own the loop, rent the model

jun 2026

Over the past year, much of the conversation around enterprise AI has shifted toward agents and the skills they can perform. Companies are building agents that research, analyse, design, sell and increasingly operate across entire workflows.

Many of these skills will create real value. But I suspect the focus on skills confuses what an agent can do today with what will still make a company different when everyone can do it.

I began thinking about this earlier in the year. In January, I wrote about the exposed company brain and the possibility that organizational judgment could begin compounding instead of repeatedly resetting. By April, in from LLM to harness to application, the argument had become more specific: models bring general capability, while applications own the recurring context in which people create, correct and approve real work.

Models will be available to everyone. Tools will become standardized and most agent skills will eventually be copied, packaged or absorbed into the platforms around them. What will remain harder to reproduce is the knowledge loop surrounding those skills: the accumulated memory of how a company sees, decides, corrects and improves.

This is why Satya Nadella's recent post resonated with me. He describes a future in which human capital and "token capital" compound through a learning loop. The opportunity is not simply choosing the best model, but building a system in which the knowledge of a company and the intelligence it uses grow together.

The "reverse information paradox" adds a question of ownership. To make external intelligence useful, a company must reveal its context, explain its exceptions and correct the system when it is wrong. The better the intelligence performs, the more the company has taught it.

But where does that new knowledge come from?

It is created wherever the company meets reality.

Friction exists wherever an internal assumption encounters something it cannot fully explain. It appears when an AI recommendation meets the judgment of an experienced person, or when short-term targets conflict with the longer-term interests of the company. It also appears at the company's edge, when a new product feature encounters actual customer behavior, or when client-facing teams hear the hesitation, objections and changing needs that never appear in a dashboard.

These moments are not simply obstacles to overcome. They are new information.

Models learn from existing knowledge. Companies learn from friction.

The people working in these moments are more than people completing tasks. They are the company's sensory system, experiencing the difference between its internal understanding of the world and the world as it actually is.

Friction is where the company discovers what it does not yet know.

Most companies capture only a small part of this. Sales objections are reduced to CRM fields. Support conversations become ticket counts. Local nuance is compressed into regional reporting. The outcome may be recorded, but the reason behind it disappears.

This creates an important paradox for automation. Companies naturally want to remove friction, but if they automate an interaction without capturing what caused it, they may remove part of their own sensory system. The work becomes more efficient while the company learns less.

Repeated friction should become a skill. Novel friction should become knowledge.

This leaves companies with three useful questions.

1. Which frictions will AI remove in the next three months, six months or a year? You probably do not want to spend years building around work the models will soon perform for everyone.
2. Which frictions become more valuable as AI improves because they reveal something the machine cannot yet see - customer hesitation, local context or experienced human judgment?
3. When those moments occur, does the learning return to a loop the company owns, or does it disappear once the immediate problem is solved?

The aim is to automate the friction that repeats while staying close to the friction that teaches.

Owning the loop means ensuring those interactions, corrections, decisions and outcomes return to the company and improve what happens next. It does not mean building a foundation model. Most companies should use the best intelligence available and replace it when something better emerges.

The model should be replaceable. The company's learning should not be.

Rent the model. Own the loop.