There's something true about AI models that almost nobody selling them wants to say out loud: they're all catching up to each other.
The gap between the best one and the second best is measured in months, and it closes. Whatever one of them ships shows up in the others within a quarter. The benchmarks keep getting beaten by smaller and smaller margins, and what's left is mostly personality and price.
Which means if your whole advantage is that you have access to a very good model, you don't have an advantage. You have a subscription — and so does your competition, at the same price, with the same raw output.
This has happened before
Every industry that ever got industrialised went through the same thing.
Electricity was once an advantage: factories that had it out-produced factories that didn't. Then everybody had it, and the advantage moved to what you built with it — how the floor was laid out, how the process was designed, how good the people running it were. Compute went through it. Bandwidth went through it. Cloud storage went through it. Every time, the thing that was scarce got cheap, and the value moved upward into everything arranged around it.
Intelligence is on that same curve — and that's about supply, not about how smart the models get. Whichever company reaches the ceiling first, several others arrive within months, and then they compete on price — which is what happens to anything you can't tell apart.
So what actually decides it
Everything wrapped around the intelligence. The harness is the part that decides whether any of it pays off.
The engine is the easy part. The advantage lives in everything that turns a huge amount of force into work that goes in a useful direction. It's unglamorous by nature, and it's six things — none of which is about being smarter.
- Memory — it knows your business, your history, your decisions, your preferences, and it doesn't start from zero every time you open it.
- Scope and permissions — it can reach what it's supposed to reach, and it can't touch what it isn't.
- A check — something looks at the output against what was actually asked for, before you see it.
- Project management — it knows what the job is, what state the job is in, and what comes next.
- The loop — it can hold a task across days instead of one reply.
- The handoff — work moves between steps without a person carrying it every time.
Nothing on that list is about reasoning ability. Every single item is about being organised.
Every session starts from zero. It doesn't know your business, it can reach anything, and nothing checks the work. Fast, confident, and quietly wrong.
It remembers last month, it knows its limits, a check catches the bad output, and the job keeps moving while you sleep.
Why the check is the whole ballgame
The biggest difference between an AI system that's genuinely useful and one that's an expensive toy is whether anything checks the work.
An unchecked system doesn't fail loudly. That's what makes it dangerous, and it's a different kind of dangerous from ordinary software — ordinary software at least has the decency to throw an error when it breaks.
An unchecked system produces something plausible, at speed, with total confidence. You accept it because it looks finished. Three weeks later you find the section that contradicts the other section, the number that was invented, the requirement that quietly disappeared somewhere in the middle. And you don't find it because a system told you. You find it in a meeting, or in a customer complaint, or when the thing doesn't sell.
An unchecked system doesn't fail loudly. It fails convincingly — and that's the expensive kind.
So the question isn't how smart it is. It's what checks it, and against what. If the answer is that you do, by reading everything, then you haven't removed the work. You've moved it onto yourself and added a machine that generates more of it.
The test you can actually run
There's one question that separates a real system from a subscription, and it takes ten seconds.
What does it remember about my business from last month?
If it knows what you decided in March and why, if it knows the offer you killed and the reason, if it picks up on Monday where you left off on Friday — you're buying a harness. That's an asset. It gets more useful the longer you have it, and it doesn't leave when a subscription lapses.
If the answer is nothing, if every conversation starts clean, you're buying a subscription. It's the same subscription your competitor is buying, and it will be exactly as good for them as it is for you.
- Intelligence
Table stakes. Everybody can buy it, everybody is buying it, and it keeps getting cheaper.
- The harness
Memory, limits, a check, a plan, a loop, a clean handoff. This is the part you build or buy deliberately.
- The asset
Only one of the two gets better because you worked. That is the one worth paying for.
The intelligence is table stakes. What you build around it is the business.
