At CES 2026, Dentsu’s Chief Trading Officer for North America made something explicit that the rest of the industry has been dancing around: agentic AI systems are now making campaign optimization decisions at a speed and scale no human trader can match. Line-item optimizations happening in milliseconds — something, in his words, “a living human being is not capable of doing today or probably ever.”
Every major holding company is building toward this. Dentsu has already overhauled its flagship platform, “Dentsu.Connect,” to automate marketing and media workflows between brands, agencies, and tech partners. The others aren’t far behind.
For performance marketers, this should set off an alarm. Not because AI is bad at optimization. Because the entity building, training, and operating that algorithm is the same entity that earns differential margins, volume rebates, and inventory commitments across the channels it recommends. That conflict of interest has always existed. Agentic AI just automates it and removes the last human review layer that was catching it.
The Conflict Isn’t New. The Speed Is.
The holding-company transparency problem is well-documented. The 2016 K2 Intelligence report commissioned by the ANA found that non-transparent practices (including undisclosed cash rebates to media agencies) were pervasive across the U.S. media buying ecosystem. Agencies were earning volume-based incentives from media owners that clients didn’t know about, and those incentives shaped which channels got budget.
Nothing about that incentive structure has changed. What’s changed is the mechanism.
When a human planner built a channel mix, it moved at human speed. A brand’s media team could review the rationale. Finance could benchmark CPMs. Someone could ask hard questions. The plan was auditable because it was legible.
Agentic AI eliminates that legibility. When an autonomous system shifts $200,000 from direct mail into a CTV inventory pool where the holding company has a volume commitment, that decision happens in milliseconds. It’s justified by an objective function the advertiser didn’t write, can’t inspect, and has no right to audit.
You don’t need to allege bad faith to have a serious problem. A model trained on historical agency data encodes whatever biases exist in that data. If the agency historically favored channels with higher margins, the model learns that as “optimal.” The AI isn’t conspiring — it’s pattern-matching. At a speed and volume that makes oversight impossible.
Your Only Defense Is Measurement You Control
Better contract language helps. Demanding algorithm access won’t work because no holding company will provide it. The only real defense is building measurement infrastructure that produces ground-truth conversion data independent of your agency’s systems entirely.
This is where holdout-based incrementality testing becomes non-negotiable.
In programmatic direct mail, the holdout methodology is structurally clean. You define a universe of households, withhold the mail piece from a randomized control group, and measure the conversion difference at the household level. No cross-device contamination. No impression-counting discrepancies. No probabilistic inference. The consumer either got the mail piece or they didn’t.
Matchback attribution adds the second layer. When an agency’s AI recommends pulling budget from direct mail in favor of a channel where the holding company has a supply-path advantage, matchback data shows you exactly what that reallocation costs in actual conversions — not modeled conversions, not platform-reported conversions, but real transactions matched to real households.
Together, these two methods create a measurement ledger the agency’s AI can’t touch. The algorithm’s objective function doesn’t matter. The rebate structure informing its training data doesn’t matter. You have your own data, tied to your own conversion events, at the household level.
Direct mail is the channel where this methodology is cleanest. Which is also why it tends to be the first channel an agency-side algorithm deprioritizes when optimizing for metrics that happen to favor higher-margin inventory.
Five Steps to Get Ahead of This
1. Establish holdout baselines now, before the AI touches your budget. If you don’t have pre-AI incrementality data for direct mail, you have no benchmark to measure against once the agency’s system starts reallocating. Withhold mail from 10–15% of your qualified audience. Measure conversion rates, AOV, and LTV at 30, 60, and 90 days. Do this before the algorithm has a chance to shift your mix.
2. Build your own channel-level ROAS reporting. Agentic AI systems produce their own dashboards, optimized to show whatever the algorithm was trained to optimize for. If the objective function prioritizes reach or cost-efficiency over true incremental ROAS, the numbers will look great while your acquisition economics quietly deteriorate. Build your own calculation from your own transaction data.
3. Require a direct mail control cell in every mixed-media test. When the agency AI proposes shifting budget away from direct mail, insist on a structured test: one segment gets the recommended mix, one holds direct mail at prior levels. Measure both against the same conversion events using your attribution stack. Any agency confident in its AI’s recommendations should have no problem with this. Resistance to testing is a signal.
4. Keep your identity resolution off the agency’s stack. Dentsu operates Merkury for Media. Publicis operates Epsilon. Omnicom runs Omni. When the entity resolving your audience identity also allocates your budget and reports your attribution, every layer of the measurement stack has a conflict. For direct mail, CRM-based identity resolution (matched against USPS NCOA data and third-party enrichment providers) gives you an audience layer the agency doesn’t control. Postie’s CRM activation connects directly to your first-party data and keeps that resolution on your side of the ledger.
5. Contractually require disclosure of channel-level margin differentials. Before autonomous systems are making real-time decisions with your budget, get on paper where the holding company earns differential economics: principal trading positions, volume rebates, preferred inventory commitments, supply-path fees. You won’t get the algorithm. But you can get the financial incentives that shaped what the algorithm was trained on — and that, combined with your own measurement data, gives you an audit trail.
The Window Is Narrowing
The shift to agentic media buying isn’t coming. It’s happening now. The brands that establish independent measurement infrastructure now will have baselines to audit against. The brands that wait will be trusting an algorithm they can’t see, built by a company that profits from the recommendations it makes, running at a speed no human can follow.
Holdout-based incrementality testing and matchback attribution aren’t just direct mail tools. They’re the only measurement methodology that produces data the agency’s AI genuinely can’t influence. That’s what makes them the right foundation for any brand trying to stay accountable in an agentic media environment.
[See how Postie’s holdout testing and matchback attribution give you measurement independence from your agency’s AI →]
