Agencies are rebuilding how they bill for AI. Make sure your AI product is more than a token bill with your logo on it.

Short answer: Agency execs told Digiday on Sep 17, 2026 that AI costs are growing, Dept called them unpredictable, and Omnicom's CFO said the holdco is building a model to measure token costs and get reimbursed. If your AI product is metered usage with your logo on it, it is easy to replace. Write down the decisions your team makes with the tool and price those. (Digiday, Kimeko McCoy, Sep 17, 2026)

TLDR

• Agency AI bills now mix seats, tokens and staff time, and Dept says the cost has been unpredictable. (Digiday, Sep 17, 2026)
• Omnicom's CFO said the holdco is moving toward "a more outcome-driven model." (same)
• A resold tool is easy to replace. Name and price the judgment.

What does an agency AI bill look like now?

Digiday described three parts: tool subscriptions such as Claude and ChatGPT, tokens to use them, and the staff who run them. Markacy pays for enterprise Claude accounts for a team of 25-plus. Crispin uses around 100 enterprise AI tools and structures most deals as flat annual fees to avoid surprise monthly costs. Dept runs a monthly AI Tool Committee that approves, tests or drops tools. (Digiday, Sep 17, 2026)

Why are agencies worried about AI cost?

Because it moves. Dept's Jonathan Whiteside told Digiday "the cost has been unpredictable in the past year" and that they "just don't know how many tokens an individual might consume on a monthly basis." (Digiday, Sep 17, 2026)

A product priced on usage inherits that problem.

How are the big agencies changing billing?

Omnicom CFO Phil Angelastro said at Goldman Sachs' Communacopia conference that the holdco is building a model to measure token costs and get reimbursed, and added: "Ultimately, we're going to move towards a more outcome-driven model from a revenue perspective." One anonymous exec said their agency is testing billing tokens to clients as materials, separate from labor. (Digiday, Sep 17, 2026)

What should the AI product actually sell?

The decisions. Which outputs get thrown out, who checks a claim before it reaches the client's customer, and who takes the call when something goes wrong. Execs in the same article said human capital is still the biggest agency expense. That is where the value sits.

Dennis Yu, the Evolved Pros Ep 001 guest, put the checking problem bluntly on X on Oct 26, 2026, after catching an agent claiming its work had been verified: "All agents cheat, so your validation rules must be bulletproof, which is MUCH harder than it sounds." That is his opinion from his own work. (Dennis Yu on X, Oct 26, 2026: https://x.com/dennisyu/status/2114845765747298579)

The AI vendors are paying to stand out too. Anthropic's advertising spend rose from $33 million in 2024 to $147 million in 2025, per Digiday, citing the company's S-1 filing. (Digiday, Krystal Scanlon, Oct 27, 2026: https://digiday.com/marketing/anthropics-ad-bill-is-soaring-according-to-ipo-filing/)

How do I name and price the judgment?

1. List the five decisions your team makes on every AI deliverable.
2. Name the person who owns each one.
3. Write the review step as a line item on the proposal, separate from tool cost.
4. Put tool cost on its own line so a model price change doesn't hit your fee.
5. Tell the client what you will refuse to ship.

What happens if I don't?

The product stays a reseller line. The day the client finds the same tool for less, the line goes.

Next step

Who: the agency owner. What: write the five-decision list for your current AI offer. When: before your next proposal goes out. How: one page, one owner per decision, review step priced separately from tool cost.