The model is the part you rent. What it knows about your business is the part you own, and the only part a competitor cannot buy.
Most AI announcements now sound the same. A company has rolled out ChatGPT, Copilot or Claude to its team, and is excited about what comes next. Two years ago that was news. Today it is the starting line. Your competitor pays the same subscription, gets the same model on the same day, and gets the same upgrade when it ships.
So the question worth asking is not which AI you have. It is what you have built around it that nobody else can buy.
The model is the part you rent
The leading models are converging. On the benchmarks Stanford tracks, the gap between the best open model and the best closed one fell from about 8% to 1.7% in a single year.1 The price is falling just as fast: running a model as capable as the original ChatGPT became more than 280 times cheaper between late 2022 and late 2024.1
That is good news, and it changes what is worth investing in. Something that gets cheaper and more alike every quarter is a utility. Choose it carefully, then treat the choice as reversible. If a better model arrives next year, switching should take an afternoon, not a project.
Why the same AI gives different results
Inside most businesses, AI already runs in ten different versions, one per person. Each of them re-explains the company every time they open a chat: what you sell, who you sell to, what you are allowed to claim, how the work is done here. The best of them get excellent results. Everyone else gets something generic, and nothing any of them learns carries over to the next person, or even to their own next conversation.
That is the most expensive way to use AI. It costs the most of your people's time, it produces the least consistent work, and what it produces disappears into a chat history. It is also where most of the disappointment comes from. A widely cited 2025 study found that around 95% of companies saw no measurable return from their generative AI pilots, and put the cause down to tools that do not learn and do not fit the way the work is done, rather than the quality of the models.2
A model knows what everyone knows. Only your system knows what your business knows.
What you can own instead
The value sits in the layer between the model and your work. It is built from four things, and none of them comes with a subscription:
Your products, your customers, the claims you can stand behind and the ones you cannot, written down once and kept current. We call it the company brain.
The judgement that today lives in a few heads, turned into agents anyone on the team can use, so the quality of the work stops depending on who did it.
Workflows that run on request or on a schedule, with a named person approving anything that goes out under your name.
Every result measured and fed back, so the system gets better with use instead of starting from zero each time.
Each of these compounds. A model upgrade makes them all better overnight, because they sit underneath whichever model you use. A competitor can buy the same model tomorrow. They cannot buy three years of your business written down.
Build it so the model can change
If the model is the replaceable part, everything else should be built so it can be replaced without loss. In practice that means three rules.
- Your knowledge lives in files and accounts you own, not inside one vendor's product. If you change models, or change partners, it comes with you.
- Your team keeps the tool they prefer. Claude, ChatGPT or whatever comes next all read from the same source underneath, so the work is consistent whoever is using what.
- Access is earned one step at a time. Every connection to your systems starts read only, and nothing is published, sent or spent without a person approving it. That is what lets you trust the system with more.
If a competitor bought exactly the same AI tomorrow, what would we still have that they do not?
If the honest answer is nothing, it is a subscription, not an investment. That is fine for some things. It should not be your AI strategy.Where to start
None of this needs a large programme. The businesses that get it right usually start small and build in the right order.
- Write down what is true. One place that holds what you sell, to whom, and what you can claim. Everything else reads from it.
- Capture one job the way your best person does it. Pick work your team does every week, and turn the best version of it into something everyone can run.
- Put one workflow on top, and measure it. Prove the return on one thing before you spread it to ten.
Then add the next job, and the next. The model underneath will change several times along the way. What you have built will not need to.
We run our own business this way, on the same Company OS we build for clients. If you want to know what your AI should own, talk to us.
- Stanford Institute for Human-Centered AI, AI Index Report 2025: the open and closed model gap on Chatbot Arena, and the cost of GPT-3.5 level inference from November 2022 to October 2024.
- MIT NANDA, The GenAI Divide: State of AI in Business 2025. A preliminary study, based on interviews, surveys and a review of public deployments.