Artificial intelligence

Setup

Roughly 4,000 paid seats sold into engineering teams at a flat $39 per seat per month, with inference bought on a variable per-token contract that repriced quarterly. The board pack reported one blended gross margin across every account, and the sales team was compensated on seats, not on consumption.

What happened

The heaviest decile of users consumed about nine times the median, and a long tail of agent retries — failed tool calls re-run automatically — accounted for close to a third of tokens burned. Because pricing was flat, that decile was served well below cost, and it grew fastest, because heavy users are also the ones who evangelise internally. Blended margin fell for four consecutive quarters, from the high fifties into the low thirties, while headline revenue growth stayed at roughly 12 percent a month and hid the problem from anyone reading only the top line.

Where it stands

The company survived by moving to a seat plus metered-overage model with a published token allowance, and by attacking retries directly: better tool schemas and a cheaper router model cut the median task from eleven agent turns to six. Margin recovered two quarters before revenue growth reaccelerated, which is the order we expect and the order we now look for. Net revenue retention held through the repricing because the heaviest accounts were also the ones getting the most value.

Heavy decile vs median usage

~9x

Tokens spent on retries

~1/3

Blended margin trough

high 50s → low 30s

Fixed-price seats sitting over variable inference is a bet that usage is predictable. Underwrite the distribution of usage, not its average, and ask for margin by decile before you accept a blended number.

Setup

A thin product layer over a frontier model, sold into legal operations at six-figure annual contracts. No proprietary corpus, no write access to any system of record, and an integration a customer could complete in two weeks. The pitch was accuracy on contract clause extraction, benchmarked against a public dataset.

What happened

Two of its three largest customers rebuilt the workflow in-house once model prices dropped roughly 90 percent over eighteen months and a four-person platform team could justify the build. Switching cost was two weeks, so it was paid without much internal debate. The vendor's benchmark lead had shrunk to noise by then, because the underlying model — the same one the customers were calling — had improved for everyone at once.

Where it stands

The surviving competitor in the same category writes reviewer corrections back into the matter-management system and holds the audit trail. Its correction set now covers several million reviewed clauses, it fine-tunes against that set quarterly, and its gross churn is about a third of the wrapper's. Its accuracy edge is smaller than the wrapper's ever was; its retention is roughly three times better.

Model price decline over the window

~90%

Customer switching cost

~2 weeks

Churn vs corrections-capturing rival

3x worse

The moat is where the corrections land, not where the prompt is written. Ask what the product accumulates that the model provider cannot, and how long a competent in-house team would need to replace it.