Capacity or Capability: What AI Agents Do to Your Agency Bill
Agencies went to Cannes selling AI agents, but what those agents do to the agency bill itself?
At Cannes last month, agencies queued up to show off their AI agents. WPP demonstrated a buyer agent for premium video; Dentsu announced a partnership to test media-buying agents in the US; the independent Butler/Till built agents with DoubleVerify to police inventory quality; Dept began renting out the engineering scaffolding it uses to build agents of its own. Digiday read the scene as a turf war: agencies racing to build agents while clients use the same tools to pull work in-house and off the market.
Prima facie, the question writes itself: in-house or agency? Spend a budget answering it and you have answered the wrong one. The scene at Cannes was the agency bundle coming apart. For decades an agency relationship sold two quite different things under a single rate card, and AI is now pulling them in opposite directions. Get that distinction right and the in-house-versus-agency argument that consumes so many marketing offsites mostly dissolves. What replaces it is a sorting exercise you can run on your own contracts this quarter.
Focus On: What the Hyundai Pilot Proves
Start with the case the turf-war reading gets backwards. Hyundai ran a pilot with the AI firm Chalice and the supply-side platform OpenX, deploying custom bidding models inside a containerised system that prices each video impression in real time. The results were hard to argue with: a 67% reduction in CPMs and a 20% fall in the cost per high-value action (a dealership visit or similar) against the next-best alternative, with the savings ploughed back into media.
Here is the detail the headlines miss. Hyundai did not fire its agency. It orchestrated three parties — its in-house agency Canvas, the specialist Chalice, and OpenX — into one system, and the gain came from proprietary logic surfacing inventory the market had overlooked. Cheaper hands had nothing to do with it. “To have this agent on my behalf, doing evaluation of every impression,” said Sean Gilpin, Hyundai’s North America CMO, “that becomes a competitive advantage.” The advantage lived in the orchestration and the model. Placing the buys was the cheap part. Read correctly, the pilot is a story about capability winning and capacity losing, and the two had always sat on one invoice.
The Split
Capacity is hands and hours: trafficking campaigns, setting them up across systems, executing programmatic buys, versioning creative, building the weekly report, running routine optimisation. You rented it because keeping that much of it on the payroll at peak demand made no economic sense. Capability is the part you could not replicate at any headcount: orchestration across a fragmenting stack, governance and accountability for what runs, privileged access to inventory and data, strategic judgment, and the outside view an internal team structurally cannot hold.
AI agents accelerate the decline of one and the appreciation of the other. On capacity, the price is falling fast. “It takes hours to set up campaigns across systems,” Dentsu’s Caitlin Gelles told Digiday; store those nuances in a database, let an agent find the efficiencies, and the hours become minutes. Dept rebuilt a Swiss brand’s e-commerce platform 3.8 times faster with agents than without. Once setup and execution drop from days to minutes, the capacity you used to rent becomes a subscription line, and the arithmetic that justified outsourcing it inverts.
On capability, the price is rising, because a more agentic stack is a more fragmented one: more tools, more autonomous actors, more ways to be wrong at machine speed. WPP’s own framing gives the game away. Its bigger bet is the governance around the agent, captured in its own phrase: “humans at the helm, agents in the loop.” Dept productised the orchestration itself: “It’s about consistency at scale,” as its technology lead put it. The scarce thing is no longer who can execute. It is who can govern a swarm of agents and answer for what it does.
As I argued in Automate or Augment, the question for a single job is whether your daily tasks sit in the automate zone or the augment zone. The same cut works one level up, on the supplier. Capacity is the automate zone of the agency relationship; capability is the augment zone. AI is draining the first and concentrating value in the second.
Two Traps, One on Each Side of the Table
The agencies have a trap, and most are walking into it: competing on the half that is commoditising. Forrester and the 4As, in research released at Cannes, found that nine in 10 agencies use generative AI and half use agentic AI for execution, while 61% still treat AI as a “cost of business” and only 31% have any plan to monetise agentic AI within two years. They are automating their own billable hours and cannot charge for the agent that did it. “The industry is at risk of mistaking efficiency for effectiveness,” warned Forrester’s Jay Pattisall — the politest available way of saying agencies are making themselves cheaper at the precise moment they need to make themselves harder to replace.
Brands have the mirror-image trap: in-housing the easy part and meeting the hard part too late. Building a buyer agent is close to a weekend’s work now. Standing behind it is not. The World Federation of Advertisers found that 71% of in-house teams cite ethical uncertainty and 57% fear copyright or data misuse, while only around 17% have integrated AI into their workflows at all. In-house penetration has climbed past 80%, yet the governed, high-stakes work keeps getting handed back out. Which is why the in-housing cycle never quite completes. The agent is easy; the accountability is the moat.
The Sort
So stop arguing in-house versus agency. The argument is a category error — it treats a bundle as if it were one decision. Take your largest scope of work and label every line on it capacity or capability. The capacity lines (setup, trafficking, execution, versioning, reporting) you either bring in-house or reprice toward the cost of compute, because that is now what they cost. The capability lines (orchestration, governance, privileged access, the proprietary decisioning logic that won Hyundai its 67%, the judgment in the room) you keep and pay for properly. Lean on them harder, in fact; they are the part that just became more valuable.
Then renegotiate on a different unit of value: orchestration and outcomes, measured where heads and hours used to sit. The cost of leaving the line undrawn is already in the data. Forrester projects that roughly a third of all agency-managed media will be bought on a principal basis by 2026 — the agency reselling inventory at an undisclosed margin rather than buying on your behalf — and that 81% of US consumer-brand marketers intend to increase that spend. It is Lampedusa in a media plan: everything changes so that everything can stay the same. The agents are new; the undisclosed margin is the oldest line in the business.
I’ll put the prediction plainly: by the next budget cycle, “agency fees” will be the least defensible line in the marketing P&L — not because agencies are dying, but because half of what sits inside that number just became a seat licence, and the other half just became too important to buy on the old terms. The agency that keeps your business will be the one honest enough to tell you which of its own services you should stop paying it for. They might not start that conversation, so start it yourself, and sort the bill before procurement sorts it for you.
Follow me
That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.