Latest / The CMO Podcast with Fexingo: Marketing Leadership, Budgets, and Executive Strategy / How CMOs Are Using AI Agents for Customer Retention
Transcript
- Lucas: There's this phrase floating around marketing circles right now — 'the agent isn't the chatbot your grandmother used.' And it's actually a useful distinction, because we're seeing a real shift from reactive customer service bots to proactive AI agents that don't wait for the customer to raise their hand. Luna: Proactive how? Like, the agent just starts texting people out of the blue? Lucas: Pretty much, but in a way that's actually helpful. The canonical example right now is Sephora's Beauty Insider VIP agent. It's an AI that lives inside their app and their loyalty program. It doesn't just answer questions — it monitors each customer's purchase history, product usage cycles, and even things like local weather or seasonal trends. And then it initiates conversations. Luna: So if I bought a certain moisturizer three months ago, it might ping me and say 'hey, you're probably running low'? Lucas: Exactly. And it goes further. It knows what shade of foundation you bought last time, and if there's a new formula for summer, it'll send a sample offer. Or if a product you buy regularly is about to be discontinued, it alerts you and suggests a replacement before you're stranded. Sephora's CMO Mary Beth Laughton presented some of this at a conference last month — she said the agent reduced churn in the VIP tier by 18 percent in the first half of 2026. Luna: Eighteen percent is big. And that's just from automated messages? There's no human in the loop? Lucas: There is for escalations. But the majority of interactions are fully autonomous. The agent learns each customer's preferences over time — it builds what Sephora calls a 'beauty genome' for every VIP member. So it knows that Luna prefers fragrance-free products, or that Lucas tends to buy during sales. It can time its messages accordingly. The key is that it's permission-based — you opt into the agent when you join the loyalty tier. And you can mute it anytime. Luna: I can see the upside, but also the creep factor. There's a fine line between 'helpful reminder' and 'big brother tracking my moisturizer usage.' How do they handle that? Lucas: That's actually where the data clean room comes in — you remember we talked about those in episode 82. Sephora built a dedicated clean room for the agent so that all the purchase data and behavioral signals are anonymized and aggregated before they feed the AI. The agent doesn't see raw purchase history — it sees a vector of preferences. And customers can request a full deletion of their profile at any time. Mary Beth Laughton framed it as a trust contract: 'the agent is your personal shopper, not our surveillance tool.' Luna: That's a good line. But let's talk about cost. Sephora has the budget and the data science team. What about a mid-size brand with fifty thousand customers? Lucas: That's where the SaaS layer comes in. Companies like Kustomer and Zendesk have launched AI agent tiers that are much more accessible. Zendesk's new 'Proactive Agent' product starts at about fifteen hundred dollars a month for up to ten thousand active customer profiles. It integrates with Shopify, Salesforce, and most POS systems. So you don't need a team of engineers — you configure the triggers and the tone in a dashboard. Luna: Fifteen hundred a month is not nothing for a small brand, but it's not insane either. And the ROI case would be pretty straightforward if it reduces churn. Lucas: Right. Sephora's pilot cost them two point one million dollars — that included custom development, the clean room integration, and six months of training data. But Laughton justified it purely on lifetime value math. The VIP tier customers have an average LTV of about two thousand dollars. If the agent reduces churn by even ten percent, that's two hundred dollars per retained customer. Multiply by their VIP base of roughly two million, and the math gets compelling very quickly. Luna: So the question becomes: is this a retention channel that will eventually replace email and SMS? Or is it additive? Lucas: I think additive, at least for now. The agent is more conversational and more contextual than a broadcast email. But it's not going to replace the weekly newsletter or the transactional shipping notification. What it does is fill the gap between those touchpoints — the moments when a customer might drift away simply because no one checked in. And that's where the real churn happens: not during the big campaigns, but in the silent weeks between purchases. Luna: You know, thinking about this, it reminds me of something our listeners might find useful — the kind of insight that sticks with you long after the episode ends. And if that's the case, it's worth mentioning that we keep this show ad-free and listener-supported. So if today's conversation felt worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. Lucas: Absolutely. Even a small contribution helps us keep digging into these topics without any sponsor influence. And we really appreciate everyone who's chipped in. Luna: Alright, back to the agent. One thing I'm curious about — how do they measure success beyond churn? Are there secondary metrics? Lucas: Great question. Sephora tracks something they call 'agent-influenced basket size.' When a customer engages with the agent and then makes a purchase within 72 hours, they look at whether the basket is larger than their average. In the pilot, agent-influenced baskets were twenty-two percent larger. The theory is that the agent's recommendations lead to add-ons the customer wouldn't have thought of. Luna: So it's not just defensive — it's actually driving revenue. That makes the ROI case even stronger. Lucas: Exactly. And there's a retention compounding effect. Customers who interact with the agent at least once a month have a 92 percent retention rate over six months, versus 76 percent for those who don't. That's a massive gap. Of course, there's selection bias — the people who opt in are already more engaged. But the data team controlled for that by matching cohorts on past engagement levels. The effect held. Luna: Alright, but what about the brands that don't have a loyalty program as robust as Beauty Insider? Can they still pull this off? Lucas: They can, but it's harder. The agent needs a rich data set to be useful. If all you have is an email address and a few purchase dates, the agent's suggestions will be generic and might feel spammy. That's why the most successful deployments so far are in industries with high repeat purchase frequency — beauty, groceries, pet supplies, subscription boxes. Brands with lower frequency, like furniture or electronics, might need to layer in browsing behavior or third-party data to make the agent smart enough. Luna: Which brings up the privacy question again. If you're using third-party data to feed the agent, you're basically doing personalized advertising in a one-to-one channel. That feels riskier. Lucas: It is. And that's where a lot of CMOs are drawing the line. The general consensus I'm hearing is: the agent should only use data the customer has explicitly shared with you. No inferred demographics, no purchase data from other retailers. Keep it within the first-party relationship. That way, the agent feels like a helpful friend, not a creepy stalker. Luna: Let's talk about the human side. What happens to the customer service team when the agent handles 80 percent of retention touches? Lucas: At Sephora, they actually redeployed their VIP support agents to higher-value tasks — handling complex returns, training the agent on edge cases, and doing proactive outreach for high-risk churn signals that the agent flags. So the headcount didn't shrink, but the work shifted from repetitive answers to judgment-intensive work. Laughton described it as 'elevating the humans, not replacing them.' Luna: That's the ideal scenario. But I imagine some brands will use it to cut costs. Lucas: Probably. And that's a choice. But the data suggests that if you cut too deep, you lose the human touch that builds loyalty. The best performing agents are the ones that hand off to a human at the right moment — when the customer is frustrated, or when the request is ambiguous. The AI is great at routine, but it's terrible at empathy. Luna: So the CMO's job becomes designing the handoff rules. When does the agent escalate? What triggers a human intervention? Lucas: Exactly. And that's a hard problem. Sephora's early version escalated too often — every time a customer typed 'no' or 'stop,' it kicked to a human, which overwhelmed the team. They had to retrain the model to understand context — like 'no, I don't need that today' versus 'no, stop contacting me.' That nuance is everything. Luna: I can see this becoming a whole new sub-discipline: conversational escalation design. Lucas: It already is. There are consultancies now that specialize in exactly that — writing the decision trees and tone guidelines for AI agents. And the good ones charge a premium, because getting it wrong means annoying your best customers. Luna: Last question: how does this change the marketing org chart? Who owns the agent? Lucas: That's a live debate. In some companies, it sits under CRM or loyalty marketing. In others, it's under customer service. The most forward-thinking ones are creating a new role — something like 'conversation experience manager' — that sits between marketing and service, reporting to the CMO. Because the agent is both a retention tool and a brand touchpoint. It can't be siloed. Luna: So the CMO becomes the steward of the agent's voice. That's a big responsibility. Lucas: It is. And it's one reason why I think this technology is going to separate the brands that understand their identity from the ones that just chase the next shiny tool. An agent that sounds like a generic chatbot might actually hurt retention. But an agent that sounds like your brand — helpful, warm, specific — that's a competitive moat. Luna: Alright, I'm convinced. I'm going to go check my moisturizer supply and see if Sephora's agent pings me. Lucas: Let me know if it does. And if we get enough listener questions about setting up your own agent, we'll do a follow-up with a step-by-step for smaller brands. Luna: That sounds like a plan. Thanks, Lucas. Lucas: Thanks, Luna.