Here is a question worth asking the next agency that pitches you a content programme, a design system refresh, or a migration where “AI will speed things up”. If AI is doing part of the work, which line on the quote covers it. If the answer is a vague gesture at the hourly rate, you are the one carrying the variance.
We spent September rebuilding how we bill for AI usage. Human hours sit on one line. Token consumption sits on its own, per model, in units a client can read. External provider spend sits on its own too, at the provider’s billing unit. That part is now boring. The interesting part, the one almost nobody has written about honestly, is what happens upstream: the quote you write before the engagement starts, when one of your inputs is no longer time but consumption.
I got the first four of those quotes wrong. Not by a little. By enough that one client’s monthly report, in September, briefly suggested we had burned nine hours against a two-hour budget, and the right figure was 0.81. The report was broken, which was the direct cause. The deeper cause was that we had been writing statements of work in a shape the new billing could not carry. This piece is about what we got wrong, what we quote now, and the short clause we would ask any buyer to put into their next AI-adjacent agreement.
Everything below is from our own books. The arithmetic sits on the pricing unit we settled on and the line-item invoice that came out of it.
The Constraint Was a Quote, Not an Invoice
Retrospective billing is a solved problem once you accept the data. The session transcript tells you how many tokens went through, which model handled them, and what the provider would charge at list. Multiply by the sale rate, divide by the unit, and the invoice line writes itself.
A quote is not that. A quote is written before the work exists. The buyer reads it and signs a page that commits them to a number. If the number is wrong by a factor of three in either direction, trust leaves the room. In September we had three live engagements: a monthly content retainer for an industrial client, a quarterly SEO programme for a regional retailer, and a one-off migration estimate for a hospitality group. All three asked the same question, in different words: what is this going to cost us this quarter.
Until that month the answer had always been a function of two things we both understood: our hourly rate and an estimate of hours. From 2026-09-14, after we migrated token consumption off time records and onto expenses, hours and tokens were measured separately. The quote had to say two numbers, not one, and we did not know how to size the second one.
Four Estimates We Got Wrong
Each of the four attempts below was tried against at least one live engagement. None of them survived. The failures share a shape, which I will come to.
1. Tokens as a fixed percentage of hours. The first move, and the most seductive, was to quote AI tokens as a flat uplift on human hours. “Add fifteen per cent.” It is easy to write, easy to read, and wrong in both directions. On deep research work, where the assistant reads long documents and compares across them, the ratio of token consumption to human hours is much higher. On routine production work, where a human is editing a near-final draft, it is much lower. We quoted fifteen per cent on a research-heavy piece and the delivered weighted-token count was closer to forty per cent of the hours rate. We ate the difference, because the alternative was a conversation with the client about why a percentage we had chosen for them was wrong.
2. Session-count estimates. The second was worse. We priced a short programme as “eight sessions at €X each”, where a session was a bounded working block in the assistant. The problem is the same one that killed every retrospective-billing attempt before we found the current unit. Session length is not delivered work. A single morning can run into one 110-minute session with heavy cache re-reads or three shorter bursts; the token consumption on the same deliverable can vary by a factor of five depending on how it is sequenced. Quoting by session is quoting by the shape of the tool, not by the thing the client receives.
3. Per-deliverable flat token budgets. The third attempt was more disciplined. We measured what two published 1,500-word articles had cost at delivery: 15,000 weighted tokens, or 0.15 units in our catalogue, €0.45 on Opus. We quoted the next six at 0.15 each. Three came in under; three came in at double. An article that references an internal catalogue, pulls in three external sources, and gets two editorial rounds against a style sheet is not the same input as one written against a brief. “One article” is not a unit. We moved to deliverable classes, which is the shape we use now.
4. A single budget that mixed hours and consumption. The fourth is the one that cost us the most credibility. A client had set a two-hour monthly cap on our agency’s assistance and we had quoted inside it in good faith. Then our weekly report, drawing on the old time-records endpoint, picked up token expense lines as if they were hours and summed them into the capacity figure. The preview that Monday read 9.1 hours against a 2.0 hour budget. The real human time was 0.81. Twenty-eight hours of apparent overage was 23.67 token units counted as hours, plus €4.68 of external API spend counted as hours. We caught it before the report was sent, but the deeper lesson was not about the reporting bug. It was that a single-budget quote assumes the two inputs are commensurable, and they are not.
Four attempts, one shape. Every one of them tried to squash a variable consumption input into a shape designed for time. Each produced either a false ceiling the buyer did not understand or a false precision they could not rely on. The fix, when we finally stopped pushing against it, was obvious.
What We Quote Now, and Why That Clause Belongs in the Agreement
The current SoW carries two budgets. A human-hour budget, in hours, at our rate. An AI-consumption budget, in units, where one unit is 100,000 weighted tokens (input plus five times output, since every model we use prices output at five times input), which is roughly ten to fifteen minutes of active assistant work. Each line states both numbers, on the deliverable class the work falls into, not per deliverable: research-heavy, routine production, review-only, and so on.
Two further things earn their place. The AI line carries a stated variance band, plus or minus thirty per cent, and crossing that band triggers a conversation, not a silent invoice. The quote names the model mix as part of the price: a line quoted at Sonnet rates cannot silently escalate to Opus without a re-quote. Buyers tend to accept both clauses once they have seen a bill broken out this way; the variance band in particular, read against a traditional fixed quote with no visibility, reads as a transparency gain rather than a hedge.
Numbers in practice, from our own September and October records. A 1,500-word article at Sonnet: 0.15 units at €0.60 per unit, so €0.09 of consumption against roughly €0.45 cost to the buyer at our sale rate. The same article on Opus: 0.15 units at €3.00, €0.45 pass-through at list, €0.90 billed. A ten-item product-description programme, in our records, lands between 0.6 and 1.1 units per item depending on research depth. For a hundred items at a mid-point of 0.8, that is 80 units, €240 on Opus and €48 on Sonnet, which is the AI line on the quote. The human-hour line is separate, costed as it always has been.
The Measured Result Is Boring, Which Is the Point
Since the two-budget model went live, the three September engagements have run through two full months without a single overage conversation that turned into an argument. The one AI line that crossed its variance band, on a piece that pulled in more external sources than the deliverable class expected, triggered a ten-minute call before the invoice went out. The client’s finance lead described that call as “the first time an AI cost has ever been explained to me before it appeared on the bill”. The engagement was renewed in November without a rate review.
Internally, 58 consumption records totalling €86.51 were migrated from time records to expenses when the model changed. The traps that surfaced during that migration are in the invoice piece earlier in this series. None of them reach a buyer. The two budgets do.
What We Would Do Differently
We would write the two-budget clause before we touched the catalogue. The order we worked in was expense categories first, then catalogue, then quoting; the order a buyer sees it in is the opposite, and the SoW is the first encounter. A buyer who signs a single-budget page and receives a bill broken out into human and AI lines will feel, correctly, that the pricing conversation happened after the fact.
We would also have classed the work by shape sooner. “Deliverable class” sounds procedural; the alternative is per-piece quoting, which fails as above. Four or five classes cover almost everything an agency our size handles: research, production, revision, operational, discovery.
Building the Two-Budget Quote on WordPress or WooCommerce
For an organisation evaluating this from the buyer side, two questions are worth asking the agency on the other end of the quote. First, where does your AI consumption get measured. Second, how does that measurement reach my invoice. The answers are implementation, not strategy, and they reveal whether the agency actually ships what it describes.
On a WordPress or WooCommerce estate that already has a project or billing surface attached, the pattern is straightforward. We model each engagement as a custom post type, with two meta fields on it: a human-hour cap in hours, and an AI-consumption cap in units. A second CPT holds deliverable classes, each with a per-unit rate and a default model tier (Sonnet, Opus, Haiku). A third CPT or a taxonomy on the first holds per-line allocations. Variance bands sit as a percentage on the engagement meta. All three CPTs are REST-exposed, so the project-management tool that runs billing can pull them through the WordPress REST API without a bespoke integration.
Measurement is the other half. Assistant session data lives, in our case, under per-project JSONL files with usage blocks on each message. A nightly job reads those files, deduplicates on message id (not optional; without it the count roughly doubles), computes weighted tokens per delivered artefact, and writes one expense record per deliverable to the billing system, carrying the task id so the piece lands on the right invoice line. Two engineers can build the WordPress side in a sprint and the measurement pipeline in two more. The traps are documented in the session-transcripts post earlier in this series.
What this costs to ship, rough order of magnitude, is three to four weeks of engineering for the WordPress side, a couple of weeks for the measurement pipeline if the assistant you use writes usable session logs, and a quarter of continuous calibration against real invoices before the catalogue stops drifting. We corrected the sale rates on three catalogue items against their first live bills; nobody got that right from list prices alone, and we would be sceptical of a supplier who claimed to.
The Short Clause Worth Pasting Into Your Next Agreement
If you are commissioning AI-assisted work this quarter, the three sentences below will do more for the clarity of your invoice than any amount of rate negotiation. “Supplier shall quote AI consumption separately from human labour, in the provider’s own billing unit, with the model tier named on each line. A variance band of plus or minus thirty per cent applies to each AI line; any line expected to cross that band shall be re-quoted before the work is done. All AI line items appear on the invoice at their delivered measurement, not at session totals, and the measurement method shall be stated on request.”
That clause does not solve quoting AI-assisted projects. It makes the shape of the problem visible, which is the part the first four attempts above all failed. If your current agency balks at writing it in, the next conversation is already more useful than the one you were going to have. If we are the agency, we wrote it before the catalogue, which is where we would start the next time. Our billing stack is in production; we are two conversations and an SoW away from running it against your work. The fastest route to that conversation is the agency page, or the invoice-line-items piece if you want to see the bill before the quote.
Last modified: November 18, 2026