“Platform AI will absorb the agency’s job.”

Earlier this month, Shekhar Deshpande, Head of Global Clients and Agency Marketing at Meta, put that line in a sponsored Adweek post vaguely titled Measurement in the Age of AI. Despite its unexciting title, Deshpande used the back half of the piece to opine on the role of the media agency in an AI-first world. It’s rich to hear the person responsible for marketing Meta to agencies opine on the value of the agency when his boss’s vision is to press a button and get an outcome from Meta. Deshpande’s perspective is helpful. It just doesn’t match reality today.

He concedes the trope is partly true. The unexamined premise underneath it is that any performance lever an agency manually operates will become automated with AI. Meta is clearly at the forefront of this, and Deshpande says so bluntly: Advantage+, its AI-enabled solution, gives no advertiser an edge over any other. He’s right about that. There’s no alpha in easy buttons. But relying on platform tools like Advantage+, PMax on Google, or Performance+ campaigns on Amazon solely for execution is not the reality of agencies at the enterprise today. For some media channels, like the Amazon DSP, we see first-hand that AI automation, whether in-platform or via third-party agentic AI, is still in its early innings. We can all agree that a near-term future state may exist where low-level performance levers are totally abstracted away. Leveraging AI to execute this at scale is not the reality on the field today.

So how do we move past this point? And how do we find a healthy place where platform AI doesn’t absorb the agency, but where platform AI and third-party agents are the foundational building blocks an agency assembles to achieve a client’s goals?

Three stages, in order: workflow, guided automation, and recursive learning loops.

Workflow first. An agency can’t rely on platform AI alone, so it has to rewrite its best practices in a manner totally unconstrained by human labour. Concretely, that means a procedure an agent can execute: how campaigns get named, how line items get built, how budgets get paced, how bids get adjusted, and in what order. If it can’t be written down as steps, an agent can’t run it.

Once rewritten, there needs to be a very real conversation at the agency leadership level about the types of actions agents can automate and the types of actions that require human review. Which bid changes run unattended. Which budget shifts get surfaced for a human to accept. And the rule for promoting an action from the second bucket to the first once the agent has proven itself on it. Automation needs to be relentlessly pursued. But it is earned one action class at a time, not granted on day one.

Then the loop. Today, most agentic tasks answer one question: how are we going to achieve this objective? And they only answer it because a human wrote the answer in by hand. A recursive learning loop is a task that also answers what happened once we took action, and what we should do next, and then uses those answers to rewrite the how. The task learns. Self-improving agentic autonomy across a set of defined goals on an advertiser is the bar agencies need to meet as we look to 2027.

This is the foundation that Deshpande misses. We’re not at this stage yet. We’re 18-24 months away from it. Admittedly, we’re only enabling this now for our customers, and despite our growth, we power a small fraction of global enterprise media. Many agencies still perceive their agentic AI bets to be hooking up an MCP into legacy tech to automate reporting and suggest actions. That’s stage zero. The share of agencies still laboriously pressing buttons and uploading CSVs into platforms is shockingly high. That this still exists highlights the remarkable opportunity for the agencies that choose to lean in now and make these agentic bets.

Once this foundation is met, Deshpande is absolutely right about where agencies should invest: validating and evaluating what the agents did, governing them across platforms, and designing the data foundations underneath them, all of which require platform-neutral expertise and pattern recognition across clients. Through these investments, “agency value migrates to deciding which outcomes are worth optimizing,” verifying the automation, and assembling the stack across platforms.

He’s right. But he’s describing the agency of 2028, one that has already automated the floor and now competes on judgment. The road there runs through workflow, guided automation, and loops, in that order, and most agencies haven’t taken the first step. The ones competing on judgment in 2028 are the ones that start building the foundation in 2026. That’s the road we build at Gigi, one channel at a time.