The model can be rented. The world it works in cannot.

Every company can rent the same models today. The API that answers a company in Antwerp answers its competitor in Austin at the same price, and that price falls every quarter. So the intelligence itself cannot be where a company’s advantage lives. What lasts is the maintained model of your own business and the market it sells into: who the companies and people are, what is true about them now, what happened before, how they relate, where each fact came from, who may see it and what came of the work. That is the world the rented intelligence works in, and no two companies have the same one.

Intelligence is a utility now

A model is priced like electricity and sold like electricity. Everyone gets the same one, the meter runs and the rate goes down. By a16z’s count, GPT-3-quality inference cost $60 per million tokens in November 2021 and $0.06 by late 2024, about 10x cheaper every year. At the deep-tech company where we run agents today, the team reaches them through Slack, through MCP inside Claude and through a chat UI. Behind all three sits a model nobody at that company owns, trained on nothing about their territory. When a better one ships, we switch, and the meeting-prep agent writes a sharper brief from the same calls, emails and deal history it read yesterday.

That is the point. A better model improves what the company has. It does not replace what was built around it. The brief gets better because the graph it reads from was already there: the accounts, the contacts, the history of each deal. Take that away and the smartest model in the world writes a confident brief about the wrong hospital.

What you own is the model of your territory

Four things stay when the model changes.

The graph. That company’s partner list was a dirty spreadsheet of about 2,700 companies. We built the territory instead from government sources, hospital lists and integrator registries, normalised it and linked it to their CRM. It collapsed to 64 ranked accounts, 48 of them with contacts already pulled. That territory is now an asset the company holds. A new rep can say “new SDR, West Coast territory, make me a list like this one” and get one on day one. The list was never in the model. It was drawn from public sources and the company’s own history, and it stays with the company.

The procedures the agents learned. The platform did not pull phone numbers natively, so we added a phone-enrichment API as a tool the agent calls before it writes to the CRM. That step is now part of how their outbound works, whichever model runs it.

The evaluations. Before we built anything we agreed the standard on real examples, including the cases that need a human. Those examples are the test the agent has to pass every time the workflow or the model changes.

The corrections. Reps give feedback in the shared Slack channel, and it changes what the agent produces next. Each run leaves a trace on the graph, so the next one starts where the last stopped. Every month of that is a month of the company teaching its agents how it works. None of it is in the model provider’s weights.

Adoption is the first advantage. Learning is the one that compounds

The immediate advantage is adoption: redesigning the work and putting agents into it. That is real. At the deep-tech company the top-50 list refreshes at 7am so the rep has a fresh list by 9am, and the meeting-prep agent has run before he opens his laptop. A competitor without that is slower every day.

But adoption can be copied. The competitor rents the same model and hires someone to wire it to their CRM. What they cannot copy is what the company has learned in the meantime, and the world model is where that learning lands. A note about a regulation attaches to the regulation, not to the call where it came up. A correction to a contact’s role attaches to the person. At a planning-software company we learned that on GitHub, commits and forks from several people at the same company are a strong sign they run an end-of-life planner internally, and that stars alone are weak. That finding lives in the pipeline that builds their migration list, not in a chat transcript that will be gone next week.

Learning that lands on the things it is about starts to connect. A changed regulation lands on the facilities it applies to, and from there reaches the open opportunities and the people who must act. That is what sight means, and it only appears when learning accumulates in one place. A company that runs its agents in chat learns nothing that lasts. The corrections die with the conversation.

Build the model of your business and rent the intelligence

For the buyer, this decides where the money goes. Most AI products ask you to pay per seat and per token to work inside their model of your business, and whatever accumulates there accumulates for them. If you leave, you take an export.

Do the opposite. Build the model of your own business, in a form your agents and your people can both work from, and rent the intelligence on top of it. At the planning-software company the team queries their world model from inside Claude by saying “use Ortelian” instead of the CRM connector. The intelligence is Claude’s. The world it is answering about is theirs. When Claude is replaced by something better, the sentence still works. It is also why we charge a deployment fee and a monthly platform fee and nothing per seat or per token. The thing we maintain is the world, not the meter.

The obvious objection is that the labs will build this too. They will not, because they sell the model. Their business is to make it better and cheaper for everyone at once, which is exactly what makes it a utility. What they add around it, memory of your chats and connectors to your tools, is memory of a user and access to a system. It does not hold the maintained fact that this integrator supplies these three hospitals, that this regulation applies to these facilities and touches these open deals, that this rep corrected that contact’s title last week. That world is yours and it is different for every company. Nobody else can maintain it, because the facts in it come from your work.

The model can be rented. The world it works in cannot.