I’ve spent close to three decades watching marketing technology solve the same problem in slightly better ways. First it was rules. Then it was segments. Then it was AI-assisted segments. Each wave promised to know the customer better, and each one did, a little. But underneath, the job stayed the same: react to what a customer just did, as fast as you can, across as many of them as your team can realistically manage.
That’s personalisation, and it has a ceiling. Not a technology ceiling, a human one. Even with a few million customers and a capable team, most marketers end up running about ten segments, because every segment is work: someone has to define it, build the content for it, schedule it, watch it, fix it when it stops working. The constraint was never the data. It was always how much thinking a team of people could do before the campaign had to go out.
That ceiling isn’t free, either. Every customer a brand can’t personalise to well enough to keep, it loses quietly, then pays again through paid media to win them back: the same customer, acquired twice. I’ve called this the reacquisition tax elsewhere, and it’s usually a bigger number than any marketing dashboard shows on its own.
What prediction actually changes
Prediction isn’t a smarter version of personalisation. It’s a different question. Personalisation asks: given what this customer did, what should we send them. Prediction asks: given everything we know about this customer, what are they about to do, and does that change what we should do right now, before they’ve told us anything at all.
The difference sounds subtle until you see it in practice. Take a customer browsing a pair of shoes and then closing the tab. The personalisation-era answer is automatic: cart abandoned, send a discount. It’s the right response to the wrong read. A system that actually knows this customer makes a completely different call. It knows she rarely responds to discounts, that she’s hesitating rather than leaving, that a better-margin alternative in her size just came into stock, and that she’s already had four messages from the brand today. So: no discount. A different product, in her voice, at a moment when she isn’t already saturated. The technology isn’t what changed. What changed is what the system was allowed to see, and what it was trusted to decide with that view.
Crocs’ move from reactive triggers to predictive propensity models of purchase intent yielded a 13X ROI from marketing investments. It wasn’t a larger personalization system driving this outcome. It was a different decision, earlier, based on a better question.
Why this requires more discipline, not less
Here is where I think most of the excitement around AI in marketing gets ahead of itself. None of this works if you hand an agent a goal and walk away. Every agent we’ve built operates inside guardrails a human set: what it’s allowed to decide on its own, what has to come back for review, and what it can never do regardless of what the numbers say. A consent flag is the clearest example of that last one. That division of labour is not a compromise on the technology. It’s the reason the technology can be trusted with real decisions at real volume.
This is also why we’ve changed our own business model to match. For years, martech has priced and judged itself on inputs, messages sent, monthly active users, segments built, because inputs were the only thing anyone could reliably measure at scale. Prediction changes that. When a system can act on a genuinely current, complete picture of a customer, you can finally hold it, and yourself, accountable for what actually matters. Did retention improve. Did lifetime value grow. Not did the campaign go out on time. That’s the shift we’ve made at Netcore.ai: in selected engagements where outcomes can be measured credibly, we tie part of our own fee to the outcomes we’re actually being hired to deliver. You can’t responsibly offer that kind of accountability on top of a system that’s still just reacting faster. You can only offer it once the system is predicting well enough that the outcome was genuinely earned, not got lucky on.
What this means for the people running marketing
The job changes, but not in the direction most people fear. It moves up, not out. Fewer hours spent building the fortieth version of a segment nobody has time to check on, more time spent deciding what a good outcome actually looks like for the business, and where the guardrails should sit. Our own customer success teams have effectively become growth engineers: less about configuring a tool, more about understanding what a client is actually trying to achieve and making sure the system stays aimed at it.
I don’t believe that this has been completely resolved anywhere, even in this case. Most of the things that appear obvious in a keynote take multiple iterations in the real world before they can stand up. And that’s okay, because the direction is where I have the most confidence: Marketing will be shifting from a reactive practice to a predictive practice, and the companies that can do it first will build a competitive advantage that won’t appear in dashboards until it’s too late.
