When an agent does the work, the marginal hour costs about forty cents of tokens. Bill that hour at $150 and you are lying. Bill it at cost and you starve. The way out is to stop selling hours entirely and price the outcome the system maintains — we charge flat monthly fees, and the math below is exactly how we set them.
Every AI-heavy shop hits this wall in year one. We hit it in month four.
Here is the model we landed on, after three wrong ones.
Why hourly billing collapses first
Hourly billing assumes effort scales with value. Agents break that assumption in one direction only: a deliverable that took six billable hours in January takes eleven minutes of review in June. If you bill honestly, your invoice falls 90% while the client gets the same artifact. If you bill six hours anyway, you are now selling fiction with a timesheet attached.
Worse, hourly creates a perverse incentive to keep work slow. Every automation you build is a pay cut you volunteered for.
We watched our own effective rate climb past $600 an hour on paper. That number is not a pricing win. It is a signal that the unit is wrong.
Clients notice the wobble too. An invoice that swings between four figures and two raises questions that a flat fee never invites.
The hour was always a proxy for value. Agents just made the proxy unusable.
Price the maintained outcome, not the task
What clients actually buy from us is a state of the world that persists: the site stays cited, content ships every week, audits run monthly, regressions get caught and rolled back. That is not a pile of deliverables. It is a condition, maintained over time, and the natural container for a maintained condition is a flat monthly subscription — not a timesheet and not a per-piece rate card.
Deliverable pricing fails almost as fast as hourly. Count the units and you invite haggling over units, and you hand yourself an incentive to pad the count.
The sentence we use on sales calls: you are not paying for the work, you are paying to never think about whether the work happened.
The flat fee also changes client behavior for the better. Nobody rations questions or hoards requests to batch into a billable hour. The relationship gets easier the moment the meter disappears.
Framed that way, nobody asks what an hour costs.
Know your floor: the robot ledger
You cannot price a subscription without knowing what a client costs you. Ours run about $85 a month in tokens and infrastructure, plus 3.5 hours of human review and judgment, which we cost internally at $200 an hour. Call it $785 all-in. Our floor rule: never price below twice true cost. For us that puts the minimum viable retainer around $1,500 a month.
The 2x is not greed. It is the buffer that absorbs the client who needs double the review hours, the model price change, and the month something breaks.
Most shops skip this arithmetic and price on vibes. Then one needy client quietly eats the margin of three good ones.
One subtlety: cost your review hours at your target rate, not at zero, even though you are the reviewer. Treating your own judgment as free is how solo operators end up underpricing every offer they make.
Track the ledger per client, monthly. It is four numbers.
Know your ceiling: the replacement quote
The ceiling is what the client would pay to get the outcome without you. A junior marketing hire runs $4,800 a month before overhead. A traditional agency retainer for the same surface area runs $3,000 to $6,000. Our $1,500 sits at roughly a third of the replacement quote, which makes the decision easy for the buyer and leaves us room to raise prices for years.
Price too close to the ceiling and every renewal becomes a bake-off. Price at a third of it and renewals are a formality.
The gap between your floor and the replacement quote is your strategy space. Ours is wide precisely because the labor is robotic.
Run the replacement math with the prospect, on the call, out loud. It reframes the fee from a cost into an arbitrage before you ever have to defend it.
If your floor sits above the replacement quote, you do not have a pricing problem. You have a cost problem, and no invoice format fixes it.
The three models we tried and retired
We ran hourly for four months, per-deliverable for three, and usage-based passthrough for two, and retired all three. Hourly collapsed as described. Per-deliverable turned every scope conversation into unit haggling and paid us to pad. Usage passthrough — billing tokens plus margin — confused clients, punished our own efficiency gains, and made invoices swing 40% month to month for no reason a client could see.
The pattern across all three failures: each one priced an input. Inputs are exactly what agentic work makes cheap and volatile.
Flat outcome pricing was the boring survivor. Predictable for the client, margin-expanding for us, renegotiated once a year.
Boring is what pricing is supposed to be.
Answer the 'it is just AI' objection with the ledger
Sooner or later a prospect says it: why pay $1,500 when the AI does the work? We answer by opening the ledger. Last month the engine shipped 34 changes, ran six experiments, rolled back two of them on its own telemetry, and a human spent 14 hours reviewing, steering, and saying no. The client is not renting tokens. They are renting a supervised system with judgment attached.
Transparency beats defensiveness here every time. Hiding the automation invites the objection. Showing the operation retires it.
The honest version of the pitch: yes, the labor is robotic, and that is why the price is a third of the alternative — not a tenth, because the judgment is not robotic.
We send the ledger monthly now, whether anyone asks or not. Ten line items, five minutes to read. Renewal conversations start from that document instead of from memory.
Clients do not pay for effort. They pay for outcomes they no longer have to inspect.
When we still sell hours
Hours survive in exactly two places. Discovery — the first two weeks of a new engagement, where the work is genuinely novel investigation — and custom builds, like a first-time integration nobody has wired before. Both are one-time, judgment-heavy, and impossible to price as an outcome because the outcome is not yet defined. Everything recurring moves to the flat fee, behind a one-time setup charge of $3,000.
Then re-price at renewal. The engine absorbs more scope every quarter — ours covers about 30% more surface than it did a year ago — and the fee should track that, typically 10 to 20% a year.
The client who balks at a renewal increase is usually the one who has not seen the ledger. Show the ledger.
Price the condition, show the receipts, keep the hours for the frontier. That is the whole model.