• by Maarten Albarda , Featured Contributor, Yesterday

"The Association of National Advertisers Media Transparency Report: 10 Years Later” virtual conference held earlier this week was a sad celebration. The famous programmatic waterfall slide of 2016 showed that just 36% of ad dollars reached consumers, while the rest disappeared to middlemen, ad tech and fraud. In 2026, worst-case scenario, only 43.3% of your ad dollars actually reach consumers (roughly a 5% improvement).

And while that mess clearly persists (that level of improvement over 10 years is a sad record), last week I wrote about the industry potentially setting up similar plumbing into the world of AI advertising and autonomous AI agents. If you thought that tracking data leakages and agency markups in the old programmatic plumbing was a headache… brace yourself.

In traditional programmatic, we track things like impressions and viewability. In the new world of direct AI infrastructure and LLM search, the base unit of currency changes entirely. According to procurement pioneer Chad Johnson, enterprise AI pricing is currently scaling exactly like cloud compute, built on input tokens (what you ask), output tokens (what it generates), and reasoning tokens (the computing power it uses to "think" before responding).

Right now, vendors are rushing to embed these layers into enterprise applications and consumer-facing search tools. In the old programmatic waterfall, we at least have log-level data, the basis for the ANA waterfall chart. With large-language-model-driven advertising and autonomous agents, that level of granularity is virtually nonexistent.

If an AI agent has to execute a multistep workflow to serve a personalized ad to a consumer, it might ping three different application programming interfaces, cache a prompt, and burn through thousands of reasoning tokens just to deliver that one "ad."  If you think supply-side-platform markup is hard to audit, try calculating the hidden token margin buried inside a vendor's application fee.

Here are some of the problems this causes. First of all, advertisers will have zero visibility on financial arbitrage. When an AI agent recommends a product or serves an ad variation inside a chat interface, you cannot audit the vendor's invoice to see what the raw computing or inventory cost actually was. So determining if you paid a fair price becomes very difficult.

And because AI models are essentially a greedy algorithm that takes the fastest mathematical shortcut to achieve an outcome, the system could be designed to naturally favor paths that optimize the vendor's margin over the advertiser’s needs.

We cannot stop the shift toward AI-assisted search and agentic workflows, but we can take some protective steps.  Here’s what to add to your checklist when you are negotiating:

Demand an outcome-based commercial model: Never pay for AI on raw token counts or computing activity. If the vendors are selling an outcome (e.g., a qualified lead, a completed contract review, a validated consumer match), force them to price the contract per action completed.

Ensure your master services agreement explicitly states that the vendor cannot use your transactional data, supplier pricing, or campaign metrics to train their public models.

Insist on a human-in-the-loop framework. Never give an autonomous agent the authority to execute financial commitments or media buying decisions without a hard human checkpoint built into the workflow.

The digital advertising ecosystem spent a decade trying to clean up the programmatic supply chain -- and we gained a 5% improvement. We stand to lose it all within the next 12 months.