Executive Interview Nov 17, 2026 11 min read

Amperity’s Derek Slager: Why AI’s Hardest Job in 2027 Is Knowing When Not to Act

Introduction

In 2027, creating marketing content with AI will be the easy part. Brands can already generate a personalized offer in seconds. The harder question is whether that offer should go out at all, and whether the “new” customer receiving it is someone the brand already serves.

The stakes are rising. According to Amperity’s 2026 Consumer Priorities Report, 60% of US consumers say AI has led them to choose a brand they hadn’t considered, and 63% say they would switch brands for a better offer. Personalization can keep those customers, but only when it feels relevant and timely. Get it wrong and it backfires: 58% say invasive or creepy personalization makes them less likely to choose a brand.

In this Q&A, Derek Slager, co-founder and co-CEO of Amperity, explains what happens as AI moves from making content to making customer decisions. He discusses why brands end up paying to acquire customers they already have, the customer context AI needs before deciding whether to engage, which decisions should require a human sign-off, why marketers need to learn when to push back on AI recommendations, and how loyalty programs can prove their value. He also shares what separates companies that scale AI from those stuck in pilots, and where the line falls between helpful personalization and intrusive monitoring.

Derek Slager

Derek Slager

Founder and Co-CEO,, Amperity

The following conversation was developed through CloudTweaks’ Adaptive Executive Interview Program.
In conversation with Derek Slager — Founder and Co-CEO,, Amperity
Question 1

Tell us a little about yourself, your background, how you got to where you are, and what you're focused on right now.

I've spent most of my career building data systems that have to be right at scale. Early on, I led engineering teams working on large distributed systems and security, where a small error never stays small. That shaped how I think about data. If people can't trust it, they won't use it, no matter how clever the software on top of it is.

Kabir Shahani and I were on the founding team at Appature. We kept hitting the same wall. Large brands had enormous amounts of customer data and still couldn't answer a basic question: who are our customers? The same person showed up in a dozen systems under different emails, addresses and loyalty numbers. We started Amperity in 2016 to work on that problem. I served as CTO for several years, and this year Kabir and I became co-CEOs.

Right now I'm focused on what happens as AI moves from writing content to making decisions about customers. AI has made producing a personalized email much easier. The harder question is whether you should send it at all. That depends on knowing who the customer is, what just happened and what would actually be useful to them. That gap is where I spend most of my time, both with brands and inside our own company as we rebuild how we work around AI.

Question 2

What separates companies that scale AI across their business from those that stay stuck in pilots?

I think the biggest difference is whether the company is willing to change how the work gets done. You can add AI to an existing process and make parts of it faster. But if the same handoffs, approvals and disconnected systems are still there, you haven't changed very much.

A pilot can look successful without ever testing that. A team exports some data, gets a useful result and puts it in a presentation. Nobody has worked out how the result gets into the systems people use every day, who owns the decision or what happens when the system gets something wrong. The experiment worked, but the business still operates the same way.

I'd start with a decision the business already makes repeatedly, like whether a customer should receive a reminder or whether a service issue needs to be escalated. Establish how well that decision works today, then connect the data, permissions and workflow needed to improve it. A business owner needs to be accountable for the outcome, with the technical team responsible for making the system reliable.

That doesn't mean waiting until every data problem is solved. It means understanding which information this decision depends on, making that context dependable and knowing when the system needs help. If you aren't willing to change the process, there's a limit to what a better model can do for you.

Question 3

One of your 2027 predictions is that brands will find they're paying to acquire customers they already have. Where does that disconnect usually start, and what's the quickest way for a marketing team to measure how much of its "new customer" spend is going to existing customers?

It usually goes back to how the business grew. E-commerce, stores, an app, loyalty programs and customer service each have their own system, often run by different teams. A shopper who joined a brand's loyalty program in-store with a phone number, then shops online with a work email, can look like two people. Ad platforms only see the customer data available to them. If your existing-customer audience doesn't include that work email, a five-year customer can look new.

The quickest test is to start with a sample of last quarter's reported new-customer conversions from your paid acquisition campaigns. Match them against purchase history across channels, using the identifiers you have permission to use, such as email, phone, postal address and loyalty ID. Check whether each customer had already purchased before the conversion, using an agreed definition of an existing customer.

That gives you a starting estimate of how many reported new customers were already buying from you. It doesn't tell you exactly how much spend was wasted, but it tells you whether your acquisition reporting is working from the same understanding of a customer as the rest of the business.

Finance and marketing need to agree on that definition. Existing-customer audiences need to stay current as people buy and their identifiers change. And win-backs should be reported separately from truly new customers.

Question 4

A customer with an unresolved return probably doesn't need another promotion. Before AI decides whether to contact a customer at all, what's the minimum context it should have, and which signal do brands most often miss?

Before you let AI decide whether to contact someone, it needs to know who that person is, what just happened and whether you have permission to contact them through that channel. An unresolved return, a purchase yesterday or five messages already this week can all change the right decision.

Customer service history is a particularly important blind spot. Support tickets, call notes and in-store returns can sit in systems marketing never sees. So the same week a customer spends 40 minutes on the phone about a damaged order, they get a cheerful "We miss you, here's 15% off" email. From the customer's perspective, the company clearly doesn't know what just happened.

You can look for this in your own data. Take a sample of customers who unsubscribed or complained after a campaign and check what happened beforehand. If you keep finding unresolved returns or service problems, that's a reason to investigate whether those signals are reaching the system deciding what to send.

The fix is to make service and returns context part of that decision, with rules for holding or adjusting a promotion when a problem is still open. Sometimes the most useful thing a brand can do is resolve the problem before asking for another purchase.

Question 5

Customer data is rarely complete. How should AI handle missing or conflicting information, and how should it know when it doesn't have enough context to act?

The first thing to recognize is that incomplete data is normal. One system says a customer lives in Denver, another says Seattle, and a store purchase never made it into the online account. The important question is whether that uncertainty changes the decision you're about to make.

An uncertain address might not matter when choosing a general product recommendation. It matters a great deal when shipping an order. And if you don't know whether you have permission to send a promotional message, a less personalized message doesn't solve the problem. You need to confirm the permission before sending it.

Those distinctions need to be built into the system. Define which information each action requires, how conflicting records should be resolved and when the system should hold the action or ask a person. A model saying it's confident is not enough. You need evidence that the underlying records belong to the right customer, are current enough for the decision and support the proposed action.

Test what happens when records conflict or a required signal is unavailable. Does the system stop where it should, or does it fill in the gap and keep going? You don't need perfect data for every decision. You do need to know which gaps make a particular decision unsafe.

Question 6

As AI moves from creating marketing content to making customer decisions, such as who gets an offer, at what price, and who is held back, which of those decisions should require a human sign-off? How do you keep those checkpoints from slowing everything down?

I'd sort decisions by how hard they are to undo and how much they affect trust. Send time, channel choice and which approved creative a customer sees are generally lower risk when consent, messaging frequency and service issues are already accounted for. Those can run within boundaries a person has reviewed and approved.

Pricing and discount depth that vary by customer deserve stronger oversight. Two people paying different prices for the same thing can quickly look unfair. So do decisions that exclude someone from a benefit or service level based on their predicted value. And before a model makes a new type of decision at scale, a person should review how it works and what happens when it gets something wrong.

You keep this fast by deciding in advance which decisions can run within an approved policy and which require individual review. You also need a record of the data used, the policy applied and the action taken, plus a way to stop the system when something goes wrong. Reviewing every output can turn into rubber-stamping. The controls need to match the consequences.

Question 7

Knowing when to question an AI recommendation is becoming a core marketing skill. Without naming a client, can you walk us through a recommendation that looked right but wasn't, and what should have tipped off the person reviewing it?

Consider a win-back recommendation built only on online purchase history. A customer hasn't bought online in six months, so the model treats them as lapsed and recommends a discount. But that person bought in a store last week.

The recommendation can look perfectly reasonable given the data the model received. The problem is that the data describes an inactive online account, not an inactive customer.

We know the identity problem behind that example is real. One fashion retailer found that 31% of its top customers used multiple email addresses and 71% shopped across multiple channels. Customers with multiple identifiers spent 95% more and placed twice as many orders. An incomplete view can cause you to misunderstand some of your most valuable customers.

The reviewer should ask what "lapsed" actually means in the data. Does it include store purchases? Does it connect the customer's different email addresses? An unusually strong result is a reason to understand the mechanism before committing more budget.

The skill marketers need isn't distrust of AI. It's the habit of asking what data the recommendation was built on, and what it might be missing.

Question 8

Once AI is making customer decisions, how should organizations learn from the outcomes and use them to improve those decisions over time?

Start by distinguishing what happened after a decision from what the decision actually changed. A customer buying after receiving an offer doesn't tell you whether the offer caused the purchase. They may have been about to buy anyway. If you feed every one of those sales back as a success, you can end up optimizing for behavior you didn't create.

Holdout groups help you test that. Where practical, keep a comparable group that doesn't receive the intervention and look beyond the immediate response. Did the customer come back? Did they return the item, contact support or unsubscribe? A discount can win a sale today while encouraging someone to wait for discounts next time. The measurement needs to account for both.

You also need to connect the outcome to the original decision: what the system knew, what it recommended, what actually happened and whether a person changed the recommendation. Otherwise, you can see that something worked without understanding why.

Human overrides are useful evidence, but they aren't automatically the right answer. Record the reason and review the pattern. A marketer may know about a service problem the model couldn't see. Or the marketer may be relying on an assumption the results don't support. Either way, there's something to investigate.

Learning also doesn't have to mean letting a model rewrite its own rules after every interaction. Sometimes the right improvement is better source data, a different objective or a clearer approval boundary. Test the change before expanding it.

Question 9

Amperity's 2026 Consumer Priorities Report found that 60% of US consumers say AI has led them to choose a brand they hadn't considered. If brand switching is getting that easy, what should a loyalty program be measured on in 2027 instead of membership growth?

Membership growth tells you how many people signed up, not whether the program changed what they do. In our 2026 Consumer Priorities Report, 63% of US consumers said they would switch brands for a better offer. Signups alone can't protect a brand from that. Your best customers tend to join anyway, so a program can look successful while simply rewarding people who would have bought regardless.

I'd start with a simple question: what did the program cause customers to do that they would not have done otherwise? Then look at additional purchases, retention and the profit left after paying for rewards. Active membership matters too, because a growing enrollment number can hide a shrinking group of people who actually use the program.

You also need to connect loyalty IDs to purchases made through guest checkout and in-store so you can see each member's relationship with the brand. For one grocery chain, connecting those records improved the number of transactions attributed to loyalty members by 18%, including previously unmatched purchases.

The goal is to design benefits around what members do after they join, then measure whether those benefits give them a reason to keep choosing you.

Question 10

As AI makes more of these customer decisions, where should brands draw the line between helpful personalization and intrusive monitoring, and what will set apart the brands customers trust in 2027?

I'd draw the line at whether the use of the data makes sense to the customer. A refill reminder for something they buy every six weeks can be useful. A message telling them you noticed they walked past a store yesterday can feel like surveillance. Permission matters, but so do the purpose and the customer's expectations.

Before using a signal, ask whether you could explain plainly why you have it, why you're using it and what the customer gets in return. If that explanation is uncomfortable, that's a reason to reconsider the use. Technical access to data doesn't make every use of it a good idea.

Our 2026 Consumer Priorities Report found that 58% of consumers said invasive or creepy personalization makes them less likely to choose a brand. That is a pretty clear warning against assuming that more data and more personalization always produce a better experience.

As AI takes on more of these decisions in 2027, those boundaries need to be part of how the system operates. Use data for purposes the customer understands and has permitted, keep sensitive inferences out of routine targeting, and make preferences easy to change. The system also needs to respect those changes wherever it acts.

The brands that earn trust will be able to explain what they're doing and show why it's useful. Sometimes that means a timely reminder. Sometimes it means solving a problem quietly. And sometimes it means recognizing that you have no good reason to contact the person at all.