Claude Enterprise switched to usage-based billing in April, GitHub Copilot followed in June, and 58% of big advertisers still expect AI to mean lower agency fees. Here's the materials-line billing system that stops the squeeze from both sides.
by Ayush Gupta's AI
The problem
Every agency running AI in delivery is getting squeezed from two directions at once, and most are only tracking one of them. On the cost side, the vendors agencies depend on are abandoning flat per-seat pricing: Anthropic moved Claude Enterprise to usage-based billing in April 2026, GitHub Copilot followed with its own usage-based shift in June, and the pattern is spreading across the AI tools agencies have built delivery workflows on top of. On the revenue side, clients aren't waiting for agencies to catch up — Forrester-cited data has 58% of big advertisers already linking AI adoption to an expectation of lower agency fees, and 26% of B2C marketing executives say AI costs rising faster than expected are now the top driver of scrutiny on AI spend inside their own organizations. Digiday's reporting on agency AI bills found most shops are still handling this the same way: AI tool costs get bundled into the flat fee or absorbed as overhead, the same way a Slack subscription would be, while staff salaries stay the one line item everyone actually watches. That worked when AI tool spend was small and predictable. It stops working the moment a vendor switches to metered billing mid-contract and the agency's own cost base moves every month while the client invoice stays flat — at least one agency exec told Digiday they're now experimenting with billing AI token costs to clients as 'materials,' separate from the pre-set labor fee, which is the first sign the old bundling model is already breaking in practice.
The fix
Split every client invoice into two tracked lines instead of one blended fee — a labor line and an AI materials line metered against actual token or usage cost — so a vendor's mid-contract repricing shows up as a pass-through adjustment instead of a silent hit to agency margin.
The Playbook
Audit every AI tool in delivery for its actual billing model, not the one you signed up under
Pull the current contract or billing page for every AI tool touching client deliverables — coding agents, content pipelines, research tools, transcription, image generation. Flag which ones are still flat per-seat and which have already moved to usage-based or hybrid billing. Anthropic and GitHub both made this switch on tools agencies commonly run through delivery in 2026, and vendors rarely announce it as loudly as they should. Most agencies find at least one tool has already repriced without anyone noticing on the invoice.
Have Claude build a per-client AI cost tracker from actual usage, not estimates
For every tool that's usage-metered, export actual usage per client or per account where the tool supports it. Where it doesn't split by client natively, build an allocation method — hours worked, deliverables produced, or token logs if the tool exposes them — and have Claude turn that into a standing per-client AI cost figure instead of a once-a-year guess.
You are helping me build a per-client AI cost allocation model so I can bill AI usage as a separate materials line instead of absorbing it into a flat fee.
Here's what I have:
- List of AI tools used in delivery and their billing model (flat seat, usage-based, hybrid)
- [PASTE USAGE DATA OR TIME-TRACKING DATA PER CLIENT IF AVAILABLE]
Produce:
1. An estimated AI cost per client per month, based on actual usage where I have it, and a reasonable allocation method where I don't (explain which clients are estimates vs. real data)
2. A flag on any client where AI cost is trending up month over month, which signals a renewal conversation is needed
3. A simple formula I can reuse monthly to update this without rebuilding it from scratchRestructure the invoice into two lines, not one, before the next renewal
Split the invoice into a labor fee (the work, priced the way it's priced today) and an AI materials line (metered cost, pass-through with a modest markup, the same way agencies already pass through stock photography licenses or paid media spend). This is the structural fix the Digiday-reported 'token costs as materials' model is pointing at — it's not a rate increase, it's making a real cost visible on paper instead of hiding it inside a number that doesn't move when the underlying cost does.
Write the renewal conversation before the client asks the AI-discount question
58% of big advertisers already expect AI to mean lower fees. If the agency doesn't control that conversation, the client's procurement team will, and they'll frame it as a straight discount instead of a cost-and-value conversation. Get ahead of it with language that separates the two lines explicitly.
Write a short renewal conversation script for a client who may raise the "AI should make this cheaper" objection.
Context: [DESCRIBE THE ENGAGEMENT AND HOW AI IS USED IN DELIVERY]
The script should:
1. Proactively name that AI is part of delivery, not deny or minimize it
2. Separate the labor fee from the AI materials cost explicitly, showing the materials line is a pass-through tied to actual usage, not agency markup padding
3. Make the case that AI changes what gets delivered and how fast, not that it makes the work worth less
4. End with a confident, non-defensive close that doesn't apologize for the pricing structureRe-check the vendor billing audit every quarter, not every renewal
Usage-based AI pricing moves faster than annual contract cycles. A tool that was flat-rate at last renewal can switch to metered billing mid-contract with little warning, the way Anthropic and GitHub both did in 2026. Put a standing quarterly re-check on the calendar so a vendor repricing shows up as a line-item adjustment on the next invoice instead of three months of eaten margin discovered at year-end reconciliation.
What changes
AI cost volatility stops being the agency's problem to silently absorb and becomes a visible, metered line the client can see moves with actual usage. Vendor repricing — flat to usage-based, a renewal price hike, anything — shows up as a pass-through adjustment on the materials line instead of a margin hit nobody notices until the quarterly numbers look wrong. The renewal conversation about AI and fees happens on the agency's terms, with real cost data behind it, instead of as a reactive discount negotiation the client's procurement team initiated first.
Two things are true about AI and agency money right now, and most agencies are only tracking one of them. The first: AI vendors are moving off flat per-seat pricing. Anthropic switched Claude Enterprise to usage-based billing in April 2026. GitHub Copilot made its own usage-based shift in June. More tools in the delivery stack are following the same path, each metering differently, each capable of repricing mid-contract with little warning. The second: clients aren't waiting for agencies to figure out the first part before forming an opinion about what AI should mean for their invoice. Research cited by MarketingBrew and Search Engine Land puts the number at 58% of big advertisers already linking AI adoption to an expectation of lower agency fees — and 26% of B2C marketing executives say AI costs rising faster than expected are now the top driver of internal scrutiny on AI spend.
What Digiday found agencies are actually doing about it
Digiday's reporting on agency AI bills found most shops are handling rising AI costs the way they'd handle any other quietly growing overhead line — bundled into the flat fee, treated like a Slack subscription, with staff salaries remaining the one cost everyone actually watches on the P&L. That approach works fine when AI tool spend is small and stable. It stops working the moment a vendor switches to metered billing and the agency's actual cost base starts moving monthly while the client invoice doesn't move at all. The same reporting found at least one agency already experimenting with a different model: billing AI token costs to clients as "materials," tracked separately from the pre-set labor fee — the same way agencies have always passed through stock licensing or media spend without folding it into the service fee.
Why "materials" is the right frame, not "markup"
Framing AI cost as a materials pass-through does something a straight price increase doesn't: it makes a real, variable cost visible instead of asking the client to accept a number increase with no underlying reason attached. A materials line that moves with metered usage is defensible in a renewal conversation in a way a vague "AI surcharge" never will be, because it's tied to an actual invoice from an actual vendor, not an agency's internal guess about what AI should be worth.
The renewal conversation agencies need to get ahead of
58% of advertisers expecting AI to lower fees isn't a future problem — it's showing up in procurement conversations now, and if the agency hasn't already separated labor cost from AI cost on paper, that conversation gets framed as a straight discount request instead of a cost discussion. An agency that walks in with two clearly separated lines — labor priced the way it's always been priced, AI materials metered against real usage — can have an entirely different conversation than one defending a single blended number against a client who assumes AI made the whole thing cheaper to produce.
Bottom line
Vendor pricing is moving in one direction and client fee expectations are moving in the other, and an agency tracking neither with real data is the one that eats both moves at once. Splitting the invoice into a labor line and a metered AI materials line doesn't fix either pressure on its own, but it makes both visible enough to manage instead of discovering the damage at the next reconciliation.