Automated Creative Optimization — Infinite Generation and Automated A/B— Mass generation vs. brand consistency, and the limits of automating regulatory review

01What DCO (Dynamic Creative Optimization) Is

First, let's pin down the idea at the center of this field precisely. DCO (= Dynamic Creative Optimization) is a mechanism that, for a single ad slot, recombines parts — headline, image, description, button color — differently for each viewer and situation. Instead of deciding on "this one ad" in advance and showing it to everyone, you prepare the parts and assemble them on the spot before serving.

If the parts are 4 headlines, 3 images, and 2 buttons, the combinations come to 4×3×2, or 24. When generative AI mass-produces these parts, the combinations balloon easily into the thousands and tens of thousands. The aim of DCO is to shift delivery, out of that vast set, toward whatever performs well. Here, "how many you can make" and "how many you can test" become linked.

In pharma, though, this "recombination" cannot be used naïvely. Wording that touches efficacy, indications, or safety is not something you may freely swap in and out as parts. Bring DCO in as-is, and approved combinations end up mixed together with unapproved ones. That danger is addressed head-on in a later section. First, let's follow the basic flow — make, test, shift — through the technology in the next section.

02How the Multi-Armed Bandit Thinks

"Test which expression works, and shift delivery toward the good ones." The foundation of this mechanism is the multi-armed bandit (= a method that tests several options while adjusting how much of your betting goes to the winners). The name comes from a slot machine with multiple levers (= a play on "one-armed bandit," the bandit with many arms). Not knowing which lever pays off most, how should you pull to gain the most within a limited number of tries — that is the problem.

There are two conflicting desires here. Trying an option you haven't tried yet — "exploration" — and continuing to use the option that has been best so far — "exploitation." All exploration, and you fail to cash in even after finding a good option. All exploitation, and you cling to whatever happened to look good first, missing the true winner. How to tune this tug-of-war is the heart of the bandit. In a 2002 paper, Auer and colleagues formalized this tug-of-war mathematically as UCB (= Upper Confidence Bound, a mechanism that preferentially tries options with few trials so far).

DCO becomes easier to grasp when you see it as this multi-armed bandit applied to combinations of parts. Treat the thousands of combinations as "arms," and shift delivery toward the arms that perform well. Generative AI makes arms without limit, and the bandit selects the winners from among them — that is what "infinite generation × automated optimization" really is.

03Mass Production by Generative AI

Now to the side that makes the arms — mass production by generative AI. Once, crafting a single headline took time. Now, specify the angle, tone, and length, and dozens of candidates appear in minutes. Images too can be made in parallel, multiple drafts each for a patient-facing or physician-facing audience. The number of arms you can test has changed by an order of magnitude.

What makes this mass production dangerous in pharma is that AI optimizes for "plausibility," not "correctness." The more readable, persuasive, and convincing the expression, the less its errors stand out. Let's restate the three traps of mass production here.

The third is a trap that arises specifically when mass production and automated optimization are joined. Even if each individual part has passed review, meaning can change in the combined context. Place the phrase "highly effective" next to a certain comparison-data image, and though each is fine alone, together they can imply an exaggerated superiority. The more arms there are, the less a human can track all of this contextual drift.

04Brand Consistency (Vibe)

Another wall for mass production is brand consistency (= the coherence by which any expression feels like the same brand; in this series we call it Vibe). Make arms without limit and shift toward the ones that perform, and expression gets pulled toward "whatever draws the most response right here." Leave it alone, and tone, color, and forcefulness of claim scatter; short-term response may be good, but coherence as a brand erodes.

A pharmaceutical brand stands not on flashiness but on accuracy and composure. Expression that earns response with strong words may post good numbers in the moment, but it chips away at a long-term asset — the trust of healthcare professionals and patients. So automated optimization must be given, in advance, not only performance metrics but a frame that asks "does this expression stray from the brand's voice?" The range in which the bandit is allowed to explore is bounded first, from the brand's side.

Concretely, the permitted tones, the expressions to avoid, and sample tones are embedded structurally into the generation instructions. Apply the brand frame at the stage of making arms, and don't generate arms outside the frame at all. Rather than letting it create freely and then fixing it, mass-produce only inside the frame from the start. This ordering is the key to protecting the Vibe while using speed to the fullest.

05The Limits of Automating Regulatory Review

Here is the core of this installment. If arms swell into the thousands, you'll want to automate regulatory review too. Detecting prohibited terms, and range-checking efficacy expressions, can be largely handed to the machine. But pharmaceutical regulatory review has a part that cannot be fully automated. To see it precisely, let's confirm the immovable foundation — the Pharmaceuticals and Medical Devices Act (PMD Act).

Misidentifying an article number itself damages trust. Exaggerated advertising is Article 66, unapproved advertising is Article 68, and information provision is Article 68-2 — this is a foundation worth memorizing outright. Beneath it sit the Ministry of Health, Labour and Welfare's Sales Information Provision Activity Guideline (= HanteiG, the Guideline on Sales Information Provision Activities for Prescription Drugs; notice of the Director-General of the MHLW Pharmaceutical Safety and Environmental Health Bureau, 2018) and the Standards for Fair Advertising of Drugs and the Like (= the yardstick for advertising expression, issued as a notice by the Director of the Compliance and Narcotics Division, Pharmaceutical Safety and Environmental Health Bureau, MHLW). Even when AI does the making, the standards applied are the same.

The limits of automation converge on two points. First, whether something is exaggerated is decided by context. A prohibited-term list can block words like "best" or "safe," but it cannot fully catch expression that creates an exaggerated impression without using those words. Second, the meaning of a combination can only be read by a human. Even when each part passes, the implication that arises when they are placed side by side is hard for a machine to judge. So regulatory review must take a two-layer structure: the machine handles a first-pass filter, and the final go/no-go remains with a human. A design that runs fully automated all the way to publication does not hold up in pharma.

06The Scarcity of the Creator

"If AI makes things without limit, do we stop needing people?" To this question, the answer is a clear no. What overflows from mass production is "plausible expression." The more it overflows, the higher the value of the judgment that picks the one that is correct, on-brand, and does not cross regulation, and sets the direction. As the labor of making drops, the eye that selects and the sense of direction become scarce.

At least three jobs remain in pharmaceutical practice. Defining the brand's voice and putting it into words as a frame for generation. Reading the context of combinations and seeing through the regulatory danger. And rendering the judgment — irreducible to any metric — of whether to take short-term response or long-term trust. None of these are things AI can mass-produce. The creator shifts role, from the person who makes parts to the person who designs the frame, selects, and takes responsibility.

07Measuring Effect

Automated optimization is inseparable from measuring effect. Click-through rate, dwell time, information requests — data can be captured in fine detail. But measuring the effect of pharmaceutical content carries a caution absent from general web marketing. Strong short-term response can harm the long-term brand.

A slightly "inflated" arm can produce better response in the moment. But if it is ever deemed exaggerated, the trust built up collapses in an instant. A drug is a "trust good (= a good whose quality is hard to verify in advance, standing on trust)." So in measuring effect, you must always look, alongside the response numbers, at whether that expression is eroding the brand's trust. Build regulatory compliance and brand consistency into the very definition of the "performance" handed to the multi-armed bandit — a design that does not let it maximize response alone. How you set the metric decides the quality of the automation.

08Connections to Other Chapters on This Site

This installment gains depth read together with the following chapters.

• AI Marketing Vol. 1 — Marketing Redefined — The whole map of advertising, CRM, content, and brand. This installment is the frontier of its "advertising × content."
• AI Marketing Vol. 5 — AI-Generated Content Strategy — The human-in-the-loop design that builds review into the speed of mass production. The foundation for this installment's mass production and automation.
• Material Review series — The practical review that finally catches the generated output and its combinations. The "exit" of automation.

Automated creative optimization has made it possible, through two technologies — DCO and the multi-armed bandit — to "make without limit, test as you go, and shift toward the good ones." Generative AI mass-produces the arms; the bandit picks the winners. This speed is a large opportunity. But in pharma, two walls do not move — brand consistency (Vibe), and regulatory review.

Apply the frame first. At the stage of making arms, hand over the brand's voice and the approved information, and do not generate outside the frame. For regulatory review, let the machine handle a first-pass filter, and keep the judgment of context and combination with the human. The PMD Act does not move even when AI gets faster — Article 66 exaggeration, Article 68 unapproved, Article 68-2 information provision. The more thoroughly you hand the automatable parts to the machine, the higher the value of the judgment you cannot hand over — selecting, setting direction, taking responsibility. What becomes scarce in the age of infinite generation is not the hand that makes, but the eye that selects. Next time, we move to how marketing changes in a world after third-party cookies have been abolished.