InterMedia, an independent agency focused on performance CTV, started running campaigns through AI agents a few months ago. The workflow connects Vibe.co's self-serve CTV platform to Claude's dashboard via MCP, then routes decisions through Olyzon to DSPs and SSPs across display and digital out-of-home. Campaign managers see reporting in the AI interface and issue commands based on what they find.
That's not a pitch deck. That's production.
According to AdExchanger's interview with InterMedia's SVP of Digital, the shift changes what "expert" means in media buying. Instead of knowing which buttons to push inside a platform, you become judicious about which changes to approve. The agent handles navigation; the human handles judgment.
For CMOs and CROs trying to model where this lands on the P&L, the question isn't whether agentic buying works. The question is how to structure risk, proof, and accountability when the agent moves faster than your measurement can validate.
The Stack That's Actually Running
Vibe.co built its Claude integration after noticing employees were already using ChatGPT and Claude for ad buying on other channels. The company eats its own cooking: Vibe uses its own platform to promote itself on CTV, so adding MCP support was a natural extension. More than 1,000 brands have adopted Vibe's MCP integration, and MCP-enabled buying now accounts for a significant share of spend on the platform.
Olyzon connects those CTV campaigns to broader programmatic infrastructure. The result is a workflow where an agent can view performance data, recommend changes, and execute across channels, all within a conversational interface.
This isn't the only stack in market. NBCUniversal, FreeWheel, Carat, and Newton announced a joint effort that pairs identity data from NBCU with audience intelligence from Dentsu to build custom buying agents. A luxury retailer is already using it. Carat is building agents tuned to specific KPIs: search lift, incremental reach, whatever the advertiser cares about.
Apostra has handled roughly $1.2 million in automated media buys over the past three months. PubMatic's AgenticOS facilitated a campaign for Rouge Care that spent $25,000 on CTV and drove over $125,000 in attributable sales, a fivefold ROAS. The Trade Desk is building a conversational interface that routes buyers to specialized agents. Magnite has a buyer agent and an orchestration layer spanning buy and sell sides.
The infrastructure exists. The spend is real, if modest. The governance frameworks are not.
The Proof Gap Problem
Here's where the CFO conversation gets uncomfortable.
Evgeny Popov laid out the math in AdExchanger: suppose a brand gives an agent $200,000 and the agent moves $1,000 an hour into a new CTV and retail media mix. The first sales readout that could judge that change takes 48 hours to mature. By then, $48,000 is committed to a decision nobody has assessed. And that first readout won't settle all $48,000 either. It produces an estimate, with error bars, about a change that's already two days old.
Popov calls this the proof gap: spend committed while the evidence needed to assess a change is still pending.
Campaign budgets cap total spend. Pacing sets the rate. Neither limits how much rides on a single untested change while the test is open. In an agentic system, the gap compounds because the agent changes bids, audiences, and channel mix simultaneously. The same weak guess drives spend across several budgets at once.
This isn't a technology problem. It's a governance problem. And it has a dollar figure attached.
Risk Budgets and Proof Contracts
Popov's framework distinguishes two instruments that most teams conflate: risk budgets and proof contracts.
A risk budget limits what an agent may commit under an unvalidated change before it must show agreed evidence or obtain fresh approval. An agent might get $12,000 to test a new mix and nothing more until the readout clears a threshold both sides set in advance. Three rules apply: the risk budget is separate from the campaign budget, it resets only when proof arrives, and the agent cannot reallocate from one test to another without human sign-off.
A proof contract assigns a defined loss to someone else when a defined part of the test goes wrong. This is the agentic version of sequential liability, the 1990s fight over whether an agency owed the publisher before the brand had paid the agency. That was credit risk. The agentic version is proof risk: who eats the loss if the outcome never arrives?
For finance teams, this means adding two line items to every agentic campaign: the maximum unvalidated exposure at any point in time, and the contractual allocation of shortfall if measurement fails to validate within the agreed window.
What Changes for the Buying Team
InterMedia's director of digital media and strategy described the shift as freeing up "time and brain power" for deeper, client-driven thinking. That's the optimistic read. The realistic read is that the job changes from platform operator to decision auditor.
The skills that matter now: understanding what the agent is optimizing toward, recognizing when its objective function diverges from business outcomes, and knowing when to override. If your team's expertise is "I know where the button is in the DSP," that expertise depreciates fast. If your team's expertise is "I know which changes are worth the proof gap risk," that expertise appreciates.
Gartner estimates that highly autonomous agentic deployments represent less than 5% of marketing use cases today, rising to only 15% by 2028. The near-term reality is agents handling execution while humans handle judgment. The question is whether your team is structured to provide that judgment, or whether they're still structured to push buttons.
The Board Conversation
If you're presenting agentic CTV to your board in the next quarter, here's the framing that survives scrutiny:
First, the technology is production-ready for specific use cases. MCP integrations exist. Spend is flowing. Early results show efficiency gains in campaign management and, in some cases, measurable ROAS improvement.
Second, the governance layer is immature. Risk budgets, proof contracts, and accountability frameworks are not standardized. Your team will need to build them, or accept undefined exposure.
Third, the skill mix shifts. You're not replacing headcount; you're changing what headcount does. Budget for retraining or expect capability gaps.
Fourth, measurement latency becomes a binding constraint. The agent can move faster than your attribution can validate. Either you accept the proof gap risk, or you throttle the agent to match your measurement cadence.
The promise of agentic CTV is outcomes-based buying at scale. The risk is that "outcomes-based" becomes "short-term signal chasing" because the agent optimizes on whatever feedback arrives fastest. Long-term brand building doesn't produce a 48-hour readout.
Model the proof gap. Set the risk budget. Define who holds the shortfall. Then let the agent run.
