up in industries with a human in the loop by law or by economics. A bank has a branch teller, a call-center agent and a relationship manager. An insurer has a claims handler. A pharmaceutical company has a field representative visiting a physician. In each, the customer interaction passes through an employee, and the engine’s job is to put the best option in front of that employee at the moment of contact. Pega’s phrase, delivered “no matter where the customer interaction occurs,” describes many touchpoints and one recommendation surfaced at each; Pega’s own page says that can mean the customer directly, with an employee only at some of them.

That design has a consequence the term hides. The engine chooses among actions a person authored. Salesforce says so plainly: recommendations “are standard Salesforce records, similar to accounts and contacts,” and a strategy applies the business logic that decides which to show. Braze says the same for its agents: “Teams choose the objective, define the action set” and “set guardrails on things like frequency and offer depth. AI decisioning then decides who receives which experience and whether they should be contacted at all…” The library of possible actions is human work; the engine ranks within it.

A DTC brand has nobody to recommend to

A Shopify store with a lean team has no teller, no retention desk and no field rep. Between the decision and the customer there is a send button, and the send button is software. That empties the seat the whole vocabulary was built around, and three things follow.

• The recommendation is the action. There is no dashboard for it to pop up on. If the engine decides that this shopper should get a reminder, on SMS, tonight, with no discount, then the engine has to send it. An engine that stops at a recommendation has recommended it to nobody.
• The action library has to be generated, not authored. A bank can maintain a catalog of a few hundred offers because a product team owns it. A store with ten thousand SKUs, a shifting catalog and shoppers at every point in their own history cannot hand-author the action set; the set is a product, an offer level, a channel and a moment, and its size is the product of all four.
• Restraint is an action. When a person receives a recommendation, doing nothing is the default and costs nothing. When the engine is the sender, the decision not to message a shopper has to be made explicitly, per shopper, and it is often the right one. The AI email agents guide argues that a marketing agent is judged by the sends it declines; next best action, run without a person, has to decline them itself.

So “next best action” in ecommerce is a misnomer twice over. It is not next, because the engine is not queuing a suggestion for later; and it is not an action until something sends it. The honest name for the thing a store needs is a decision that executes.

What changes in the engine

Take the enterprise design, remove the person, and the engine has to grow three parts it did not previously need.

• Execution. The decision has to reach the shopper on the channel it chose. Relvino decides offer, timing, channel and creative for each shopper and executes across email, SMS/MMS and pop-ups, with no hand-off. The agentic marketing guide places this against the incumbents’ agent features, most of which still stop at a draft or a suggestion.
• Guardrails in place of a reviewer. In the enterprise version, the employee is the last check: an agent can decline a recommendation that feels wrong. Without that seat, the brand sets the limits once (margin floor, quiet hours, frequency, consent, channels, voice), and the engine decides inside them. Braze’s description of guardrails on “frequency and offer depth” is the same idea; the difference is whether a person still sits after them.
• A prior for the first decision. An engine that learns only from the brand’s own history starts every new store, and every new shopper, from nothing. Relvino’s decisions draw on a Large Retail Model trained across 10K+ retailers and 7M+ data points, on 1.78M real shoppers, so the first decision for a new store is informed by its category before its own results arrive. Each decision takes under 80 milliseconds, and the outcome of each feeds the next.

Where the enterprise framing still applies

To be fair to the term, two situations keep the seat occupied. A brand with a real customer-service or clienteling team (luxury, high-consideration goods, B2B accounts) does have a person between some decisions and some customers, and a recommendation surfaced to that person is the right shape for those interactions. And a brand that wants a human review of every offer before it ships is choosing the enterprise design deliberately; the cost is that the review queue becomes the bottleneck the engine was bought to remove. For a lean Shopify team running email, SMS and pop-ups, neither applies, and the engine that only recommends is an engine that needs a hire.

The two designs, side by side

In the list below, the recommends-to-a-person column is assembled from the Pega, Salesforce, Tiger Analytics and Braze statements quoted above (Salesforce’s in part from a search snippet, noted in the evidence); quoted phrases are theirs and the rest is our reading. The decides-and-executes column describes Relvino.

• Who receives the output · Next best action that recommends to a person: “users” (Salesforce), sales representatives (Tiger Analytics), agents at touchpoints (Pega) · A decision that executes: The shopper, on the chosen channel
• The action set · Next best action that recommends to a person: Authored: “standard Salesforce records”; “Teams choose the objective, define the action set” (Braze) · A decision that executes: Generated per shopper from product, offer level, channel and moment
• Last check before the customer · Next best action that recommends to a person: An employee, where the vendor names one · A decision that executes: Guardrails the brand set once
• The do-nothing option · Next best action that recommends to a person: Free: the person ignores the recommendation · A decision that executes: An explicit per-shopper decision to decline
• Where it learns from · Next best action that recommends to a person: The enterprise’s own data and models · A decision that executes: A Large Retail Model across 10K+ retailers, then the store’s own results
• Latency that matters · Next best action that recommends to a person: Time to surface a recommendation at the touchpoint · A decision that executes: Under 80 milliseconds per decision, then the send itself
• Built for · Next best action that recommends to a person: Banks, insurers, telcos, pharma field teams (the vendors’ own examples) · A decision that executes: Direct-to-consumer stores on Shopify
• Example · Next best action that recommends to a person: Pega, Einstein Next Best Action, Braze AI decisioning, by their own descriptions · A decision that executes: Relvino

For a brand that has read this far and runs Klaviyo, Braze or Salesforce today, the practical question is whether its current tool decides or suggests. The Klaviyo vs Salesforce and Klaviyo vs Braze comparisons read each vendor’s AI page on exactly that point; the Braze alternatives guide sorts the wider field the same way. The engine that executes can be proven the only honest way, a 14-day pilot beside the current setup, and the Klaviyo vs Relvino page lays that out with pricing. Two adjacent guides cover the decisions the term folds together: lifecycle marketing, which is what the stage programs were trying to do, and send time optimization, the one decision most platforms already automate, and the wrong clock they automate it on.

Frequently asked questions

What is the concept of next best action?

A decision engine looks at everything known about one customer at one moment and chooses the single most valuable thing to do next: an offer, a message, a channel, a wait, or nothing. Pega describes it as automating real-time decision-making with AI so that recommended actions are delivered consistently wherever the customer interaction occurs. The concept shows up across banking, insurance and healthcare, where the recommendation is surfaced to an employee who then acts on it.

What is the next best action on Salesforce?

Einstein Next Best Action is Salesforce’s implementation. A recommendation is stored as a Salesforce record, like an account or a contact, and a strategy decides which ones a person sees on a record page. That is a different mechanism from Pega’s and Tiger Analytics’ recommendation feeds, but the same shape: a system surfaces a pick, and a person decides whether to act on it.

What is NBA next best action?

NBA is the abbreviation the enterprise vendors use for next best action; the same idea is also sold as next best offer and, at Braze, AI decisioning. All describe a system that ranks the possible actions for one customer and picks one. The distinction that matters is what happens to the pick: in the enterprise products it is shown to a user, and in an autonomous system for ecommerce it is sent to the shopper.

What is a next best action engine?

The software that makes the pick. It scores the available actions against real-time behavior, the customer profile and history, and a set of business rules, then returns one. Salesforce and Braze rank within a set of actions a person on the team put there. Relvino skips that step: it builds the set itself for each shopper and each catalog, chooses one action or none, and sends it, deciding in under 80 milliseconds per shopper.

How long does it take to migrate from a next best action setup on Klaviyo to Relvino?

About 30 minutes for the technical cutover: connect the Shopify store, point the sending domain, connect SMS, and shopper data ingests automatically. There is no action library to rebuild and no recommendation queue to staff; the brand sets guardrails once and the engine decides and executes per shopper inside them. The result is proven in a 14-day pilot run beside the current setup, scored on revenue per shopper.