The Unbilled AI Overhead the Agency Absorbs, and Why

🇹🇷 Türkçe: Bu yazının Türkçesini oku →

An invoice is a contract between what a client paid for and what they received. The hard part of transparent AI billing is not the arithmetic. It is drawing the line between what the client bought and what the agency bought for itself, because the AI costs look like one shape and turn out to be two.

We have been itemising AI usage line by line on client invoices for about a month now. The first four parts of this series walked through why we started charging for AI at all, four wrong ways to price a token, the pricing unit we settled on, and putting it on the invoice line by line. This piece is about the other side of that line: the AI work that stays on our side of the meter and does not reach a client.

Thirty-plus years ago on a Mac Plus, with QuarkXPress and a photocopier for proofs, the thing that separated a job you quoted properly from one that bled quietly through the month was knowing which hours were craft and which were production. Craft came out of the fee. Production got billed. The boundary was common sense and almost nobody got it wrong. The same question returns now in a different shape, with a different machine on the other end, and the common sense is not settled yet.

What follows is one agency’s version of the boundary. Four kinds of AI cost that stay with us, the arithmetic that makes that honest, and the handful of questions I think a buyer should put to any supplier about to add “AI” as a line on an invoice.

What a Session Consumes Is Not What a Client Bought

Three sessions last month, two 1,500-word articles delivered, metered at list price: $161. Reading the receipt out loud to the founder of this agency was uncomfortable, and his reaction, cleaned up for print, was essentially this: two short pieces of writing, and the meter reads two hundred and eighty dollars.

He had every reason to object. Over 110 million tokens of cache reads were recorded in a single one of those sessions. Cache reads are the model re-reading its own context, because the session was long enough that it had to pick up where it had left off dozens of times. That is work the model does to stay coherent across a long run, not work delivered to a client.

What was actually handed over, measured against the input and output of the artefact itself, was in the neighbourhood of $0.45 of Opus or $0.09 of Sonnet for two 1,500-word articles. The gap between $0.45 and $161 is the thing we are asking the client not to fund. Everything in this post is downstream of accepting that gap exists.

Four Things That Look Billable and Are Not

In roughly ascending order of how expensive each one was to learn:

Cache Rereads (Session Length Is Our Problem)

If the agent re-reads the entire project context thirty times across a single session because that is how long the work took, those re-reads are the cost of us using a long session. The client did not ask for a long session. They asked for the output. On Anthropic’s own pricing, cache reads are priced at a tenth of the base input rate, but a hundred million of them still produce a real bill. Nothing about that bill is work the client commissioned.

Retries and Backtracks (Craft Is Ours)

Half an attempt rewritten because the first pass had the wrong angle is not a billable line item. It is the agency’s craft. There is a conservative version of this where retries get billed as “work in progress”, and we tried it for a day. It collapses the moment a client reads the invoice and asks, with reason, why they are paying twice for the same thousand words. The right unit is the thing delivered, not the attempts behind it.

Research-Side Token Traffic (Already in the Hourly)

Discovery, reading a brief, loading our own notes: the AI runs during research the same way it runs during delivery, and the tokens look identical in the usage log. We bill research as human hours (we were already doing that long before AI arrived), and the AI’s role inside those hours stays on our side. Charging a client twice, once for the hour the person spent researching and once for the tokens the model consumed during it, is not transparency. It is a double count dressed as detail.

Model-Driven Gatekeeping (About $19 a Month, Absorbed)

We run a pipeline that pre-triages incoming client activity through a small model: pulls the queue, filters obvious work, parks the rest for human approval. It costs around $19 a month in API usage across the whole book. The temptation was to add it to invoices under “AI operations”. We refused. The triage is the agency doing its own gatekeeping. The client did not commission the gatekeeping; the client commissioned the delivery. Nineteen dollars is small money and the principle is bigger than the number: anything the agency runs to decide whether to do work is agency cost, anything the agency runs to produce the thing a client asked for is client cost. One sentence saves a day of arguing about edge cases.

What the Invoice Actually Sees

Everything above means the invoice only sees one measurement: the token weight of the delivered artefact. Input tokens plus five times output tokens, where the “five times” is not a convention but an invariant. On every current Anthropic tier, from Haiku through Opus, output is priced at exactly five times input (Anthropic’s current public pricing). The ratio collapses two numbers into one.

weighted tokens = input + (5 × output)
1 invoice unit  = 100,000 weighted tokens

For a 1,500-word article: around 5,000 input and 2,000 output, which is 15,000 weighted tokens, or 0.15 units. On Sonnet that converts to roughly nine cents of model cost, on Opus around forty-five cents. The invoice carries it at two times cost, same multiplier we apply to every external provider line, so it reads as around eighteen cents or ninety cents on the client’s end, with the margin stated plainly rather than hidden.

The measurement itself comes out of the raw session JSONL files that the client-side agent writes to disk, de-duplicated on message.id because 8,355 of the 19,181 usage-bearing lines in our transcripts turned out to be duplicates of the same message appearing in multiple reads. A naive sum doubles the bill. If this measurement is in your supplier’s hand, that one word, “de-duplicated”, is worth asking about by name.

What We Eat Per Month, and Why the Trade Still Works

A month of running the delivery-only model gives us a shape of what the agency absorbs:

• Cache reads across all client sessions, in the low tens of dollars at list price, varying with how long individual sessions run.
• The triage model, around $19 a month, flat across all clients.
• Research-time AI, folded invisibly into the hourly rate and not separately metered.
• The overhead of maintaining the measurement itself (the scripts, the de-duplication, the per-client reports), paid as our own time.

Set against what we bill clients, which runs in the low three figures per client per month of itemised AI usage so far, the absorbed overhead comes to a few percent of the model spend. It is not nothing, and in a tight month it moves the needle slightly. But the alternative is a line on an invoice that reads “AI operations, forty dollars” and a client who, correctly, asks what that is and gets back an answer the agency cannot defend without reading the usage log.

There is a ratio question I did not expect to find useful. One hour of active AI work, measured by weighted tokens, runs at around 480,000 of them, which translates to roughly €14 of model cost at Opus list rates or around €12 at our customer-facing price including margin. An hour of senior human time on the same project bills at €90, before overheads. The ratio for a given hour is around 1:7. That is not a benchmark, every job varies, but it is a sanity check. If an AI line on a monthly invoice is approaching the human line, something is wrong at the measurement, not at the rate.

Three Things I Would Do Differently Next Time

First, start the conversation at the delivery base, not at the session total. The $161 figure was a diagnostic. It was also the number the founder saw first, and the first conversation had to recover from it. Had the first number been “the article cost nine cents of model, we are billing eighteen”, nobody would have objected to anything. The arithmetic has to arrive before the receipt does.

Second, write the de-duplication rule into the first script, not the fifth. More than four in ten usage-bearing lines in the raw transcripts were duplicates. Every attempt to compute a billing base before we fixed that was roughly two times inflated. Three of our four failed pricing attempts would never have been failed had the measurement itself been clean. The lesson keeps repeating in a thousand smaller forms: a broken ruler teaches nothing useful about the plank.

Third, treat gatekeeping overhead as a fixed category from day one. Spending a day debating whether $19 of triage was client cost or agency cost was time nobody paid for. The rule, now written into the spec: anything the agency runs to decide whether to do work is agency cost. Anything the agency runs to produce the thing a client asked for is client cost. The sentence closes the argument before it starts.

What to Insist On If You Are Buying Agency AI Work

If you are the person signing an agency contract that includes AI line items on the invoice, there is a small set of questions worth putting to the supplier. None of them require you to be a developer. All of them separate a serious billing model from one that hopes you will not look closely.

Ask whether the AI line items measure the session or the delivered artefact. If the answer is “the session”, or if the person on the other end of the call does not understand the distinction, you are being asked to fund their tooling. Ask whether their usage data is de-duplicated on message identifier. The word “message.id” should mean something to them, or at the very least they should be able to describe the de-duplication rule in a sentence; if they cannot, the invoice is almost certainly double counting. Ask what sits on their side of the meter. “Cache reads”, “retries”, “research time”, “internal gatekeeping” are the right answers. If everything the model touched is being passed through, the invoice is a bill for the agency’s craft, which you were already paying for in the hourly rate.

Ask for the arithmetic of one line, picked at random. “This item is 0.3 units at €3 each. What does that mean?” You should get back a sentence like the model produced about a thousand words of output from a brief of about two thousand five hundred, that converts to roughly thirty-two thousand weighted tokens, 0.3 units, two euros of model cost at list price and three at our rate. If the sentence does not come, the measurement does not exist.

Inside a real WordPress or WooCommerce estate, the same test lands as a brief. If AI is going to touch the catalogue (product descriptions, variations, enrichment), the support stack (replies, triage, routing), or the content layer (editorial drafts, translation, summaries), the engagement should name, before the first sprint, which bytes count as delivered output and which bytes are the supplier’s own paperwork. A hundred-SKU test batch, measured in isolation, takes two to three days to set up and a week to run honestly. The invoice that comes out of that run is the first one that reads as a real cost rather than a vendor’s best estimate, and it is the one worth basing a quarterly budget on.

A buyer who skips that step will not know whether they are paying their agency to think or to produce. Those are different numbers, and most agencies would prefer you did not notice.

The Boundary Is Worth Paying For

Transparency has a cost in both directions. For the agency it means absorbing four or five percent of our AI consumption as overhead rather than passing it through. For a client it means knowing precisely what they funded, down to a few thousand tokens. Twenty-plus years of invoices have taught me that the version of the bill a client can read without wondering what is hidden under it is the version that keeps the relationship alive at the end of year three. The one that renews the contract is the one you did not have to defend. If you are planning this for your own supplier relationship, or building the measurement for the first time inside your own commerce estate, we built the arithmetic that makes it work, and are happy to walk a team through it.