Every few weeks a new AI model tops the leaderboards, and the question we hear from businesses follows right behind it: which one should we be using? It is a fair question. It is also, most of the time, not the one that decides whether AI does useful work for you.
What usually decides it is the harness — everything wrapped around the model. For a business trying to get real work done, the harness matters more than the model, and the AI harness vs model debate is worth settling before anyone signs up for anything.
What an AI harness actually is
A model on its own is a remarkable engine for reading and producing language. It does not know anything about your business, cannot see your documents, cannot look anything up and cannot act. Everything that turns it into something useful sits around it: the instructions it is given, the context it can see, the tools it is allowed to use, what it remembers between tasks, and the checks that catch its mistakes before anyone relies on them.
That surrounding system is the harness. Put the same model into two different harnesses and you get what feels like two different products — one that answers confidently from nothing, and one that finds the right document, uses it, and shows its work.
Why the wrapper decides the result
When a business tells us AI got something badly wrong, the model is rarely the culprit. The assistant was looking at an outdated document, or at nothing at all. It had no way to check a figure it was unsure of. Nobody told it what a good answer looks like for this particular customer. Those are harness failures, and swapping in a newer model fixes none of them.
The reverse holds too. A capable but unremarkable model with current, well-organised context, the right tools and a check on its output will beat a leaderboard-topping model working blind on the tasks businesses actually care about, almost every time. Most of the quality lives in what the assistant can read, not in the engine doing the reading.
A brilliant model that cannot see your business is still a brilliant stranger.
What to ask when you are buying
If you are evaluating an AI product or assistant, the useful questions are almost all about the harness. What can it see, and how current is that information? What can it actually do — search your documents, read a calendar, update a record — and with what permissions? How is its work checked before a person or a customer relies on it? And can the model underneath be swapped when a better one arrives, or are you tied to one provider's engine?
That last question matters more than it sounds. Models are improving and changing hands quickly, and a system built so the engine can be replaced is one you will not have to rebuild every year. A harness that only works with a single model is vendor lock-in with extra steps.
Where the model still matters
We would be overstating it to say the model never matters. For long, reasoning-heavy work — untangling a complicated contract, planning across many moving parts, writing and checking code — the stronger models are genuinely better, and it shows. The gap is real at the edges.
But it narrows quickly, and it is rarely where a business is losing value today. Most are losing it in the plumbing: context that is stale, tools that are missing, output nobody checks. This blog is a small example. The cover illustrations here are generated and then checked against a written standard before any of them publishes, and it is the checking that keeps them consistent, not the choice of image model. The same thinking is why we expect the future to belong to sites and systems that agents can connect to properly, rather than to ever-bigger engines guessing from the outside.
So the next time someone asks which model to use, ask a different question first: what is it going to be able to see, do and prove? That is the part worth getting right, and it is the part we spend our time on — see our method.
Keep reading
