The build vs buy AI decision most owners get backwards

I think most Alberta owners are shopping for the wrong thing when they buy AI, and Harvey and Thomson Reuters just proved it with real money. Here’s where I land on build vs buy AI. You rent the model, the same one any competitor can sign up for. What you own is the layer on top, built from your own jobs and judgment, and that’s where I’d put your money and attention.

Harvey, the legal AI company, told its own story in August. Its biggest tool, Tenet, still runs on a rented base, a model called Kimi K3 post-trained for long legal work. Two of its smaller tools moved somewhere else. Review Table now runs on a model Harvey post-trained itself, and Harvey says it costs roughly a tenth what the frontier version cost per cell. Firm Knowledge, built on an open base with its own retrieval layer, cut cost per query by 90%.

Thomson Reuters made the same bet the same month. It put $40 million into a model called Thomson, started from an open source base instead of a frontier lab’s API. Forty million dollars is not a number any Alberta business will spend on a model, and that is not the point. The point is what the money bought. Thomson Reuters was not chasing a smarter general assistant. It was buying the part built on 175 years of its own legal and tax content, the part nobody else can copy by paying for the same software.

You will never train a model, and the same call still lands on your desk

Nobody reading this is going to fine tune an open source model in a data centre. That part of the story does not apply to a construction firm in Red Deer or a law office in Lethbridge. The part that does apply sits one layer down. Every AI tool you buy still splits into the same two pieces Harvey and Thomson Reuters just paid to separate. One piece is the model, the raw intelligence running underneath. The other piece is everything built on top of your own data, your own cases, your own judgment calls.

The model gets cheaper and smarter whether you do anything or not. GPT, Claude and Gemini improve every few months at a price that keeps falling, because three large labs are racing each other on your behalf for free. That part of your stack is a rental, and renting something that improves on its own is a good deal. The mistake is spending your scarce attention shopping for the best version of it, the way a buyer agonizes over which identical gas station to use.

The part worth owning never improves unless you build it

The layer that compounds is the opposite kind of asset. It is the pricing logic your best estimator carries in his head and nobody wrote down. It is the actual history of jobs your firm has run, the ones that went over budget and the ones that came in clean, and why. It is the audit trail that shows a client or a regulator exactly what an agent touched and when. None of that gets smarter on its own. A frontier lab is never going to improve your own client history for you. You either build it, or a competitor who did build it starts winning the bids you lose.

Illustrative example, not a client case study: picture two insurance brokers in Calgary buying the identical AI underwriting assistant from the identical vendor. One treats the subscription as the whole project and plugs it in. The other spends the first month feeding it three years of its own claims history and writing down the five judgment calls its best underwriter makes that the software cannot see on its own. Both brokers are renting the same model. Only one of them owns anything a rival could not get by signing up for the same tool next week. If I owned either brokerage, I’d want to be the second one.

You can test this split today in how you evaluate any AI vendor pitch. Ask what happens to your account history, your client file and your own prompts if you cancel next month. If the honest answer is that it all stays locked inside their product, you were never building anything of your own. You were paying rent and calling it ownership.

This sits one layer above the file your own team should already own. Owning the workflow protects you if a vendor disappears tomorrow. Owning the judgment behind it is what keeps a rival from copying you even while the vendor stays exactly where it is.

It also changes how you should be paying for AI in the first place. Buying by the finished job instead of the token forces a vendor to prove their rented model is doing the work, rather than billing you for effort a cheaper model could have matched. And the same split shows up in which model actually protects your client data, where the cheapest line on the price sheet is rarely the one worth owning.

Before your next AI purchase, ask which half of it you are actually buying. If the answer is only the model, you just bought what every competitor can buy tomorrow afternoon.

My take: rent the model, and spend your own effort on the part nobody can buy off the shelf.

If you’d like a second set of eyes on which half of your own AI budget falls into each camp, that’s where an AI strategy roadmap starts. I’d have that conversation before the next renewal comes up.

Build vs buy AI, the questions owners ask