Fern Halper, Founder of AI Foundations Group and VP of Research at TDWI, has been doing AI since the Bell Labs days. Her take: most AI failures are data failures wearing a disguise.
Back in the 1990s, Fern Halper and her team at AT&T Bell Labs could already predict which customers were about to cancel their phone service. The model worked. The problem was nobody could use it. When they brought it to the call center, the answer was basically “great, but how do we operationalize that?” There wasn’t the compute, and there wasn’t the process. Thirty years later, the compute is cheap and the models are everywhere, but that same lesson is still the one most companies are learning the hard way.
I realized there that we could build a model that could predict, say, customer churn, but actually getting that model into the business process was really hard. And that lesson stayed with me, that the algorithm was really only part of the story.
— Fern Halper, Founder, AI Foundations Group
Fern is now the Founder of AI Foundations Group, where she mentors executives and boards on AI decisions, and she still leads research at TDWI. She also wrote a book called Data Makes the World Go ‘Round, which is about exactly what sits behind the “why can’t we just build a chatbot like that?” request. Jessie and I had a great conversation with her, and I came away with a few things I think every marketing and data team should hear.
What does a strong data foundation for AI actually look like?
Here’s the thing. A lot of people think of a data foundation as the place where data lives. A warehouse, a lake, a lakehouse. Fern’s definition is much bigger than that. For her, it starts with what you want the AI to do, and then asks whether your data is fit for that purpose, integrated, governed, and accessible across the company.
The data foundation is really more than a place to store the data. It’s the ability to find and integrate and prepare and govern and serve the data with enough business meaning to actually use it responsibly.
— Fern Halper, Founder, AI Foundations Group
Why does this matter so much? Because the generic stuff, like summarizing documents or drafting emails with Copilot, is available to everybody. Fern talks about a “value ceiling.” Companies can squeeze some productivity out of AI that way (and plenty of them aren’t even measuring it), but the companies that treat their data foundation as infrastructure for AI are the ones that break through. They measure impact and they get more value.
The real value from AI comes from actually using your company’s data, because that’s where your proprietary advantage lives.
— Fern Halper, Founder, AI Foundations Group
That’s a big deal, and it’s something I say to our customers all the time. Your own data is the moat. Everything else is a commodity.
Why isn’t data cleanup a one-time project?
This part of the conversation hit close to home for me. Fern pointed out that one of the biggest mistakes companies make is treating the foundation like a spring cleaning. You clean up the CRM for the pilot, the pilot looks great, and then everyone moves on. Meanwhile campaign data changes, business definitions change, and the data drifts.
Another mistake that they make is treating the foundation basically as a one-time cleanup job. Even if you have a clean pilot data set, it doesn’t always stay clean.
— Fern Halper, Founder, AI Foundations Group
And when it drifts, the AI gives you a plausible answer built on stale data, or on a metric that means something different in another department. Then everybody points at the model.
The organization blames the AI when the failure is really much more early on in the data path.
— Fern Halper, Founder, AI Foundations Group
In the identity world, we live this every day. Data is a living thing. People move, change jobs, change emails. When a customer tells me they only want their data refreshed once a month, I have to explain that by the end of that month a meaningful chunk of it has already decayed. Fern made the same point about models. Traditional predictive models decay as conditions change, and the freshness of the upstream data matters just as much as the model itself.
How should companies manage and validate unstructured data for AI?
And here’s where it gets interesting. According to Fern Halper’s research at TDWI, the share of companies using unstructured data (documents, emails, chats, call notes) is now on par with structured data. Generative AI made that possible. One of the biggest use cases she sees is running call center notes through an LLM to pull out the classes of issues driving churn.
The problem is that governance hasn’t caught up.
I’d say that there’s a 20% gap between people trusting their unstructured data and trusting their structured data.
— Fern Halper, Founder, AI Foundations Group
You can’t just dump a pile of call notes into a prompt and call it a system. If you want to operationalize it, you have to chunk the data, extract information, validate what the AI pulled out, and define what “accurate” even means for a call center note. Fern recommends building an ontology: the entities (customer, agent, product), the intent (billing dispute, technical glitch), the sentiment, and the resolution. That’s how you get consistent answers instead of a different result every time you run it.
She also called semantics the hot topic of 2026, which I loved. It used to be a BI conversation about metric definitions. Now it’s about business meaning across structured and unstructured data, and the relationships between them.
My own rule of thumb lines up with this. I tell people to treat AI like a brand new employee who has never done the job. Give it the full explanation, even if that explanation runs pages. The more specific you are, the better it performs.
What’s the difference between data governance and AI governance?
Fern Halper draws a clean line here. Data governance asks whether the information is understood, fit for purpose, protected, and used properly. AI governance asks what the model or system is doing, how it’s tested, what it can produce, and who is accountable. With agentic AI, accountability gets a lot harder. You need to know where the data came from, which agents touched it, how it changed along the way, and what action came out the other end.
One of the things that organizations don’t do well with data governance is that they put it in at the end of a project.
— Fern Halper, Founder, AI Foundations Group
By then, data has already been copied into a dozen tools and the app is built. Fixing it is painful. The companies that succeed see governance as an enabler, not a compliance tax. They bring in security, legal, and the business people who know what the data is supposed to do. Many are moving to a data product model, where every data set has an owner, documentation, and users it has to satisfy.
So what can a small marketing team do right now? Fern’s answer was refreshingly human. When two systems disagree about the same customer, get the people in a room, trace it backwards, and agree on a definition. The single view of the customer has been the holy grail for 30 years.
If you don’t have a single view, at least you need to know that you don’t have a single view, and how you’re going to treat it in multiple systems if the view isn’t the same.
— Fern Halper, Founder, AI Foundations Group
The bigger picture
What I appreciated most about this conversation is that Fern isn’t selling hype. She’s been watching companies try to make AI work for three decades, and the pattern hasn’t changed. The algorithm is the easy part. The hard part is the data, the semantics, the governance, and the culture that lets people disagree productively about what a “customer” is.
We’re building an MCP server at BDEX right now, and every conversation like this reminds me why the quality of the data underneath matters more than the shiny interface on top. If you feed an agent stale, ungoverned data, you get confident, wrong answers at machine speed. If you feed it clean, fresh, well-defined data, you get something that actually moves the business.
Fix the foundation first. The AI will thank you.
Connect with Fern Halper: Learn more about AI Foundations Group at aifoundationsgroup.com, check out her book and blog at datamakesworld.com, or connect with her on LinkedIn.
Want cleaner data powering your AI? BDEX helps companies build on identity data that’s fresh, verified, and ready to use. Visit bdex.com and click “Talk to an Expert.”
Watch the full episode: What It Takes to Make AI Work on YouTube
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.
