Predictive Marketing — Anticipating "What You Want Next"— The practice of getting ahead of expectations with data, and where that tips into feeling creepy

01What Prediction Is — Not "Guessing Right," but "Placing a Bet"

First, let's get the words exact. The "prediction" in predictive marketing is not fortune-telling the future. It looks at how similar people, in similar situations, moved in the past — and from that accumulation, it bets that "this person is likely to move this way next." It is not right-or-wrong; it is a weighting of probabilities.

Say a physician keeps viewing materials in a particular disease area. Many physicians who moved the same way in the past went on to ask for detailed clinical data — so the bet is that the next thing to offer is that data. It is not 100%. It will sometimes miss. Prediction, precisely stated, is not certainty but "prioritizing the next move."

The accuracy of this "bet" is set by the volume and quality of data. But the more accurate it gets, the more a separate problem appears: the person being predicted notices they have become "the object of a bet." That is where pleasant and unpleasant part ways. The next section looks at the mechanism.

02Expectation and Data — Read Through the Expectation-Disconfirmation Model

Why is "getting ahead" delightful at one moment and creepy at another? Explaining this by feel leads to design errors. A classic of marketing research, Oliver's expectation-disconfirmation model (= the framework holding that satisfaction is determined by the gap between prior expectation and actual experience, 1980), explains it cleanly.

The core of the model is simple. A person's satisfaction is set not by "the experience itself" but by the difference between "prior expectation" and "the actual." Exceed the expectation and you get satisfaction (= positive disconfirmation); fall below it and you get dissatisfaction. Predictive marketing is a technology that touches this "expectation" directly. So mishandled, the source of satisfaction inverts into the source of dissatisfaction.

In other words, the true nature of the creepiness is the moment the person realizes "they hold information I never meant to reveal." As Kahneman described in Thinking, Fast and Slow (2011), people sense "I'm being watched" with fast intuition (System 1). The body goes on guard before the reasoning does. Predictive design must not misread this intuitive reaction.

03Implementation in CRM — The Idea of "Next Best Action"

The central concept when translating prediction into CRM is next best action (= the single best move to take toward each customer right now). Rather than blasting the same notice to everyone at once, you look at each person's state and decide, case by case, "what is the best thing to do for this person, right now." Prediction supplies the raw material for that.

The skeleton of the implementation splits roughly into three layers.

• Data layer — Gather past contact history, browsing, and responses. What materials, when, how much viewed. If this is thin, prediction cannot stand.
• Prediction layer — From the gathered data, produce "the likely next action" and "areas of high interest" as probabilities. The part the model handles.
• Action layer — Taking the prediction, decide what to actually do. Send a material, have a rep make contact, or deliberately do nothing. Being able to choose this "do nothing" is a condition of good implementation.

The point of next best action is shifting the mindset from "send because you can" to "send because it's best." Once you can predict, you are tempted to raise the number of contacts. But the best thing for the other person is often "leave them alone for now." Use the power of prediction not for the volume of pushing but for the precision of restraint. This is where the pharmaceutical character is tested.

04Application to MR Activity — Prediction as "Support," Not "Replacement"

Where prediction bites directly in the pharmaceutical field is in the activity of the MR (= the sales representative who conveys drug information to healthcare professionals). Which physician, when, which information to deliver. This used to rest on the MR's experience and instinct. Prediction reinforces that instinct with numbers. But it does not replace it.

What prediction offers goes only as far as "candidates." The final decision on what to convey rests with a person, in light of the approved indications and efficacy and the context of the physician and patient in front of them. Even if prediction ranks "information likely to work" at the top, you must not choose it if it exceeds the approved scope. Prediction is a tool for prioritization, not an indulgence for regulation. This line is the same stance as the generative-content piece (Vol. 5).

05The Creepiness of Over-Prediction — The Point Where "Convenient" Flips to "Surveillance"

The more accurate prediction gets, the more, ironically, the risk of creepiness rises too. Take this lightly and you lose accumulated trust in a single stroke. Three typical ways over-prediction turns unpleasant:

What the three share is that we are making light of the person's awareness that they are "being predicted." What divides pleasant from unpleasant is not accuracy itself but two things: "are we respecting the range of the person's expectations?" and "can we explain why this suggestion?" Over-prediction happens when you cross both of those. To draw the line, you need the next two things — the law, and explainability.

06Drawing the Line Against the Personal Data Protection Law

Pushed to its essence, predictive marketing is the act of "gathering an individual's data and inferring that person's behavior." So you have to observe the Act on the Protection of Personal Information (= the law governing the protection of personal information, setting the rules for acquiring, using, and providing personal data to third parties) not by feel but at the level of its provisions. Here are the points, narrowed to a practical yardstick.

Medicine especially is a field where interest in a disease can invite inference of "medical history," which is special-care-required personal information. A person keeps viewing materials for a particular disease — that alone can let you infer their own or a close relative's medical history. So in predictive design, the practical discipline is to draw the line of "how far to gather, and what may be inferred" on the foundation of the law, and one notch more cautiously than the law. Not "gather because you can, infer because you can."

07Accuracy and Explainability — "Being Able to State the Reason" Over "Hitting Often"

When evaluating a prediction model, one tends to look only at "how often it hits." But in predictive marketing, and in medicine especially, being able to explain "why this prediction" (= explainability) is weighed as heavily as accuracy. Getting ahead without being able to state the reason leaves eeriness even when it's right.

There is often a trade-off here. The more complex the model, the easier accuracy rises, but the harder it becomes to explain how it reached its conclusion. Conversely, a simple, visible mechanism can fall short on accuracy. In pharmaceutical CRM, you must not tip this scale all the way to "accuracy first." For both physicians and patients, being able to explain, in a coherent chain, "why did this information arrive?" is a condition of trust.

In practice, you split the two. Raise the accuracy of the prediction. But at the point of delivery, translate the reason for the suggestion into a form a person can explain. "You viewed materials in this area last week, so we've brought the related latest data" — if you can explain it this plainly, the creepiness drops sharply. Accuracy you cannot explain, we don't use in pharma. Set that as the principle.

08Effects and Pitfalls

The effect of predictive marketing is clear. You can concentrate the limited time of MRs on the people with high interest. You cut off-target contacts and lower the other person's burden. Designed well, both sender and receiver gain. But the pitfalls are just as clear.

• Bias toward short-term response — Immediate numbers like "opened" and "clicked" are easy to capture. But whether they are eroding long-term trust is hard to see. Good response and damaged trust can coexist.
• Self-fulfilling prediction — Predict "this physician is highly interested in A," send only A, and the chances to touch B or C vanish. Prediction narrows reality, and narrowed reality reinforces the prediction again. The danger of fixing the field of view.
• Deferring regulation — When prediction ranks "information likely to work" at the top, checks on the approved scope and personal data protection easily fall behind. The allure of accuracy shoves aside the caution of regulation.

What these share is the structure of being dragged by "what's measurable" and losing sight of "what's important but hard to measure." Trust, breadth of view, legal compliance — none show up easily in short-term numbers. So in measuring the effect of prediction, alongside the response figures, always re-ask "does this getting-ahead respect the range of the person's expectations?" Because the numbers being good and the thing being right are two different matters.

09Connections to Other Chapters on This Site

This piece deepens when read together with the following chapters.

• AI Marketing Vol. 1 — Marketing Redefined — The whole map of advertising, CRM, content, and brand. Prediction sits at the core of CRM.
• AI Marketing Vol. 5 — AI-Generated Content Strategy — If prediction decides "what to offer," generation is "how to make it." The stance of a person carrying the final decision is shared.
• Hibi Kore Kojitsu (Daily Journal) — Essays treating a person's expectations and anxieties as a human-to-human affair. Reading the "creepiness" of prediction from the side of feeling.

Predictive marketing is the technology of reading ahead to the "next thing you want" and offering it. Done well it becomes a "thoughtful" experience; missed, an "annoying" one; and seen through too far, a "creepy" one. The dividing line lies not in the height of accuracy but in whether you respect the range of the person's expectations, and whether you can explain why this suggestion. As Oliver's expectation-disconfirmation model teaches, satisfaction is set not by the experience itself but by the gap against expectation. Precisely because prediction is a technology that touches that expectation directly, it demands care in handling.

To use it in pharma, two foundations are non-negotiable. One is the personal data protection law — especially in a field where disease interest lets one infer medical history, which is special-care information, "gather because you can" will not do. The other is explainability — prioritize "being able to state the reason" over "hitting often." Design prediction not as a tool for raising the volume of contact but as a tool for raising the precision of restraint. Use it not to watch the other person but to lighten their burden. That is the condition on which getting ahead turns into trust. Next time we move to whether AI-made "personas" can speak for medicine — the ethics of virtual influencers.