In 1993 the Istanbul print shop I was helping as a teenager still priced per hour, counting the typesetter’s time on each job. Then designers started setting their own type inside QuarkXPress at home, and the hour the shop sold stopped being the hour the client was paying for. The invoice took three years to catch up. Agencies and consultancies are in the same gap today, and the thing being sold is tokens.
If you are signing off on marketing or software work, this matters more than the opinion columns about the death of the billable hour. The question for you is not where billing ends up five years out. It is what your next invoice looks like, and that is already decided by what agencies are quietly doing with the AI line item right now.
So I went looking for named positions. Below are agency executives who have said on the record how their shop handles AI token cost, grouped into the three camps the market is sorting into, plus one sceptic and one warning. I have added our own view on what to actually ask in an RFP.
I Asked Nine Agency Bosses Where the Token Lives on the Invoice
These positions are drawn from Sam Bradley’s reporting in Digiday, 3 March 2026, in which eleven senior agency and buy-side executives discussed how their firms handle the token bill. Grouped by position, every quote verbatim.
- Johnny Rohrbach, co-founder and director of partnerships and operations, Silverside AI
- Kyle Smith, chief technology officer, Merge
- Taryn Crouthers, CEO, Big Spaceship
- Lisa Herdman, SVP and executive director of video investments, RPA
- Chris Neff, global chief AI officer, Anomaly
- Will Hanschell, CEO and co-founder, Pencil (Brandtech)
- Krish Kuruppath, EVP and head of tech, Horizon Media
- Domenic Venuto, chief product and data officer, Horizon Media
- Peter Rice, VP of data strategy and analytics, Kepler
- Josh Archer, chief data and intelligence officer, Lerma/
- James Londal, founder and CTO, Avelin
- Jess Lewis, global chief technology and data officer, Crossmedia
- Ruben Schreurs, CEO, Ebiquity
Camp One: Pass-Through at Cost, With the Meter Visible on Every Line
The first camp treats tokens the way agencies long treated media and third-party production: a cost stated plainly and passed through, with no mark-up or a small handling margin. Josh Archer at Lerma puts the stance in one line: “We’re not marking it up, we just show it as a line item similar to other costs.” Merge’s CTO Kyle Smith takes a comparable position; his agency meters AI usage case by case and bills the client for it. The invoice tells the client exactly how much compute their work consumed.
Johnny Rohrbach at Silverside AI is in the same camp with a different mechanism. Because “you have to use a lot of different models and all of those have different token economics,” his firm charges a seat-based subscription rather than meter every call. The buyer’s experience is similar: a line item for AI, not a hidden overhead. The difference is that one is per-use and the other per-user, and that difference matters when your usage is spiky.
What this camp is really saying: expect the token on your statement, expect “at cost” to be the frame, and expect the model to shift between metered and seat-based depending on where you sit in the usage curve.
“We’re not marking it up, we just show it as a line item similar to other costs.”
Josh Archer, chief data and intelligence officer, Lerma
Camp Two: The Agency Carries the Token Until the Client Gets Value
The second camp absorbs the token into the agency fee, treats it as production cost, and does not break it out. Taryn Crouthers at Big Spaceship puts it cleanly: “We’re treating it similar to a production cost.” That is the position of a shop that is not willing to charge for inputs it cannot yet prove delivered something.
Lisa Herdman at RPA sharpens it: “We can’t expect to charge our client for something we don’t know is actually going to work for them.” This is a position on risk, not sentimentality. The agency pays for its own learning curve. The client sees a flat fee against a deliverable. The variance in what it cost the agency to produce is the agency’s variance.
For a buyer this looks like the quieter invoice, which is nice until you notice the trade-off. The agency has to price for a wide range of possible token spends, which usually means pricing on the pessimistic side. In a long stable engagement that evens out. In a short pilot you are paying for the agency’s uncertainty rather than their efficiency.
Camp Three: The Invoice Stops Talking About Tokens Altogether
The third camp refuses the input question entirely and sells the outcome. Will Hanschell at Pencil frames the discipline it imposes: “You want to compete on merit, not on lowest price.” His argument elsewhere in the piece is that usage-based pricing aligns the agency’s incentives with the client’s, because both sides want the campaign that works, not the campaign that burns the most tokens.
Peter Rice at Kepler says it plainly: “Impact is the metric that we care about. Token usage is the fuel to deliver that impact.” Fuel is a running cost; impact is the product. Nobody prices a flight on the kerosene burned. James Londal at Avelin pushes from another angle: “The bigger economic shift isn’t token arbitrage. It’s labor compression.” If the real change is that the same campaign takes a quarter of the headcount, pricing by inputs misses the point.
Ruben Schreurs at Ebiquity, speaking for the buy-side auditor view, agrees from the client’s seat: “It’s less about the input and more about how incremental business returns are being delivered.” He expects buyers to start demanding token-usage audits where contracts allow it. Even the camp selling outcomes should assume the camps selling inputs will be independently verified.
“Impact is the metric that we care about. Token usage is the fuel to deliver that impact.”
Peter Rice, VP of data strategy and analytics, Kepler
The Honest Sceptic: Is Pass-Through Just the Latest Mark-Up
Chris Neff at Anomaly is the camp of one worth listening to. He said directly that passing on token costs “feels like a money grab.” Not because tokens are not a real cost, but because the industry has a history of inventing line items that outlast their justification. His scepticism is the one you want in an RFP: the person who asks whether this is a pass-through or a new centre of margin dressed as transparency.
I am not with Neff on the principle, having spent the autumn building our own version of transparent AI line items on invoices, and the honest point is that transparent arithmetic raises more questions than it dodges. But I am with him on the warning. If an agency moves to pass-through, you should see the arithmetic. If you cannot, “transparency” is doing the opposite of its job.
Platform Cost Recovery Is a Different Conversation Entirely
Domenic Venuto at Horizon Media describes a fourth model that looks like billing but is not: “It’s cost recovery, not revenue.” Horizon’s internal AI product Blu is onboarded with a nominal fee designed to recover setup effort, not extract margin. His colleague Krish Kuruppath notes token usage only turns into a significant cost once the user base is in the tens of thousands.
This is a platform economics argument, not an agency pricing one. It belongs here because the line between product and service is blurring across mid-market work. If your agency is also running a product you use, you should expect two different billing conversations. Mixing them is where the dispute happens.
The Warning Everyone Should Pin Above the Procurement Desk
Jess Lewis at Crossmedia offers the sentence that stays with me: pass-through agencies risk becoming “solely focused on cost and margin, rather than focused on value.” This is the race-to-the-bottom warning, directed at camp one. If the invoice becomes a conversation about cents per thousand tokens, the agency has quietly moved from thinking partner to intermediary, and the client eventually finds a cheaper intermediary.
The camps are not equal. Pass-through wins transparency; production-cost wins simplicity; outcome wins the long-term relationship. A thoughtful shop runs all three at different moments and tells you which one it is quoting under.
Safe Bet, Outrageous Prediction
The format calls for one of each from the chair.
Safe bet. Within twelve months, every mid-market marketing or software contract above a certain threshold carries at least one explicit AI-related line item, whether that is tokens, seats, usage-based fees, or an outcome-tied bonus. The default of “it is in the hourly rate” is already illegible to buyers’ finance teams.
Outrageous prediction. By 2028 the billable hour disappears from AI-dense engagements, replaced by a two-column invoice: metered inputs on one side (tokens, API, compute), outcomes with a defined measurement on the other. Human hours get rolled into one column or the other, not standing alone. This is already happening at the top of consulting; the only question is when it reaches mid-market shops.
Stop and Start: Two Short Lists
Agencies should stop: hiding AI cost inside a flat hourly rate on AI-dense work; marking tokens up at 2× without saying so; writing the AI line in model-vendor units a buyer cannot decode; pretending one pricing model fits every engagement.
Agencies should start: declaring the pricing model up front, in plain language, before the SOW is drafted; publishing the arithmetic for one representative deliverable; letting the client pick metered or absorbed at pilot and reviewing at renewal; auditing their token usage monthly so they can defend it.
Buyers should stop: treating “AI is included” as lower risk than “AI is a line item”; comparing agency fees on headline hourly rate when half the input is now metered; asking only once at RFP and never again.
Buyers should start: asking which of the three camps above the agency is quoting under, by name; requesting an audit clause for metered inputs that materially affect the fee; reviewing the model annually, not every three years.
What a Mid-Market Commerce Team Should Put in the RFP Next Monday
Our side of this is what we would build and what we would ask, inside a real WordPress or WooCommerce estate the team already owns. Four concrete items for a procurement team about to sign a twelve-month engagement where AI sits inside the deliverable:
- Pricing-model declaration clause. Require the agency to state, in one paragraph, which camp above they are quoting under. If they quote under more than one for different workstreams, require them to say which workstream is which. Thirty minutes of their time saves six months of invoice disputes.
- Unit arithmetic on one representative deliverable. Pick one piece of work the engagement will produce often, a landing page or an email campaign, and require the agency to show the token arithmetic for one worked example at list price, before the SOW is signed. If they cannot, their meter is not instrumented and the pass-through promise is theatre. We publish ours: input plus five times output equals weighted tokens, 100,000 weighted equals one unit.
- Audit clause with named provider units. Require that any metered line uses the provider’s own billing unit (OpenAI, Anthropic, Google tokens) rather than an agency’s internal “AI credit”. Credits are where mark-up hides. Provider units reconcile against your own direct observation.
- Measurement baseline before the pilot. Instrument the WordPress or WooCommerce estate for the outcome you want the agency to improve, before they start, in GA4 and in your own order data. Pilot is four to six weeks of baseline then a quarter of observation. Outcome-based pricing without measurement is a signature on an argument you cannot win.
Done well this takes a procurement lead and an in-house engineer about three working days, and roughly a sprint of agency time to answer. Cheap insurance on a twelve-month agreement.
What Is Not Resolved, Honestly
The three camps do not agree on what counts as transparency, margin, or a reasonable outcome. Pass-through defines transparency as visible units; outcome defines it as visible results. Neither has fully absorbed the auditor view Schreurs is quietly pushing, that metered inputs will be independently verified whether agencies invite that or not. In a few years I expect one camp to win on vocabulary while the other two win on practice, which will be its own kind of mess.
For now notice that the agencies that have published their position are already a self-selected minority. The invoice you will receive from the shop that has not said anything yet is still being drafted in a meeting room, with somebody in finance guessing which camp lands softest with their largest client. If you are that largest client, you should have an opinion about which camp it should be, and send it over before the SOW arrives. We are happy to help you draft that opinion. Our own line-item invoice is public at thewpclan.com/ai-invoice-line-items/, and if you want our view on your particular stack, talk to us.
Last modified: October 29, 2026
