Jean Perez runs data, analytics and AI at Valuedynamx, the Collinson company that powers rewards for some of the world’s biggest airlines, banks and hotel groups. The lesson from the conversation: the recommendation that looks wrong to a human is sometimes the model doing exactly what it should.
Here’s a story I can’t stop thinking about. Jean Perez’s team at Valuedynamx had human testers come back and say the AI recommendations made no sense. The testers had never shopped at the merchants being suggested. Some had never even touched the category. So was the model broken?
Nope. The model was working perfectly. It just didn’t have much personal data to work with on a test account, so it leaned on what similar customers were doing. That one story says a lot about how AI personalization actually works, and why gut checks aren’t always the best way to judge it.
How do you move an AI pilot into production?
Jean has been with Valuedynamx for nearly a decade, running the full data journey: ingesting client data, cleaning and governing it under different regional rules, and then turning it into reporting, insights and data products.
The way Jean’s team takes an AI project from pilot to production depends on the risk involved. Low risk use cases get lighter governance and can go from prototype to pilot to production in a few steps. Higher risk ones, like pricing decisions, go through much more scrutiny: fairness for the customer, regulatory requirements in each region, and data sharing rules with the countries where clients and their customers live.
At a point in which we need to start involving real data, real customers and starting to set up the pilot, we need to have all of these boxes ready and effectively approved before we move at that stage.
— Jean Perez, Director of Data, Analytics and AI, Valuedynamx
The part I liked is that the model doesn’t get judged only on accuracy. A lot of that can be tested in prototyping with data cohorts and synthetic data. The real test is the pilot, where Valuedynamx measures the recommendation engine against a control group that gets no personalized recommendations. Did conversion go up? Did it go up as much as the model predicted?
And there’s a human side to “working” too.
There is a big difference between a customer receiving a recommendation and thinking, this is great, I didn’t know about this, to receiving a recommendation and thinking, how on earth did I get that? Am I being observed?
— Jean Perez, Director of Data, Analytics and AI, Valuedynamx
That’s the trust line. Delight on one side, creepy on the other. Every brand doing personalization is walking it whether they admit it or not.
Why do AI recommendations sometimes defy human intuition?
This is where it gets interesting. Jean Perez walked us through the two engines behind most recommendation systems at Valuedynamx.
Content based filtering looks at you. Your transactions, how you interact, what you’ve done before. It recommends the next thing based on your own history.
Collaborative filtering looks at the crowd. If customer one behaves a lot like customer ten, and customer ten just started buying pet food, customer one might get a pet food offer too. It’s communal knowledge.
Most companies pick a fixed blend. Valuedynamx doesn’t.
It’s not a general, say, 50/50 split. It is truly, I’ll be looking at you, David, and saying what is a perfect split for David, and how that can be exposed in our recommendations.
— Jean Perez, Director of Data, Analytics and AI, Valuedynamx
So back to those confused testers. Their accounts had gaps in the data, so the model decided the crowd knew more about them than their own history did. Totally rational. Just not what a human expected to see.
Jean made another point I think every data team should tape to the wall.
A lot of the times what happens is not necessarily the model misbehaving. It’s mostly either a problem about data feeds or rules. What rules did we put for AI to follow, and do we have clean data to utilize?
— Jean Perez, Director of Data, Analytics and AI, Valuedynamx
That’s a big deal. When AI looks wrong, look upstream first. Nine times out of ten it’s the inputs. That’s the world we live in at BDEX every day.
How much data does a loyalty program actually generate?
I asked Jean Perez for real numbers, and the answer is huge.
If we had a client with a million customers, a bank, we will be receiving easily somewhere between 3 to 5 million transactions a day from them.
— Jean Perez, Director of Data, Analytics and AI, Valuedynamx
And that’s just bank transactions. On top of that, Valuedynamx tracks what happens on its own platform: points redeemed, items clicked and wishlisted, how currency is earned and spent. Jean estimates another million or so touchpoints a day across their data products, for a single client of that size. All of that gets fed back into the models: which offers you acted on, when, how long you took to buy, and how much you spent. Their commerce platform connects loyalty programs to more than 50,000 retail and travel partners, so the scale compounds fast.
Under the hood, the Valuedynamx stack is AWS for ingestion, processing and hosting, AWS Bedrock for AI models, Databricks for data models, lineage and access control, and Tableau for reporting back to clients and merchants. (We’re an AWS shop at BDEX too, so I felt right at home.)
How should brands balance loyalty value and marketing budgets?
Valuedynamx sits in the middle of a three way juggling act. Consumers want real value: cash back, points, experiences. Merchants want to reach specific customers, win back lapsed ones, or grow basket size from $150 to $200. And the bank or airline running the program needs it to pay off.
Merchants are getting a lot more demanding about that last part.
It’s not enough saying, can you put my name in front of customers for 20,000. They want to make sure that those 20,000 are well spent and they want to see a clear return on investment.
— Jean Perez, Director of Data, Analytics and AI, Valuedynamx
Jean also pointed out how much loyalty has changed. It used to be a supermarket club card where you silently earned points and got a voucher every so often. Now it’s earning dinner out from your weekly shop, or getting access to concerts, the symphony or the ballet. Jessie made the point that this is all driven by economics. Acquiring new customers is expensive, so keeping the ones you have is the smarter play.
What happens to loyalty when AI agents do the shopping?
This was my favorite part of the conversation. Jean Perez said they’re already exploring agentic commerce at Valuedynamx, and the reasoning is hard to argue with.
LLMs have been used to do the shopping. They’re not going to be caring about, okay, this person normally shops in this merchant. They’ll be trying to find the best price.
— Jean Perez, Director of Data, Analytics and AI, Valuedynamx
If an agent is optimizing for price, loyalty becomes the main way to beat it: cash back, exclusive perks, experiences you can’t get elsewhere. And Jean’s view is that you should be able to redeem your points right inside the LLM conversation, because that’s where shopping is going.
I threw out an idea on the show: an MCP server that runs your loyalty program, so shopping agents talk to it directly. The recommendations go to my agent, not to me. Jean didn’t think that was far off at all.
The bigger picture
What I keep coming back to is that “this recommendation looks weird” is not the same as “this recommendation is wrong.” Jean Perez’s team built models that blend individual and crowd signals per person, measure everything against control groups, and can explain why a result looks the way it does. That’s what grown up AI looks like.
But it only works if the data feeding it is clean. Jean said it plainly: when things break, it’s usually data feeds or rules, not the model. And if agents become the shoppers, the brands with the best data will be the ones that agents can actually talk to. That’s a future I’m genuinely excited about.
Connect with Jean Perez and Valuedynamx: Visit valuedynamx.com or find Jean Perez on LinkedIn.
Ready to put clean identity data behind your personalization? Visit bdex.com and click “Talk to an Expert.”
Watch the full episode: When AI Recommendations Challenge Human Intuition
This article was adapted from an episode of Deconstructing Data, BDEX’s weekly podcast on data-driven marketing. Tune in live every Thursday at 4:15 PM Eastern on LinkedIn.
About BDEX: For companies that need clean identity data to power their products, BDEX offers unmatched quality and execution. Visit bdex.com and click “Talk to an Expert” to get started.
