A founder recently described cutting her company’s AI model costs by 97 percent in a single month. She switched from a frontier model — one of the expensive, headline-grabbing ones — to an open-weight alternative (a model whose underlying parameters are publicly released, so anyone can run it without paying per-query fees). The switch took a weekend. The savings were not marginal.
That conversation has stayed with me, because it reveals a gap that most executive conversations about AI still paper over: cheap intelligence is arriving fast. The ability to use it well is a different problem entirely.
The story the market tells, and the story your operations tell
Right now the market values AI intelligence as though it were scarce — something only a handful of labs can supply, worth enormous multiples. That narrative is not obviously wrong as a short-term financial story. But it runs ahead of what companies actually experience on the ground.
What companies experience is model costs falling, alternatives multiplying, and a growing sense that the bottleneck is not “which model” but “what surrounds the model.” The decision about which AI to use gets made in an afternoon. What takes months — and what most organizations have barely started — is building the structure that makes any AI useful at scale.
What the harness actually is
The term for this is the harness, borrowed from a useful framing encountered in recent thinking. The harness is not the AI itself. It is the organizational layer that makes the AI do anything trustworthy: the context it has access to, the documents and data it can see, the permissions governing what it can act on, the review standards that determine when its output is good enough to ship, the memory of past decisions, the budgets and decision rights, and — crucially — the accountability for what gets produced.
Without the harness, you have a capable but untethered thing. With it, you have a working part of your organization.
The distinction matters because labs can sell you models, tools, and engineers to help with integration. What they cannot sell you is your own operating context or your own judgment about output quality. Those are yours. They are also the hard part.
The question leadership cannot delegate
Here is where organizations stumble, including ones that run genuinely sophisticated AI pilots: they treat the harness as a technical problem and hand it to engineering or IT. Then they are surprised when deployment stalls, or when employees quietly stop trusting the AI’s outputs, or when a compliance question surfaces that nobody has an answer for.
The structural decisions — when is this output good enough to ship? who is accountable if it is wrong? what does this AI have permission to act on? — are not engineering questions. They require judgment about the business, and they require leadership to own them. A CTO can build the system; only the leadership team can decide what the system is allowed to do and what counts as acceptable.
This is not an abstract governance concern. It is the difference between an AI pilot that runs in controlled conditions and AI capability that actually integrates into how the company works.
Two worlds at once
What makes this moment genuinely difficult is that organizations have to operate in two worlds simultaneously.
In the public narrative world, AI intelligence is a scarce, expensive, transformational asset. Valuations are stratospheric, announcements are breathless, and the pressure to be seen doing something with AI is real and relentless.
In the operational reality world, model costs are collapsing, open alternatives are becoming increasingly capable, and the companies that will extract lasting value are the ones that invest in the harness — governance, context, standards, accountability — rather than in the chase for the most impressive model.
These two worlds are not impossible to hold at once, but they require a deliberate choice about where you focus. Chasing the headline model while neglecting the organizational structure is how you end up with expensive AI that nobody trusts and nothing changes.
The scarce asset may not be what you think
The outlook for where this goes is not pessimistic. The arrival of cheap, capable intelligence is genuinely good news for organizations willing to build the structure to use it. The ones that do will find that the competitive advantage does not come from the model — everyone has access to roughly the same models — but from the harness they have built around it.
That harness, once built, is not easily copied. It reflects years of operating context, accumulated judgment about quality, and organizational decisions that outsiders cannot simply replicate. In a world where the intelligence itself becomes a commodity, the harness may be the thing that actually holds lasting value.
The question worth asking now is not “which AI should we buy?” It is “do we understand the structure we need to build, and are we the ones building it — or is it being built for us by default?”
Text summarized and optimized using Anthropic’s models and reviewed by a human.