Thought Leaders

The Smartest AI Decision Might Be Not Sending the Message

Derek Slager
Add Unite.AI to your preferred sources on Google
Key PointsOptimizing for clicks rewards sending more messages even when restraint would better…
View more
  • Optimizing for clicks rewards sending more messages even when restraint would better protect the customer relationship.
  • A well-written promotion sent while a refund is unresolved is a decision failure that better content cannot fix.
  • Human review should focus on ambiguous context and high-consequence actions rather than routine policy cases.

Generative AI has made it much cheaper to produce a personalized message. It hasn’t made it easier to decide whether that message should be sent.

That distinction matters as AI takes on more of the decisions behind customer interactions. A system can produce a relevant offer, predict that someone will click and still make the wrong decision. The customer may be waiting on a refund. They may have an unresolved complaint. They may have already bought the product under a different email address.

Better writing won’t fix any of those problems. The system needs a better understanding of the customer, an objective that reflects what the business actually values and clear limits on what it can do.

What Are You Asking the System to Optimize?

Sends, opens, and clicks are easy to count. That makes them useful operating metrics, but it also makes them tempting substitutes for the thing a business actually wants: a customer who sees value in the relationship and keeps coming back.

An AI system optimized for total clicks, with few constraints on contact frequency, may find that sending more messages improves its score. Each additional send can look reasonable in isolation. Across campaigns and channels, the same customer can end up hearing from the brand far more often than anyone intended.

The cost shows up outside the metric being optimized. In Validity’s 2026 holiday consumer survey, one-third of respondents said holiday email volume alone was enough to make them unsubscribe. That doesn’t tell us AI caused the overload. It tells us that increasing activity can come at the expense of the relationship.

An unsubscribe is visible. The frustration that preceded it often isn’t. If the system gets credit for a click today but no penalty for making a customer less likely to return next month, it has little reason to account for that tradeoff.

Restraint Is a Decision Too

Think about a customer who returned an order last week and is still waiting on a refund. A marketing system sees a recent purchaser with a high likelihood of responding and queues a promotional email. From the customer’s perspective, the brand is asking for more money before returning what it owes.

The return may live in the order system, the refund status in a service tool and the purchase in a marketing profile. Those records may not even be connected to the same person. The model can be working exactly as designed and still be making a decision from an incomplete account of the customer.

Connecting those records changes what the system can know. It doesn’t automatically change what the system will do. The decision process also needs to account for an unresolved refund and allow holding back a promotion as a valid outcome. Otherwise, you’ve given the system more information without changing the behavior it is rewarded for.

This is why customer context and objectives have to be designed together. Context explains what is happening. The objective and decision rules determine what should happen next. A return alone doesn’t mean all communication should stop. The customer may need a refund update. The distinction between a useful service message and another promotion matters.

A poorly written email is a content problem. A well-written email sent while the customer is waiting for their money is a decision problem. You can’t solve the second by getting better at the first.

A Signup Is Not the Same as Loyalty

Loyalty programs offer another version of the same problem. If success means more enrollments, a system may learn to recruit people with a discount. Enrollment rises. Whether those customers return without another discount is a different question.

The useful questions take longer to answer. Are members coming back? Is their spending profitable after the cost of the incentives? Does the program change their behavior, or mostly reward purchases they would have made anyway?

Answering those questions requires connecting loyalty activity with purchase and service history over time. It also requires a comparison that helps separate the program’s effect from behavior that was already likely. A larger membership count cannot tell you that on its own.

Human Oversight Has to Change the Decision

People are responsible for deciding what the system should optimize. Retention, margin and customer effort involve tradeoffs that leadership has to own. A model can help evaluate those tradeoffs. It can’t take responsibility for choosing them.

The same is true of the limits on action. A useful starting question is: how hard would this be to undo? Holding a promotional message for review has different consequences from changing a price, issuing a credit or canceling a reservation. The consequences, the customer’s permissions and the reliability of the available context should determine what can run automatically and what needs approval.

An open refund doesn’t need to become a manual review task every time. If the policy is clear, a rule can hold the promotion while service communication continues. Human review belongs where the context is ambiguous, the rules conflict or the consequences exceed the system’s authority. If the information needed to make that distinction is missing, the system needs a defined fallback rather than permission to guess.

Reviewers also need more than a confidence score. They need to see which customer records informed the recommendation, how current those records are, what the system is trying to achieve and which constraints apply. A confident recommendation based on stale or mismatched records is still a bad recommendation.

Overrides should become evidence. If people repeatedly stop the same kind of promotion, record why. The pattern may reveal missing service data, an identity problem or an objective that rewards the wrong behavior. Fixing that pattern is more valuable than asking a person to catch the same mistake indefinitely.

Measure Whether Holding Back Helps

Teams can start with one decision rather than trying to redesign every customer interaction at once. Take a promotional send, identify the context it requires and define the conditions under which it should be held. Keep necessary service messages outside that suppression rule.

Then test whether the change helps. For eligible promotional messages, compare the existing approach with a policy that holds back under defined conditions. Track repeat purchases, retention, complaints and margin over a period that fits the buying cycle. Sends and clicks still help explain what happened, but they shouldn’t be the final verdict.

The test may show that fewer promotions protect the relationship. It may show that a particular restriction is too broad. Either result is useful. Silence shouldn’t get a free pass any more than sending should. Both are decisions whose consequences can be measured.

None of this removes the engineering work. Customer records have to resolve to the right person. Recent events have to reach the decision process in time to matter. Permissions have to constrain the action, and teams need a record of what happened and why. A better objective cannot compensate for missing context, just as better context cannot rescue the wrong objective.

Build for the Relationship

The question for leaders is straightforward: what does your system count as a good decision? If the answer is a click, it may be doing its job while making the customer experience worse.

Trusted customer context gives people and AI a more dependable basis for deciding what to do. Clear objectives and controls determine how that understanding gets used. The result should be a system that can explain why a message is useful now, recognize when it isn’t and learn from what happens next.

Sometimes the smartest decision an AI system can make is to send nothing at all.

Derek Slager is co-founder and co-CEO of Amperity, where he leads the company's AI-first transformation across both product and the way the company operates. He co-founded Amperity to give marketers and analysts customer data they could trust, and built the patented identity resolution and real-time profile architecture behind Amperity's trusted customer context. Earlier, he was on the founding team at Appature and held engineering leadership roles in large-scale distributed systems and security