As Marketers build new AI brains, whose interests are being encoded?

I spent a lot of my twenty-seven-year career inside the holding-company world, latterly running EssenceMediacom North, a WPP agency and the largest media operation outside London. I owe a lot to that world and to the people still in it so I have some “skin in the game” here as they say across the pond.     The Sony-WPP case has brought the client-agency relationship sharply into focus, but what’s in danger of being overlooked in this furore is how the relationship between advertisers and the firms they trust with their budgets is being rebuilt right now, mostly in software.     So while the headlines focus on the past, my priority (and that of the whole team at TAU) is to ensure the decisions made by brands and agencies today set the industry on a path to a bright future for agencies and clients.

Marketing has always relied on agents acting on a business's behalf, with traditional agencies the obvious example. The term itself is exact: an agent is meant to act in your interest, not its own. What’s emerged over the last year is that a second kind of agent, the software sort, is being handed the same decisions. The rules that govern billions in ad spend are being written into AI agents.

While most of the discussion about this shift is about speed and efficiency, the urgent question is are either kind of agent working for the business, or for themselves?

This urgency comes in two forms: a challenge to address, and an opportunity to take.

The challenge

Every algorithm optimises for something. Once an incentive is written into code, whatever it optimises for runs on its own, at scale, and keeps running until someone thinks to look under the hood.

The specifics of the WPP and Sony case are contested, and WPP denies them. But the principles it exposes hold for almost every enterprise relationship with a holdco agency, and they will still hold long after this news cycle passes.

The shift raises the stakes because AI agents are moving from reporting on decisions to making them. Give these systems clean, granular, honest data and they improve with every cycle. Feed them a plan that hides its own incentives, and they won't fix the problem; they'll run that flawed logic faster and more consistently than any human could. The danger to guard against isn't a system behaving unpredictably. It's a highly capable system doing exactly what it was built to do, inside a structure whose real incentives the advertiser cannot see.    Najoh Tita-Reid put this well on the Uncensored CMO podcast when she explained how (poorly implemented) AI can become a "colonizer of inefficiency". She describes how AI rarely shows up as a takeover. It starts as a pilot: “we're only automating the reporting, nothing to worry about.” Then the reporting starts to shape the business strategy, the strategy shapes the team and the organisation’s structure.

That's why transparency and control at all points are so critical. Hand reporting or planning to a system you can't see into and you've given a third party a foothold in decisions that were yours. Transparency and control are what keep you in the driving seat.

The Sony case offers a second lesson worth holding onto: the standard checks the industry trusts to keep everyone honest could miss the real problem. It took a client running its own forensic investigation to find what the routine checks missed. In a human system that blind spot is expensive. In an agentic one it gets built into the machinery, and the software inherits the very layer nobody was inspecting.

This creates a cost that compounds. The data your campaigns throw off is the training material your own AI needs. If the planning, buying and decision-making all live inside someone else's system, that intelligence gathers there rather than with you. You get faster execution, but you don't get smarter in any way you own, and each cycle deepens the dependency rather than building your own capability.

The opportunity

Beneath the warning signs is a more encouraging reality: the window to settle these questions is unusually open right now.

The leverage comes from a structural rebalancing of the agency market. As AI opens up access to data and scenario planning, it removes the need for the deep analyst hierarchies holding companies traditionally charged for. At the same time, as TAU sets out here, as ad spend concentrates inside the tech hyperscalers whose algorithms set their own terms, the buying scale of legacy agencies no longer carries the weight it once did.

Faced with those headwinds, and with AI compressing days of work into minutes, the billable hour is losing its footing. Holding companies are being pushed towards output- and outcome-based pricing to keep up. Together these forces open a real chance for brands to reopen and renegotiate terms, and contracts like these are rarely reopened.

If you're paying for an output, you're the one who gets to define it: the outcome, the baseline, the evidence, the data you keep hold of, and what happens when a standard slips. And because those same terms are about to be written into the systems doing the work, the contract is where you decide what gets encoded. Set them while the agreement is open, or accept whatever the software comes to treat as normal. This is the moment to read the agency agreement again with the next few years in mind, not the last few.

There is a catch. An outcome can be defined in ways that flatter whoever is measuring it. As TAU Marketing Solutions CEO Robert Webster sets out in If in doubt, verify, a recorded sale and a caused sale are not the same thing, and a surface number can pass for proof without being it. So outcome-based buying only works in your favour if you scrutinise the data beneath the headline; otherwise you're giving an agency, or an algorithm, permission to grade its own homework.

None of this is an argument against working with agencies. The strongest partners, including the independent agencies we work alongside, are already building this level of transparency, because they can see it is how the relationship holds up once AI takes on more of the execution. A good agency has little to hide here and a great deal to gain from proving its value in full daylight. I believe that the agencies that will thrive in this new era will spend more time on a client's growth and business strategy as they don't need to spend 80% of their time making the trains run on time.

The pattern I've seen doesn't change: the brands that come through a shift like this in control are the ones that treat transparency as something to own, not something to check once a year. We'll publish the practical side of that shortly: the detail a plan should carry, the service levels worth putting around an outcome deal, and the intelligence a team should keep rather than rent. And if you're building a 2027 plan and want an independent read of it before you sign, that's a conversation I'm glad to have.

The agentic buying stack is being assembled right now. Someone's incentives are being written into its logic, someone's data is training it, and someone will own the intelligence it produces. The plans you sign for 2027 are the first lessons those systems learn. The only question that can't wait is whose side the agent is on when they do.