Latest / The CMO Podcast with Fexingo: Marketing Leadership, Budgets, and Executive Strategy / How CMOs Use AI Agents for Real-Time Campaign Optimization
Transcript
- Lucas: So a few weeks ago, I was talking to the CMO of a direct to consumer brand called Olli Goods — they make premium home cleaning products, think refillable sprays and concentrated pods — and she mentioned something that made me stop. Luna: What was that? Lucas: She said they ran a seventy-two-hour flash sale last quarter and used something she called an 'AI agent' to reallocate their Facebook and TikTok budgets every fifteen minutes. Not a human tweaking bids — an autonomous system that decided which audiences got spend and which creatives got pulled. Luna: Every fifteen minutes? That's a pretty big leap from the standard 'set it and forget it' lookalike audiences most brands still use. Lucas: Exactly. And the results were striking — they boosted return on ad spend by thirty-four percent compared to their previous flash sale, while reducing wasted impressions by about twenty percent. But the interesting part isn't just the number — it's the shift in how the CMO thought about her role. Luna: Let me guess — she spent more time setting rules and less time clicking around in the ad manager. Lucas: Nailed it. She said her focus moved from 'optimizing the campaign' to 'designing the agent's boundaries.' What budget floors to set, which creative variants the agent could rotate, what frequency caps to enforce. The agent did the minute-by-minute work; she managed the constraints. Luna: So in practice, what did that look like? Walk me through the setup. Lucas: Sure. They used a platform from a startup called — appropriately — Agentic. Think of it as a middleware layer that plugs into the ad APIs. They gave it three objectives: maximize revenue from existing customers, acquire new customers at a cost per acquisition under thirty-five dollars, and avoid showing the same ad to anyone more than four times. Luna: And the agent just… went? Lucas: Within those guardrails, yes. It tested different combinations of audience segment and creative — like a 'kitchen cleaner' video versus a 'bathroom spray' static image — and shifted budget toward the combos that performed best in the last fifteen-minute window. By hour forty, it had essentially converged on two winning variations and was pouring most of the budget there. Luna: That sounds efficient, but also a little nerve-wracking. What if the agent had gone off the rails? Lucas: That's the key question. The CMO told me they had a human override — a 'kill switch' — plus a hard cap on daily spend per channel. But she also admitted that during the first six hours, the agent actually underperformed the manual benchmark. It took a while to learn. If you're a CMO with a quarter-end revenue target, do you have the patience to let it learn? Luna: That's the tension, right? The agent might be better over seventy-two hours, but if it's worse at hour six, someone's going to pull the plug. Lucas: Right. And this is where I think the conversation about AI agents in marketing needs to get more specific. A lot of people are still lumping them in with generative AI content tools like copywriters or image generators. But an agent is fundamentally different — it's not producing creative, it's making allocation decisions. Luna: So it's more like an autonomous trader than a ghostwriter. Lucas: Exactly. It's deciding where the money goes, which audiences get seen, and which messages get retired. That's a different risk profile. If a language model writes a mediocre product description, you lose a conversion. If an agent misallocates budget, you could burn through a quarter's ad spend in a weekend. Luna: Do you think most CMOs are ready for that? I mean, the Olli Goods story is a win, but they're a relatively small DTC brand. What about a Fortune 500 company with multiple brands and hundreds of campaigns running simultaneously? Lucas: That's the frontier. I've talked to a couple of larger CPG companies that are experimenting with agents, but only in limited categories or geographies. They're not handing over the whole portfolio. The pattern seems to be: start with a single campaign in a single market, learn the failure modes, then expand. Luna: What failure modes have come up? Lucas: One that keeps surfacing is the 'winner bias' problem. An agent that optimizes for short-term ROAS will eventually favor safe, proven audiences and exclude experimental ones. So you end up over-optimizing for today's conversion at the cost of tomorrow's brand growth. The CMO at Olli Goods actually noticed that — after twenty-four hours, the agent had almost completely stopped showing ads to cold audiences. Luna: So they had to manually inject a 'discovery budget'? Like a tax for exploration? Lucas: Exactly. They built it into the agent's objective function: at least fifteen percent of spend must go to audiences that have no purchase history. That's a fascinating example of how the human role shifts from optimizer to rule-maker. You're not pulling levers anymore; you're coding the constitution the agent operates under. Luna: I think that's a really powerful framing. And it raises an interesting question about the skills CMOs need going forward. If your job becomes designing constraints for autonomous systems, do you need to understand machine learning, or just clear strategic thinking? Lucas: I'd argue both, but more of the latter. You don't need to know gradient descent, but you do need to be precise about what you want. Vague objectives like 'grow revenue' will produce bad agent behavior. Sharp objectives like 'grow revenue from repeat customers while maintaining a minimum margin of forty percent' — that's something an agent can work with. And that's a marketing leadership skill, not a coding skill. Luna: So the CMO becomes a kind of constitutional designer. I like that. Lucas: It reminds me of something the Olli Goods CMO said. She told me: 'Before, I was a pilot. Now I'm an air traffic controller.' Her job is to set the flight paths and handle emergencies, not to manually steer every plane. Luna: That's a great line. And it lines up with something I've been reading about — there's a concept called 'human-in-the-loop' AI, but a lot of marketing tech vendors are starting to push 'human-on-the-loop' instead. Meaning the human oversees, but doesn't intervene unless something goes wrong. Lucas: Which brings us back to the trust question. How much autonomy are you comfortable giving an agent? For Olli Goods, the answer was 'a lot, but only for seventy-two hours.' They didn't let the agent run indefinitely. That time-boxing is probably a good practice for anyone starting out. Luna: And what about the creative side? If the agent is deciding which creative to use, does that change how creative teams work? Lucas: It does. At Olli Goods, they produced about twice as many creative variants as they normally would — twelve different videos and static images instead of six — because they knew the agent would test them and find the winners quickly. So the creative team wasn't replaced; they were asked to produce more variety upfront, with the understanding that only a few versions would ultimately get most of the budget. Luna: So the agency model might shift from 'produce one great campaign' to 'produce many candidates, then let the agent decide.' That changes the economics of creative production. Lucas: Absolutely. And I think that's going to be one of the bigger structural shifts in marketing over the next couple of years. The CMO's role isn't just about managing agencies and approving creative anymore — it's about designing systems that learn and adapt faster than any human can. Luna: Speaking of that faster pace — one thing that struck me about the Olli Goods case is that the agent made decisions every fifteen minutes. That's not a pace humans can sustain. So the CMO had to accept that she wouldn't see every decision. She had to trust the guardrails. Lucas: Right. And that takes a different kind of confidence. Most CMOs are used to being the final decision-maker on budget moves. Handing that to an algorithm — even one you designed — requires a leap. But if the alternative is missing out on a thirty-four percent improvement in efficiency, the leap starts to look worth it. Luna: Before we move on, I do want to flag something that I think is important context for our listeners. We're able to have these deep conversations about marketing strategy because of a small group of listeners who support the show directly. Lucas: That's true. And if these episodes have given you an idea you've actually used — a framework, a tool, a way of thinking — then consider chipping in at buy me a coffee dot com slash fexingo. It keeps the show ad-free and independent, and it literally funds more of these deep-dive conversations. Luna: Yeah, a few dollars from people who find value here makes a real difference. No pressure, just wanted to mention it. So — back to agents. Lucas: Let's talk about the risks more concretely. One thing that worries me is what happens when multiple brands in the same category use similar agents. If every cleaning brand has an agent optimizing for the same audiences on Facebook, you could get bidding wars that drive up costs for everyone. Luna: A tragedy of the commons, but automated. Lucas: Exactly. And agents are fast enough that they could escalate spending in minutes. The CMO at Olli Goods said she actually saw their cost per click jump during the sale — she suspects a competitor was running a similar sale with an agent. So the macro effects are still unknown. Luna: That's a systemic risk that no single CMO can solve on their own. But it does reinforce the need for those guardrails we talked about — budget caps, frequency caps, maybe even coordination with industry peers. Lucas: Right. And I think we'll see platforms like Facebook and TikTok start offering agent-specific features — maybe even 'agent detection' that shows you when other agents are bidding against you. It's an arms race that's just beginning. Luna: So where does this leave the average CMO who's listening right now? If you're not at a DTC brand with a seventy-two-hour flash sale, what's a practical first step? Lucas: I'd say pick a specific, time-bound campaign — a product launch, a seasonal push, something with a clear start and end. Define exactly two to three objectives with hard metrics. Set a small budget cap. Then let an agent run within those constraints for a week. Compare it to a control group where you manage manually. You'll learn more from that experiment than from reading a hundred articles about AI. Luna: That's solid advice. And if the agent underperforms, you've only risked a small budget and gained insight. If it outperforms, you've got a case for expanding. Lucas: Exactly. The Olli Goods CMO told me that their first agent experiment actually lost money. They had to tweak the objectives and add the exploration mandate before it worked. So patience and iteration are part of the process. Luna: I think that's the right note to end on. AI agents aren't magic — they're tools that need good design and good oversight. But when those are in place, they can do things humans simply can't, like reallocating budgets every fifteen minutes across a thousand audience segments. Lucas: And that's the future we're heading toward. The CMOs who learn to design good constitutions for their agents are going to have a serious advantage. It's not about being replaced — it's about upgrading your role.