Whose Agent Is It, Anyway?

Here's the thesis, up front: in the agent era, advertising stops being a placement problem and becomes a loyalty problem. The two labs racing hardest toward AGI just staked out opposite positions on it — and the split tells you more about where AI is headed than either company's roadmap does.

At the 2026 Super Bowl, Anthropic spent some of the most expensive airtime on the planet on six words: "Ads are coming to AI. But not to Claude." Not a tagline. A bet.

Around the same time, OpenAI moved the other way — piloting ads on ChatGPT's free and Go tiers, then expanding fast, Korea included.

Same industry, same moment, opposite calls. Underneath both is one question: what happens to advertising once AI stops showing you options and starts handing you an answer?

Search Showed You Everything.
AI Shows You One Answer.

Search never hid the comparison. You searched, you got a results page, you did the work of choosing. A sponsored link might sit at the top — but you always knew what you were weighing against what.

Agentic AI collapses that whole process into a single output. "Find the right insurance for my family." "What's the best ETF to buy right now." You're no longer asking for options. You're asking for a verdict.

That collapse is the opening commercial interest has been waiting for.

The Agent That Inherited the Principal's Chair

Economists have a name for this setup: the principal-agent problem. It shows up whenever the party doing the work knows more than the party who hired them — and their interests quietly diverge. Hand your judgment to an AI, and you've built that dynamic by design. Except this agent also answers to its own shareholders.

Somewhere behind every model sit developers and advertisers with incentives of their own. An AI that looks like it's working for you may, structurally, be working for them.

The moment AI stops retrieving and starts deciding, trust stops being about information and starts being about the agent itself. Which makes the operative question, going forward, not how smart is this model — but whose interests was it built to serve.

The Ad Isn't in the Results Anymore.
It's in the Reasoning.

Search-era advertising competed for a slot on the page. Sponsored links were labeled, and the user still scanned, compared, decided. The ad could shape the information. The judgment stayed human.

Agentic AI moves the intervention upstream — into the reasoning itself. A product surfaces first, "naturally." One true fact gets emphasized over another, equally true, fact. None of it needs to look like advertising. All of it can move a decision. And almost none of it gets noticed.

Princeton put numbers on this. Researchers had people choose an e-book using either a search engine or a conversational AI. In both, 20% of listings were secretly marked "sponsored." Search intervened by ranking sponsored items higher. The AI went further — actively talking up sponsored products while quietly undercutting the rest.

The gap: 22.4% sponsored-pick rate for search. 61.2% for the persuasion-tuned AI. Slap a "Sponsored" label on it and warn users up front, and the number barely moves — 55.5%. Design the model to mask its intent, and detection drops below 10%.

Conversational AI isn't a modestly better sales channel than search. It's a categorically different one.

The Delegation Spectrum

Not all AI use is equally exposed. What determines the risk is exactly one variable: how much judgment you hand over.

Specify your own criteria — Google Flights, "morning departure, nonstop, under $230" — and AI has almost nowhere to intervene. Let it compare for you, Amazon Rufus-style, and it can shape what you notice first, though the final call is still yours. Ask it to recommend, and you stop comparing altogether — its one answer becomes the only candidate. Let it execute — Buy for Me, discretionary AI investing — and you're not in the loop at all. You're reading a receipt.

Each step right on that spectrum is a step further from human judgment and a step closer to whatever incentives the model was trained on.

Disclosure Has a Ceiling

The obvious fix is disclosure — label the ad, name the advertiser. It's worked for two decades in search, because Google's structure made "organic" and "paid" cleanly separable.

It doesn't survive the agent. An AI answer fuses recommendation, summary, comparison, and a call to action into one seamless output. There's no page to split into two columns. The more judgment you've delegated, the less a "Sponsored" tag buried in the footer actually buys you.

We've Seen This Movie Before

Recommendation engines got here first. YouTube and Instagram started surfacing content nobody searched for, and ads migrated from the results page straight into the algorithm. Platforms called it "content you'll like." How much of that ranking was paid placement versus genuine relevance was — and still is — invisible from outside. Regulators eventually stopped asking for ad labels and started asking for algorithmic transparency instead.

Influencer marketing showed the flip side. Mandatory sponsorship disclosure didn't dent trust — audiences who already liked a creator kept buying the recommendation anyway. Disclosure read as a badge of influence, not a warning label.

AI has none of the influencer's tells. A creator's biases surface over time — you learn their leanings, their sponsors, their blind spots. AI answers every question in the same confident, consistent voice, forever. You have far less to go on.

Labeling, on its own, was never going to be the fix.

AI Needs a Duty of Loyalty

This has already moved past theory. The Consumer Reports Innovation Lab and Stanford's Digital Economy Lab, through their Loyal Agents project, are arguing that AI agents should carry the same duty of loyalty the law imposes on human agents.

A lawyer or an asset manager is legally required to put the client first, conflicts or not. An AI agent acts on your behalf constantly — and no such standard exists for it yet.

That's the real shift in this debate: AI advertising stops being a labeling question and becomes a fiduciary one.

What Insurance Already Solved

Insurance ran this experiment already. Agents earn commissions that vary by product — a built-in incentive to sell what pays best, not what fits the client. Textbook principal-agent conflict.

The first fix was disclosure — publish the commission, let the client see the number. It didn't work. A disclosed number doesn't resolve a conflict; it just teaches everyone how to route around it.

What actually worked was restructuring the incentive. Commissions got spread out over up to seven years, so an agent only gets paid in full if the client stays on the policy. The reward moved from volume sold to client retained — which is another way of saying it moved from the agent's interest to the client's.

AI advertising is headed for the same fix. The lever was never the size of the "Sponsored" label. It's whether an advertiser can pay for rank, and how tightly an AI company's revenue is coupled to what's actually good for the user. Regulation will follow the incentive, not the disclaimer.

Where Regulation Goes Next

Regulation never arrives finished. It solves the easiest problem first, hits a wall, and moves the target. AI advertising will take the same three steps.

Disclosure comes first — flag the ad, name the advertiser. It does real work as long as users are still comparing and deciding for themselves.

It stops working the moment AI starts choosing for you. So the target shifts to recommendation structure — can a sponsor pay for rank, can performance-based deals move a result. Same move insurance regulation already made, on a different substrate.

Once AI starts executing — buying, investing, committing money — the target shifts again, to delegated authorityitself: mandatory approval before autonomous action, limits on performance-linked contracts, hard boundaries on how much a model is allowed to decide without asking.

Disclosure, then structure, then authority. Search-era regulation policed information. This era polices judgment.

The Real Competition

Which is the answer to the opening question. Anthropic and OpenAI didn't just make different calls on ad formats. They made different bets on what kind of agent they want to be. Search was a race to surface information fastest. This is a race to become the agent people actually trust with a decision.

That's the same test every AI startup is now being run through, whether they've noticed or not. Model quality won't be the moat much longer. What will matter is how much judgment users are willing to hand you — and how confident they are about whose interest that judgment serves once they do.

So the work, right now, is concrete: build recommendation logic that can explain itself under pressure, and go find the exact point where your revenue model and your user's interest quietly diverge — before a regulator, or a competitor, finds it for you.

Trust was never going to be a compliance line item. In this era, it's the moat.

About Kakao Ventures

Founded in 2012 and backed by Kakao — Korea's leading tech platform — Kakao Ventures is one of Korea's most active Seed-stage venture capital firms, with approximately $280M USD in AUM. We partner with founders before the path is fully defined, when conviction in people matters more than proof in numbers.

Our portfolio includes Lunit (AI cancer diagnostics), Rebellions (AI semiconductors), and Dunamu (operator of Upbit, one of Asia's largest crypto exchanges).

If you're building at the edge of what's possible — we'd like to hear from you.