Price the AI before you give it away

For twenty years the economics of software rested on a convenient fact. Once a feature was built, serving it to one more customer cost almost nothing. A second user of a scheduling screen or a reporting page added a sliver of hosting and support, and gross margins of 75 to 85 per cent followed. Pricing teams learnt to bundle, because bundling cost little and kept customers, and boards learnt to read gross margin as a measure of how good the product was.

AI features break that fact, and I do not think the industry has absorbed it yet. When a veterinarian at one of our 7,400 hospitals finishes a consult and asks our notes assistant to draft the medical record, we pay two outside providers: one to transcribe the conversation and one to run the language model that turns the transcript into a structured note. The cost is small per note. It is not small per year, and it grows with every customer who likes the feature. That is the part that should worry a CFO. In this product, success is what drives the bill.

A feature with a meter running

We launched the assistant in April 2024 inside our top tier, with no extra charge and no usage limit. The reasoning was familiar. Competitors were offering similar tools free, adoption would prove the product, and price could come later. By early 2025 about 3,100 hospitals were producing 2.6 million notes a month. Each note cost us $0.31. The monthly bill had passed $800,000 and was rising faster than the subscription revenue attached to it, and our subscription gross margin had fallen from 81.2 per cent to 77.4 per cent in a year.

None of this was a surprise in the sense of being unknowable. The cost per note was in the provider invoices from the first month. What was missing was a habit. Nobody had been asked, before launch, to write down what the feature would cost at full adoption and who would pay for it. Software companies have rarely needed that habit. Our planning documents had headings for engineering effort, launch date and expected adoption, and none for cost per use, because for nearly everything we had ever built the answer was close to zero.

The free-bundle argument deserves a fair hearing, and I gave it one. Customers who adopt an AI feature may be less likely to leave, and a competitor's free tool is a real threat. But the argument rests on an assumption that grows more expensive each month it is held: that customers who have been given unlimited use will later accept a limit. They resist, and reasonably so. The longer a company waits, the more its customers treat the free version as part of what they bought, and the more any later price feels like a withdrawal rather than an offer.

Where the cost is booked decides who manages it

During the beta, most of our inference bill sat in research and development. That was defensible while the feature was experimental and a handful of hospitals were testing it. It stopped being defensible once thousands of paying customers relied on it, because the cost was no longer the price of building something. It was the price of serving it. From the third quarter of 2024 we moved it into cost of revenue.

The reclassification mattered more than its size suggests. In research and development, the bill was an engineering expense managed against an engineering budget and reviewed by people whose job was to ship features. In cost of revenue it became part of gross margin, which our product leaders, our chief executive and our investors all watch. Within a quarter the conversation changed from "how much are we spending on AI" to "what does a note cost, and what do we charge for it", and engineers began asking how to make a note cheaper.

I have seen proposals to keep such costs out of the margin by presenting an adjusted gross margin that treats AI spending as an investment. I opposed that at Brindlewood and would oppose it on any board. An adjustment that removes a cost incurred every time a customer uses the product is not an adjustment. It is the income statement of a different, imaginary product. If a feature's cost really is temporary because engineering expects to reduce it, the honest course is to report the cost and show the curve falling.

An allowance, a meter and a floor

The pricing we introduced in July 2025 has three parts, and I would recommend the structure to other software companies even where their numbers differ. First, an allowance. Our top tier includes 250 notes per veterinarian per month, which covers about four in five clinicians, so most customers never think about a meter. Second, a meter. Notes above the allowance cost $0.40 each, and hospitals on lower tiers can buy the assistant for $149 per veterinarian per month. Third, a floor. Every AI feature has a minimum gross margin, agreed between finance and product before launch, below which its price or its design must change.

The meter is not mainly a revenue device; overage income is modest. Its purpose is to make heavy use visible to both sides, so that the customer sees what the feature consumes and we see where the cost sits. The floor matters more to me. It is a written agreement, made before launch, about the point at which we will act. Without one, every quarter of margin decline can be explained away as temporary, and usually is.

Pricing alone did not restore the margin. Engineering did most of the work once cost became a design requirement rather than an afterthought. Simpler consults are now routed to a smaller, cheaper model, silence is trimmed from the audio before transcription, and repeated passages in common procedures are cached. The cost per note fell from $0.31 to $0.11 by mid-2026, and subscription gross margin returned to 80.4 per cent.

There was a price for this approach, and I will not hide it. About 140 hospitals left or downgraded, citing free alternatives, and took roughly $3.1 million of annual recurring revenue with them. Some customers were surprised by their first metered invoice, which was our fault: we should have shown them their usage before we billed for it, and we spent $0.9 million on credits learning that.

For directors the questions are simple, and they do not require any technical knowledge of language models. For each AI feature, what does one use cost today, and what will it cost at full adoption? Where is that cost booked? What allowance sits in each price tier, and who agreed the margin floor? A management team that can answer those questions is running a software business with a new kind of cost. One that cannot is running a promotion and hoping the bill stays small.