Put AI Tokens on the Materials Line Before You Rewrite the MSA

The SOW looked clean. The token burn did not.

Seats, tokens, and people are three different bills. Digiday's Kimeko McCoy put the anatomy in print on September 17, 2026. If an exec got a regular bill for third-party AI tool costs, it would show subscriptions for tools like Claude, ChatGPT, and Google's AI suite, tokens to use those tools, and the people running them.

Most agencies Digiday spoke with still carve AI spend out of existing tech budgets rather than a new P&L line. Human capital remains the dominant expense. Tokens are the unpredictable middle.

I have bought enough agency work to know where overages hide. If the SOW has no token line, you will eat the spike or the agency will bury it. That is the invoice problem. It is not the same as rewriting the MSA for AI-rewritten work. Contract language is one fight. The materials line is how the month actually closes.

Markacy pays for enterprise Claude for a team of 25-plus, Digiday reported, plus ChatGPT for some staff who request it. Dept runs Claude, ChatGPT, and Google because client preferences vary. Jonathan Whiteside, global EVP of technology at Dept, told Digiday those three drive the biggest token costs, and that token burn is hard to predict because nobody knows how many tokens an individual might consume in a month. Dept runs an AI Tool Committee to approve, test, drop, or certify tools, and is building systems to help staff pick models for a task.

Omnicom CFO Phil Angelastro, speaking at Goldman Sachs' Communacopia conference, said the holdco is building a business model to manage token costs, measure them, and get reimbursed, moving toward more outcome-driven revenue over time. Digiday also reported at least one agency executive, speaking anonymously, whose shop is experimenting with billing AI token costs to clients as materials, separate from pre-set human labor. Another experiment embeds token cost into the human cost of work.

So what for multilocation operators buying agency work. Ask where tokens live on the invoice before you approve the next scope. The fire is not whether the agency uses AI. The fire is whether unpredictable metered usage shows up as a line you can see.

Ramp research Digiday cited said about 31% of companies spend more than $10,000 per month on AI, a signal that AI is becoming a formal budget line. Digiday's agency execs declined specific spend figures. I am leaving their numbers off the table.

So what again. The emotional outcome is whether you can sit in a quarterly review and explain an AI cost without a shrug. Transparency builds trust. A buried token spike kills it.

So what a third time. Opaque token burn becomes a relationship tax. Either you eat the surprise or you train the vendor that mystery fees are fine.

Cost of inaction is concrete. Leave tokens off the materials conversation and you will either surprise-finance an overage or fund opacity. The operator who puts tokens on materials early will negotiate cleaner scopes while peers still argue about MSA poetry.

Every sales conversation comes down to one thing: trust. If your prospect trusts you, they buy. If they do not, fear wins and the deal dies. An invoice that cannot explain token burn fails that test.

Keep it simple. Define your customers. Define your solutions. Trust the data you can stand on. The data is seats, tokens, and people as separate cost shapes, plus holdcos already building reimbursement rails.

On my side of the table I still own pay. I will not approve a clean labor SOW that leaves metered AI as a mystery fee.

Bring this into the Evolved Pros community this week. Who on your vendor list can show token burn today. What line on the invoice would catch an overage before month-end. When you next renew an agency SOW, how do you put AI tokens on the materials line before you reopen the MSA.

The Bottom Line

Put AI tokens on the materials line before you rewrite the MSA. Digiday dated the agency AI bill anatomy September 17, 2026. Seats plus tokens plus people. Omnicom's Angelastro is building measurement and reimbursement toward outcome-driven revenue. At least one shop is already testing tokens as materials separate from labor.

If the week produced a clean MSA rewrite and a messy token surprise, you fixed the contract and ignored the invoice.