Want AI Agents That Work? You've Gotta Do The 'Unsexy' Stuff First

Full analysis

Abbey Klaassen, 114 days into running independent performance agency Tinuiti, went on AdExchanger Talks with Allison Schiff and made a claim that cuts against the whole agentic-AI hype cycle: the agent is the easy part. The hard part is the plumbing underneath. Clean data, ETL, governance, workflow design. The unsexy stuff. And in a Basis-sponsored mid-roll, Mark McEachran explained the protocol that makes the agent layer possible in the first place. Both point the same direction, so let's take them together.

What's being decided here: nothing, formally. This is a practitioner conversation, not a deal or a launch. But it names a bet that every agency, DSP, SSP, and measurement vendor is now placing, whether they say it out loud or not: when AI agents can talk to every ad platform through one common connector, what's left worth paying for? That bet is easy to reverse for any one company. Hard to reverse for the industry, because once buyers standardize on a protocol, they don't un-standardize. The deadline is set by how fast Google, Meta, and Amazon ship working MCP servers, and per McEachern, they already have.

The Market Analyst. Klaassen is talking her own book, and it's a good book. Tinuiti built two-plus years of measurement infrastructure before the agentic wave, so of course she says the infrastructure is the moat. The interesting question is whether that moat holds against people with more money. Holdcos and integrated DSPs can buy or build clean data pipes faster than an independent agency can scale revenue to defend the pitch. MCP (the Model Context Protocol, the open connector that lets an AI agent talk to any ad platform in plain language) is the part that reshapes the board. It commoditizes the connection between buyer and platform. When the connection is free, the "we have the relationship" moat that agencies and managed-service platforms sell gets thinner. For a generalist: the pipes between the buyer and Google or Meta used to be custom and expensive; now they're becoming a standard part anyone can plug into.

The Skeptic. Steelman the case against Klaassen. If clean data is the real differentiator, why hasn't clean data already won? Every agency has claimed a proprietary data stack for fifteen years. Most were dashboards with a login. The Institute for Real Growth stat she leans on, eight in ten leaders say marketing could create more value but only one in five say that value gets captured, comes from a group whose name announces its conclusion, with no methodology in the transcript. Treat it as marketing, not evidence. And the "holistic ROAS" pitch, combining several marketing-mix models with brand-equity modeling into one de-duplicated number, is exactly the kind of black box a CMO with shrinking tenure can't audit. Digiday's own data puts median CMO time in office down a third over sixteen years. Short-tenure CMOs buy simple stories, not de-duplicated enterprise value models.

The Operator. Try to run this Tuesday morning. MCP servers from the walled gardens are read-mostly today. McEachern said it plainly: letting one agent shift budget across platforms on its own is where the industry says it's heading, not standard practice. So what actually breaks first is trust, not connectivity. The moment an agent has its hands on the wheel, moving spend between Google and Meta without a human clicking approve, someone has to own the guardrails. That's a governance job, and governance is where 90-day surprises live. An agent that reallocates budget on a bad signal at 2am doesn't get caught until the reporting lands. The plumbing Klaassen sells is the same plumbing that decides whether the agent does something stupid at scale.

The Customer / End User. Two customers here. The CMO wants fewer vendors and one honest number. Klaassen's integrator pitch, one agency to orchestrate the dozens of measurement partners, creative shops, and retail media networks, is a real pain point, not invented. Vendor overload has hard fee costs and hidden management overhead. But the second customer, the retail media networks, should read this as a warning. If agencies build de-duplicated cross-channel ROAS in-house, RMN self-reported numbers stop being the scoreboard. Albertsons Media Collective already leans on incrementality; the ones that don't will get graded against a number they don't control.

The CFO. The MadTech survey says 58% of ad execs expect agentic buying to scale within a year. That's the number that governs spend decisions right now, and it's an expectation, not a result. If you're funding an agent program on that expectation, the payback depends entirely on whether the data underneath is clean enough to trust the agent with money. Building that foundation is the expensive, invisible line item that no board wants to fund because it demos badly. The vendors selling "fully agentic end-to-end everything," Klaassen's own pick for most overhyped term, are selling the cheap, visible part. The bill for the boring part comes due either way.

Where the council splits. Klaassen and the Skeptic disagree on whether clean data is a durable moat or just this decade's version of a proprietary-dashboard claim. The Market Analyst thinks the moat is real but stealable by richer players. The Operator and the CFO agree on the thing nobody's funding: governance for agents that move money is where this actually gets hard, and it's precisely the work that demos poorly and gets deferred.

What it hinges on. One belief: that MCP standardization is real and moving fast. If it is, the connection between buyer and platform stops being a moat for anyone, and value migrates to whoever controls clean data and trustworthy guardrails. Google, Meta, and Amazon have shipped servers. The protocol sits under the Linux Foundation. This is not vaporware. The council leans toward Klaassen being directionally right on the plumbing and oversold on Tinuity's ability to own it alone.

What to de-risk before betting on it. Anyone building an agentic buying stack should test one thing before scaling: can the agent move budget across two platforms and can you catch a bad reallocation before it costs real money? If the answer is no, you've bought the agent and skipped the foundation.

Prediction: By the close of the 2027 upfront negotiations in June 2027, no top-10 US advertiser will let an AI agent shift budget between Google, Meta, and Amazon fully autonomously without a human approval step in the loop.

Confidence: Medium. The walled gardens' MCP servers are read-mostly today and governance lags the pipes.

Why: McEachern of Basis said directly that agents moving budget across platforms on their own is directional, not standard practice, and that the moment they get "hands on the wheel," guardrails become the whole conversation. The 58% of execs in the MadTech survey expecting agentic buying to scale within a year are describing optimization and assistance, not unsupervised money movement between rival walled gardens who have every incentive to bias an agent toward their own inventory. Large advertisers with shrinking CMO tenures and audit exposure will not hand three competitors an unsupervised budget dial that fast; the opposite outcome would require both the protocol to mature past read-only and enterprise legal and finance teams to accept autonomous cross-platform spend inside eighteen months, which is faster than any prior buying-automation shift has moved.

Revisit by 2027-06-30: We're right if no top-10 US advertiser has publicly deployed fully autonomous cross-platform budget reallocation across Google, Meta, and Amazon without human sign-off by the 2027 upfronts. We're wrong if at least one has.