Most Enterprises Shouldn't Train Their Own Models

Every company wants to own its intelligence.

The fashionable answer is to own the model: run an open model, train it on proprietary data, and stop depending on someone else's API.

The instinct makes sense, but it puts ownership in the wrong place.

A stronger model will arrive. The real ownership test is whether it can inherit your best processes, your hard-won exceptions, and every lesson from production. If it cannot, you owned a snapshot.

The intelligence is in the loop

A model does not resolve a support ticket, approve a claim, or review a contract by itself. A system does.

That system decides what the model sees, which tools it can use, what it remembers, when a person should step in, and how the answer is checked. Engineers often call this surrounding system the harness.

The model supplies general capability. The harness turns that capability into dependable work.

Evidence is starting to bear this out. In a 2026 study of coding agents, researchers kept the model fixed and improved the system around it over ten iterations. Performance rose from 69.7% to 77.0%. The improved harness also made three other model families better. Most of the gains came from tools, memory, and the software coordinating them. Rewriting the instructions alone made performance worse.

Another project, GEPA, studied records of what an AI system did, diagnosed its failures, and improved its instructions. On a set of six tasks, it outperformed reinforcement learning on five while using up to 35 times fewer attempts.

These results do not mean that prompts always beat training. They show how much intelligence can live in the system around a model, and how improvements there can survive when the underlying model changes.

Training is not wrong. It is late.

In controlled experiments, reinforcement learning helped systems handle unfamiliar variations better than training on examples alone. In a legal-agent project, Harvey and Applied Compute used reinforcement learning to raise the share of individual requirements passed from 85.3% to 91.3%. The share of answers that passed every requirement more than doubled.

But the most instructive part is what happened first.

The team validated its evaluator against more than 2,500 criteria. It made evaluation cheaper. It repaired broken tool calls, improved tool descriptions, limited noisy output, and managed the agent's context. Only then did it train the model.

The harness made better strategies available. RL made their execution more reliable.

Training baked a proven strategy into the model. The understanding came first.

That is the right sequence for most enterprises:

Use the system to discover what works. Train the model after the behavior becomes stable.

What most companies are missing

Most companies do not have a model-training problem. They have a learning-loop problem.

An agent makes a mistake in production. Someone notices. A prompt gets patched. A few examples are tested. The change ships. Three weeks later, another fix quietly brings the old problem back.

The company was busy, but it did not learn.

A real learning loop looks different:

1. Capture production failures with enough context to understand what happened.
2. Find the cause, whether it was the model, instructions, tools, memory, or workflow, and turn the failure into a test.
3. Make a versioned change, replay realistic cases, and require a person to approve high-impact updates.
4. Ship the change and watch production to prove the fix held.

This is where company-specific intelligence accumulates.

Over time, the test suite becomes a record of what the company has learned and a guard against learning the same lessons the hard way.

Training scales your judgment

Reinforcement learning does not decide what a business should value. It gets better at producing whatever the business rewards.

A support agent rewarded for speed can close tickets before customers are satisfied. A claims agent rewarded for avoiding false approvals can reject everyone. A sales agent rewarded for conversion can make promises the company should never keep.

Research on reward-model overoptimization demonstrates this problem in a controlled setting: optimize hard enough against an imperfect score, and the score can keep rising while the real result gets worse.

If a company cannot write a trustworthy evaluation for a task, it is not ready to write a reward function for it.

Training becomes compelling after the workflow is stable, repeated, and well measured. At that point, it can make a proven behavior faster, cheaper, more consistent, more private, or newly possible.

Before that, training is an expensive way to harden assumptions.

Own the learning loop

The most valuable AI system is the one that can encounter a new failure on Monday, turn it into a test on Tuesday, and ship a safe fix on Wednesday. It should not have to learn the same lesson twice.

Model choice will keep changing. The durable advantage is a system that turns each important failure into a lesson the next model can inherit.