The Model Is the Wrong Thing to Own

Everywhere you look, enterprises are being told to own their intelligence.

I agree. But too many companies have mistaken owning intelligence for owning a model.

For many teams, ownership now means running an open model, training it on proprietary data, and escaping someone else's API. That is one way to own a model. It is not how a company owns intelligence.

An enterprise should own the part no model provider can supply: the knowledge of how its business actually works. It should build a loop that makes that knowledge compound.

Models keep leapfrogging

Frontier capability isn't improving in small annual steps. On METR's software-task benchmark, the trend is exponential. The length of task a frontier agent can complete with 50% reliability has doubled roughly every six to seven months.

That doesn't mean models can suddenly do every job. It does mean that work that looked out of reach a year ago can become ordinary surprisingly quickly.

The economics are moving just as fast. Frontier models aren't always cheaper in absolute terms. Better reasoning can cost more, and companies spend more as they give models more work. But the cost of a fixed level of capability keeps falling.

Stanford's AI Index found that the price of querying a model with GPT-3.5-level performance fell more than 280-fold in about 18 months.

Intelligence isn't cheap. A unit of useful intelligence is getting much cheaper.

For a growing share of enterprise work, raw model capability is no longer the main bottleneck. The harder problem is teaching a general model how work actually happens inside a particular company.

Own what model companies cannot

Models can know how claims usually work. They cannot know why your company approves one exception and denies another.

They can understand contracts. They cannot know which terms your legal team will concede, which risks matter most, or how those decisions change by customer.

They can talk to customers. They cannot know which answers create trust in your market and which technically correct answers cause people to leave.

That knowledge is earned through use. It lives in the unusual cases, the human corrections, the policies nobody wrote down, and the outcomes that reveal whether an answer was actually good.

Enterprises should own that accumulated understanding in a form that can survive any one model:

1. Record the work, the decisions, and the outcomes.
2. Capture the corrections and edge cases that reveal how the business really operates.
3. Turn recurring lessons into reusable examples, instructions, checks, and workflow changes.
4. Bring that knowledge to each new model and measure whether it performs better.

Suppose a claims agent recommends denying a borderline claim. An expert overrides it because a state-specific rule changes the answer. That reasoning becomes a reusable example and a quality check, so the next model inherits the lesson instead of repeating the mistake.

Examples, instructions, decision rules, and quality checks can move from one model to another. Fine-tuned weights generally cannot cross model families.

That is the loop to own: an expanding body of knowledge about how the company works, carried forward as models change.

Training is usually the wrong first move

Training can be powerful. The problem is timing.

There are also good reasons to run your own model. Some data cannot leave a controlled environment. Some workloads need predictable latency or cost. A narrow task repeated millions of times may be cheaper on a smaller, specialized model. Those are legitimate deployment and economic decisions. They do not change what the enterprise should accumulate: domain knowledge that survives whichever model runs underneath.

An enterprise can spend months teaching today's model a behavior that the next frontier release performs out of the box. It can also encode a flawed process before it has enough evidence to know what good looks like.

In a legal-agent project, Harvey and Applied Compute used reinforcement learning to improve performance substantially. But first they validated how quality would be judged against more than 2,500 criteria, repaired tool failures, improved instructions, and managed the information available to the agent.

The company learned what worked before asking the model to learn it.

Research points in the same direction. In a 2026 study of coding agents, researchers held the model constant and improved the software, tools, and memory around it. On Terminal-Bench 2, a benchmark of 89 real-world command-line tasks, first-attempt success rose from 69.7% to 77.0%. That 7.3-point gain eliminated roughly a quarter of the remaining failures. The finished system then improved three other model families by 5.1 to 10.1 points without being redesigned for them.

GEPA studied records of AI behavior and improved the system's instructions. It outperformed reinforcement learning on five of six tasks while using up to 35 times fewer attempts.

None of this proves that instructions always beat training. It shows that much of a company's advantage can live in knowledge and workflows that are easier to inspect, improve, and carry forward.

Training makes sense once a behavior is stable, repeated at scale, and measured against an outcome the business trusts. At that point, putting the behavior into the model can improve speed, cost, privacy, consistency, or capability.

The inheritance test

The durable asset isn't a particular model. It is the knowledge that makes any capable model valuable inside the company.

When a stronger model arrives, can it inherit the company's best examples, edge cases, decisions, and ways of judging quality? Can it become useful without forcing the organization to start learning again?

If so, the company owns its intelligence. If not, it owns an implementation with an expiration date.

Model providers will keep competing to make general intelligence more capable and less expensive per unit of useful work. Enterprises should spend their energy making that intelligence understand their world.

Own your domain knowledge. Let the models compete.