Let’s talk about why the smartest technology decision your company made this year is the same decision every one of your competitors made — and why nobody is pricing the thing that makes it dangerous.
Start with the price.
A year ago, using a frontier model at volume was a serious line item. It was a board conversation. You measured it, you argued about it, you had a budget for it. Then the price fell out of the floor. By August this year the effective price American businesses were paying per million tokens was sitting around sixty-eight cents, down from a dollar-fifteen in March. Roughly forty percent gone in five months.
When a price collapses like that, the strategic answer becomes obvious in every room at the same time. Buy more. Bolt it into everything. If your competitor is doing it and you are not, you look slow.
So everyone did the same thing.
Now picture a mid-sized insurance brokerage in Sandton. Fifteen people. Good book of business, decent margins, no engineering team to speak of. Eighteen months ago they put a chatbot on the website. Then a tool that reads policy schedules. Then a claims triage assistant. Today the quotes are drafted by one model, the client emails summarised by another, the underwriting checklist generated by a third, and the receptionist’s inbox is being triaged by a fourth — all of it reaching back into the same cloud, all of it wired in through the same three or four platforms.
They didn’t build that because they were careless. They built it because every one of those tools was rational on the day they bought it. And every broker in the country did the same arithmetic and reached the same answer.
That is what a monoculture looks like. Not one company doing something stupid. A thousand companies doing something sensible, in the same direction, at the same time.
Here is the part that should make you sit up.
There are maybe four model families that matter for real work. A handful of agent frameworks underneath them. And now — this is the connection almost nobody made — a small number of identity providers issuing the credentials those agents use to act on your behalf. In August, Okta opened the door for agents to register as first-class identities in the same directory as your human employees. That was sold as a security upgrade, and it is one. It is also concentration. Every actuator in the economy now points back at the same accounts desk.
And the volumes are not small. Enterprise agent deployments roughly doubled in a single quarter — the modal company went from somewhere in the twenties or thirties of live agents to somewhere in the high seventies. Meanwhile the share of those agents actually being monitored barely moved. So the thing we just standardised across the entire economy is also the thing almost nobody is watching.
We have seen this film before. In July 2024 a single bad update from one security vendor took down airlines, hospitals, banks and airports in the same hour. Nothing was hacked. No enemy was in the building. The system simply had one throat, and someone squeezed it. That was a weekend.
Now point the same architecture at judgment instead of uptime.
Because that is the real difference this time. That outage broke your check-in system. This one breaks your opinion.
If every loan decision in a market runs through the same three models, trained on similar data, calibrated to similar averages, then the errors are not independent. They correlate. The lender who thinks they diversified because they bought two vendors is running the same brain twice. When the model has a bad week — a version update, a poisoned evaluation, a quiet drift in how it reads a particular kind of applicant — it does not go wrong for one bank. It goes wrong for all of them, on the same day, in the same direction.
This is the oldest risk in finance wearing new clothes. When every fund holds the same position, the market looks calm right up until the moment everyone tries to leave at once, and then there is no other side of the trade. Concentration is comfortable for years and catastrophic for an afternoon.
So what does an operator actually do about it?
Three things.
First, stop counting tools as diversification. Two different logos on the invoice is not two different brains, and it is definitely not two different failure modes. Ask the harder question: if this model had a bad week, what else in my business would break on the same morning? The answer to that is your real exposure.
Second, keep a fallback that is actually alive. A second provider you have never once run in production is a slide, not a plan. Push a small slice of real work through it every week. Keep the workflow written down somewhere you own, in plain language, so that swapping the engine underneath it is a Tuesday afternoon job and not a quarter-long project.
Third — and this is the one that actually holds — the thing that cannot be cloned is not your stack. It is the data you have that nobody else has, the way you do the work nobody else does, and the judgment sitting in the heads of the people you kept. Every model on earth can be rented by your competitor tomorrow morning. Your claims history, your relationships, your scar tissue — that is not on the shelf.
Because here is the thing hiding under all of it. The efficiency everyone is chasing is real. The output genuinely is better than it was two years ago. The problem is that when a whole industry standardises on the same answer, that answer stops being a competitive advantage and starts being a single point of failure wearing a consensus.
You did not diversify anything. You joined the crowd.
The question was never how good the model is. It is who else is standing exactly where you are standing, and what happens the day the floor moves.
You’ve got this.