First-generation personalization engines mostly worked. The services army required to operate them was the real product — and when it left, the ability to answer "is this working?" left with it.
There's a version of the personalization story everyone in martech has already agreed on: the tools overpromised, the results underdelivered, and buyers got burned. It's a tidy story. It's also wrong about where the failure actually happened.
The engines were mostly fine. First-generation personalization platforms could genuinely do what the demo showed — behavioral targeting, dynamic content, product recommendations, real-time decisioning. The technology worked. I know because I spent years on the implementation side of those platforms, and the machines did what we configured them to do.
The failure was in what it took to configure them.
Here's the purchase math nobody put on the pricing page. You paid six figures for the software license. Then you paid another six figures — often more, often annually — for the professional services army required to make the software do anything.
Not to fix it. To operate it.
Every new campaign was a scoping call. Every rule change was a statement of work. Every "can we try showing returning visitors something different" was a ticket in someone's queue, estimated in hours, billed at consulting rates. The platform was sold as a product, but it functioned as a services engagement with a login screen.
And the buyers went along with it, because during the engagement, things happened. Campaigns launched. Dashboards moved. The quarterly business review had slides.
Then the services contract ends, or the budget line gets cut, or the agency rotates off the account. And the shiny platform is still there — still running, still billing — executing the same product recommendation logic someone configured eighteen months ago.
Nobody left on the team knows how it works. Nobody knows why those rules exist. Nobody knows what would break if they touched it. So nobody touches it.
The system degrades into an expensive appliance running stale logic. And here's the part that should bother you most: nobody can tell you whether it's driving lift. Not "the lift is small" — unknown. The measurement discipline lived with the consultants too. When they left, the ability to answer "is this working?" left with them.
That's the actual failure mode of the personalization era. Not bad technology. A cost-of-change so high that the system could only evolve while a services team was attached to it — and could only be trusted while that same team was grading its own homework.
If you have a personalization tool running today, three questions:
1. When was the last time the decisioning logic changed, and who changed it?
2. Could someone on your current team explain why a given visitor sees a given recommendation?
3. If you turned the whole thing off tomorrow, what number would move — and do you know that from a holdout, or from a vendor dashboard?
If the answers are "when the consultants were here," "no," and "we've never run a holdout," you're not running personalization. You're paying rent on a decision someone else made two budget cycles ago.
We built AXO around the belief that the cost of change is the product. Decisions run at the tag, in-session, against behavioral signals — no CDP project, no identity graph, no integration program that requires a bench of billable specialists to keep alive. Zones, segments, and variants are authored and adjusted directly, and every configuration ships with holdout-based lift measurement by default, so "is this working?" is a report, not a services engagement.
If a personalization system only works while an army is standing next to it, the army was the product. We'd rather sell you the machine.