AI in Marketing: Science Should Serve the Art

Every decade or so, marketing reinvents itself and announces that this time, finally, it will be able to talk to each customer as an individual. Database marketing promised it in the nineties. CRM software promised it in the aughts. Programmatic advertising promised it in the tens. None of them quite got there. Now AI is making the same promise, and the question worth asking — as always — is what's different this time, and what isn't.

A recent conversation hosted by Amazon Web Services brought together Tom Burrell, a retention strategy consultant with over two decades of experience working with subscription and recurring-revenue businesses, and Chris Morris, a media strategist focused on the technical end of the advertising industry. The AWS branding is worth acknowledging: this is a platform conversation, produced by a company that sells cloud infrastructure to the brands these two consult with. That doesn't make the ideas wrong, but it does mean the framing leans toward implementation rather than critique. Keep that in mind as you read.

With that caveat on the table, both Burrell and Morris say things that ring true to anyone who has watched the gap between marketing's promises and marketing's delivery widen for the past thirty years.

What the model actually gets wrong

Burrell's central argument is about churn models, and it's more interesting than it sounds. A churn model, for those not fluent in the dialect, is a statistical tool that identifies which customers are likely to cancel or stop buying. You build it by studying customers who have already left, find the patterns they shared, and then apply those patterns to your current base to flag who might be next. Then you typically send them a discount to make them stay.

Burrell's complaint is that this approach is structurally broken for businesses with flexible, month-to-month relationships — streaming services, subscription apps, gyms, gambling platforms. His argument: by the time someone is showing the signals your churn model was trained to detect, the decision to leave has already been made in their head. The discount doesn't fix the underlying problem. It just costs you money.

"The people who take that offer are the people who would have stayed anyway," Burrell says. "And the control group — the people who are going to leave — they leave anyway."

What he proposes instead is a system built on what he calls embeddings. This is the same underlying technology that powers large language models like ChatGPT — a way of representing behavior not as a checklist of events but as a kind of map of relationships between signals. Where a traditional model might flag "customer hasn't logged in for 14 days," an embeddings-based system tracks the shape of the whole relationship: frequency trending down, sessions getting shorter, certain features going unused. It catches the slope before the cliff.

The gym analogy he uses is the clearest illustration. A rules-based system says: no visit in two weeks, send a re-engagement email. An embeddings system notices that someone who used to come three times a week has quietly drifted to twice, then once, without ever triggering a hard rule. You intervene while they still care — maybe with a new class recommendation, maybe with encouragement — rather than chasing someone who has already mentally cancelled.

This is genuinely new capability. The computational cost of tracking behavioral drift at individual level, in real time, across a large subscriber base would have been prohibitive even five years ago. Burrell says some subscription apps are already running what he calls "retention 3.0" systems like this. His description of the mechanics — cloud data infrastructure feeding behavioral signals into pre-built embedding models, with AI agents then deciding which intervention to trigger across which channel — is plausible and consistent with how these systems actually work. I find this part of the conversation credible, for what that's worth.

What I'm less certain about is how widely it's being done versus how widely it's being discussed. There's a long tradition of practitioners in AWS-hosted conversations describing cutting-edge implementations that turn out to be rare, complex, and expensive to replicate. Burrell's advice to start small — focus on onboarding, the moment where he says disengagement runs deepest — is sensible precisely because it acknowledges that most brands aren't close to running this end-to-end.

The paid media problem

Morris's half of the conversation concerns the acquisition side — how brands find new customers through paid advertising — and his diagnosis of where that industry has ended up is worth sitting with.

His argument about walled gardens is essentially this: the major digital platforms have built closed ecosystems where advertisers hand over money and creative assets, the platform's algorithm decides who sees what and when, and the advertiser gets results back without ever seeing the mechanism. Morris describes this as a progressive loss of control dressed up as efficiency. The platform optimizes toward whatever outcome the advertiser specified — which is fine if you specified the right outcome, and a problem if you didn't.

"If you don't specify the right outcome for your brand," Morris says, "it will optimize towards the behavior and not necessarily the final outcome."

What I find genuinely clarifying here is his pivot on one-to-one advertising outside your own customer base. Morris admits, with some candor, that the industry has been overstating this capability for years. Stitching together a single person's identity across platforms and channels — knowing that the person who clicked your ad on one platform is the same person browsing your website from another device — has always been harder in practice than in the pitch deck. Privacy legislation is making it harder still.

His conclusion is a kind of professional about-face: maybe the answer isn't more sophisticated targeting. Maybe it's better creative reaching a broader audience, with AI helping you understand the cultural moment rather than the individual click. "Use AI to create Don Drapers," he says, "don't use it to create people who are doing that one-to-one communication."

The Don Draper line is doing some work there. It's memorable and it's selling an idea — that brand advertising is the sophisticated move, performance marketing is the trap — in a way that consultants tend to find more compelling than CFOs do. But the underlying argument isn't wrong. There's substantial evidence in marketing effectiveness research that brand-building investment pays out over longer timeframes than cost-per-acquisition metrics capture, and that chasing the cheapest click tends to hollow out a brand's pricing power over time. Morris is pointing at something real, even if the framing is slightly more dramatic than the evidence requires.

The agentic question nobody can answer yet

Both men spend time on what happens to advertising when AI agents start making purchasing decisions on behalf of consumers. The scenario: you tell your AI assistant you want new running shoes, and it searches, compares, and presents options without you ever seeing an ad. At that point, Morris argues, performance advertising as currently practiced becomes largely irrelevant. You can't retarget an algorithm. You can't catch someone "in market" if the algorithm is market-aware on their behalf.

What matters instead, in this world, is whether the consumer has a pre-existing preference for your brand — whether they'll tell their agent "find me Nike specifically" rather than "find me running shoes." That's a brand-building job, not a performance marketing job. And it means the creative, cultural work that the industry has been de-emphasizing for fifteen years becomes strategically important again.

I think this argument is directionally right and temporally uncertain. The agentic shopping future Morris describes exists in prototype; it doesn't yet describe how most people buy most things. Whether the timeline is three years or ten matters enormously for where brands should be putting money right now.

The organizational problem nobody wants to talk about

The most practically useful part of the conversation is about what it actually takes to implement any of this inside a real company. Burrell's observation that organizations often create a separate "AI unit" and then discover it doesn't understand the business well enough to improve it rings true. The people who know what a good retention intervention looks like are the CRM marketers who have been building customer journeys for years. The people who understand data infrastructure are the engineers. Getting both groups to build something together, and then giving up enough control to let an agent make the actual decisions — that's the organizational challenge, and it's harder than any of the technical pieces.

There's also the straightforward political problem: a marketing manager with quarterly targets is not going to hand decision-making authority to a system that needs months to learn before it delivers results, when their job security depends on this quarter's numbers. Burrell acknowledges this honestly. His advice is to start with a contained pilot at the moment of highest drop-off — and he identifies the early weeks of a new customer relationship as where the steepest disengagement typically occurs — rather than trying to transform everything at once.

That advice sounds modest for a conversation about AI transforming marketing. But modesty is what implementation actually looks like. The gap between what's technically possible and what organizations can absorb is where most technology transformations quietly stall — and it's where the real story of AI in marketing is currently being written, away from any hosted conversation.

---

Bob Reynolds is Senior Technology Correspondent at BuzzRAG.