Derek Slager, Co-Founder and Co-CEO of Amperity – Interview Series

Derek Slager is co-founder and co-CEO of Amperity, where he leads the company’s AI-first transformation across both product and the way the company operates. He co-founded Amperity to give marketers and analysts customer data they could trust, and built the patented identity resolution and real-time profile architecture behind Amperity’s trusted customer context. Earlier, he was on the founding team at Appature and held engineering leadership roles in large-scale distributed systems and security.

Amperity is an AI-powered customer data cloud designed to help enterprises unify fragmented customer information from online and offline sources into accurate, real-time customer profiles. Its platform combines machine learning-based identity resolution, data management, analytics, predictive modeling, and activation tools, enabling marketing and data teams to better understand customer behavior, personalize experiences, measure revenue impact, and coordinate actions across channels. Founded in 2016, Amperity serves more than 400 brands worldwide, including Alaska Airlines, Virgin Atlantic, Wyndham Hotels & Resorts, and DICK’S Sporting Goods.

You’ve worked across advanced analytics, customer insights, product management, ML/MLOps, and AI strategy, helping build foundational products before Amperity. How has that path shaped the way you think about the relationship between customer data quality and AI performance?

My work in the industry has really centered on one problem: how do you get AI to reason correctly about people when the data describing those people is always incomplete, inconsistent, and changing? Early on, in analytics and product roles, I watched teams build increasingly sophisticated models on top of customer data that was fundamentally unreliable: duplicate records, mismatched identifiers, stale attributes. No amount of modeling sophistication fixed that; it just hid the problem more effectively.

That instinct is what led us to build Amperity: identity resolution isn’t a nice-to-have layer on top of AI, it’s the foundation everything else sits on. When I look at enterprise AI today, I see the same pattern playing out at a much larger scale. Teams are pouring resources into model selection and prompt engineering while treating the customer data feeding those models as a given. It isn’t. Get the identity layer wrong, and every downstream AI system inherits that error and amplifies it.

We often hear enterprise AI discussions framed around model selection, but you argue that trusted customer identity data is becoming just as important. Why is the conversation shifting from “Which model is best?” to “What data can AI actually trust?”

For the first few years of the generative AI wave, model capabilities were really the bottleneck, as the tools simply weren’t good enough yet. Now, that’s changed. Leading models today are remarkably capable, and the gap between them has narrowed for most business use cases. What hasn’t caught up is the data feeding those models inside the enterprise.

You can put the most advanced model in front of a customer service agent or a personalization engine, but if it’s reasoning over three duplicate profiles for the same person, an outdated email address, and a purchase history missing half the channels that person shops in, the output is bound to be wrong. The conversation is shifting because leaders are feeling that gap directly. They’ve adopted a great model and still aren’t seeing the ROI they expected. The next real advantage won’t come from a better model. It’s going to come down to whether an organization can give its AI systems something trustworthy to reason over in the first place.

What are the most common ways inaccurate or fragmented customer identity data can undermine AI outcomes, even when the underlying model is technically strong?

A few patterns show up constantly. The most basic is duplication. The same customer exists as five or six different records across channels, each with a partial view of that person, so an AI system might send a win-back offer to someone whose subscription is active under a different record. Then there’s staleness. Data that was accurate a year ago but hasn’t been refreshed, so a model is optimizing based on who a customer used to be rather than who they are now. There’s also over-merging, which is more subtle: identity systems that are too aggressive about matching end up combining two different people into one profile, and now an AI agent is confidently making decisions on behalf of someone who doesn’t actually exist. And finally, there are plain gaps. Loyalty activity in one system, support history in another, browsing behavior in a third, with no single system holding the full picture. Individually, each of these looks like a data hygiene issue. In an AI context, they compound because the model doesn’t know what it doesn’t know. It just produces an answer with total confidence.

As generative AI becomes a new front door for product discovery, what changes for brands when consumers increasingly rely on AI systems to compare options, evaluate loyalty, and make purchase decisions?

Usually, brands controlled the front door. You built a website, optimized for search, refined your funnel, and customers came to you directly or through results you could influence. Now, that front door is a conversation with an AI assistant that a brand doesn’t control and can’t fully see inside of.

Someone asks an AI tool to compare loyalty programs or recommend a hotel, and the brand’s job now is to be legible and trustworthy enough, in that moment, to be the answer the AI gives. That’s a genuine shift in what marketing must optimize for. It means the quality, structure, and accuracy of a brand’s own data and content matter more, because AI systems are synthesizing from whatever signal they can find.

Amperity’s research suggests that only a minority of consumers now go directly to familiar brands without considering AI recommendations first. What does this mean for brand loyalty, personalization, and the traditional customer journey?

We surveyed 1,000 U.S. consumers earlier this year, and one number stood out to me: fewer than a quarter said they go straight to a brand they already know without first factoring in an AI recommendation, and 60% said AI has already led them to a brand they hadn’t previously considered. That erodes the default advantage established brands once had. Loyalty hasn’t disappeared, but the funnel above that decision has changed. Brands are being evaluated and shortlisted by systems they don’t control, based on whatever data is available about them.

Personalization strategies built on the assumption that ‘we already know this customer, so we don’t need to earn their attention’ will underperform. At the same time, the customer journey itself has gotten less linear. A customer can arrive at a brand’s site already having been ‘sold’ by an AI’s summary of reviews and comparisons, which means the brand’s own experience needs to hold up against expectations it never got the chance to shape directly.

Identity resolution has existed as a data problem for years. What makes it more urgent now that AI agents, copilots, and generative AI interfaces are becoming part of daily enterprise workflows?

Identity resolution has always mattered for things like preventing duplicate mail or making sure a loyalty member gets proper credit for a purchase. Those are real problems, but relatively forgiving ones where a human usually catches the mistake. Agents remove that safety net.

When an AI agent is empowered to actually take action, whether that’s issuing a refund, applying a discount, recommending a next-best offer, or updating a preference. It’s doing so on top of whatever identity resolution it’s been given, at machine speed and often without a human in the loop. Get the identity wrong, and the agent doesn’t just misfire once. It can misfire systematically and immediately, at scale, before anyone notices. That’s what makes this moment more urgent than before.

How should companies think about customer identity as part of their broader AI governance strategy, alongside privacy, security, model monitoring, and compliance?

I’d put customer identity right alongside those other pillars, not underneath them. Most AI governance conversations focus on the model: is it biased, secure, compliant and being monitored for drift? Those are necessary, but they assume the input the model is reasoning over is sound. If the identity layer is wrong, you can have a perfectly governed model producing decisions about the wrong person, or a model that appears biased when the real issue is that its data was fragmented in ways that disadvantaged certain customer segments.

Practically, this means the same rigor organizations apply to model monitoring needs to apply to the identity and data layer underneath it. Who owns the customer’s golden record? How is match quality measured and monitored over time? What happens when a match is wrong? Those should be governance questions with the same priority as anything happening at the model layer.

Many organizations are investing heavily in AI pilots, but their customer data remains siloed across marketing, commerce, support, loyalty, and offline systems. What practical steps should leaders take before scaling AI into operational decision-making?

The pattern I see most often is an organization running several promising AI pilots — a chatbot here, a personalization experiment there — each plugged into whatever data happens to be closest at hand for that team. It works well enough to get funded, then hits a wall when someone tries to scale it into an actual operational decision, because the pilot’s data never reflected the full customer.

The first practical step is an honest inventory: know where every source of customer truth lives (marketing, commerce, support, loyalty, offline). How much do they overlap? Where don’t they? Second, invest in identity resolution before investing in more pilots. It’s what lets every future AI initiative reuse the same trustworthy foundation, rather than each team rebuilding a partial one. Third, build a small number of cross-functional owners for that unified customer data, rather than leaving ownership to whichever team happens to need it for their current pilot. Organizations that get this sequence backward end up with a portfolio of AI pilots that never leave the pilot stage, because none of them can be trusted with a real decision.

Where do you draw the line between personalization that feels useful and personalization that feels invasive, especially when AI systems can infer more from unified customer profiles?

The line is less about how much a brand knows and more about whether what they do with that knowledge feels earned and relevant in the moment. A recommendation based on someone’s actual purchase history feels helpful. A recommendation that reveals an inference the customer never explicitly shared feels invasive even when it’s accurate, because it exposes how much is being shared.

Our report found that a large majority of consumers say they’re more willing to engage with personalization when they trust how their data is being used, and a majority say personalization that feels invasive makes them less likely to choose a brand at all. My rule of thumb for teams building on AI-driven inference is simple: if you wouldn’t want to explain, in plain language, why the system knows what it knows, don’t act on it yet. Personalization should feel like being recognized, not surveilled.

Looking ahead, what will separate enterprises that get real value from AI from those that remain stuck in experimentation mode: better models, better data infrastructure, stronger governance, or a new operating model altogether?

Model quality was never going to be the durable differentiator. It evolves too quickly, and most enterprises have access to roughly the same models. Governance matters, but it’s a constraint, not a source of value on its own. The two things I actually think separate the enterprises pulling ahead are data infrastructure and operating model, which are connected.

The companies getting real value from AI are the ones that have already invested in having a single, trustworthy, continuously updated view of the customer, so any new AI capability can be pointed at that foundation instead of stitching together its own. But that only works if the operating model changes too: if AI initiatives are still owned team by team, with each group defining its own version of the customer, you’ll keep introducing the same data quality problems. The organizations that break out of experimentation mode are the ones that treat trusted customer data as shared infrastructure and build their AI operating model around that assumption from the start.

Thank you to Derek for the insightful interview. Readers can learn more about Amperity and explore Derek Slager’s latest insights on his Unite.AI author page.