Sovereign intelligence

Every company that uses AI is making a decision about ownership, whether it knows it or not. Each time a model works on a company's records, something is learned. The question is who keeps it.

Two things that get bundled together

The first is the thinking: the general ability to read, write and reason that large models supply. The second is the learning: what a company knows about its customers, prices and exceptions. A distributor may have spent years learning which customers pay late and which products sell together in the rainy season. As agents take on the work, that know-how moves into rules, memory and trained models. The terms on which it lives there matter.

The thinking is becoming a commodity. The learning is what separates you from your competitor.

The power plant objection

Early factories generated their own electricity, then almost all bought from the grid. Why should intelligence be different?

Because of what flows back. Electricity flows one way; nothing the factory does teaches the power station about its business. With AI, what is learned from your records can flow on: into general models, into products that serve competitors, into a vendor's bargaining power at renewal. Even where contracts forbid training on your data, the prompts, rules and corrections that make AI useful tend to live inside the vendor's system, in forms you cannot easily take away.

The thinking is like electricity. The learning is like the design of the production line.

What owning your intelligence means

• Your data stays yours. Stored where you can see it, and exportable.
• Your learning stays yours. The rules, corrections and memory go with you if you leave.
• Your dependence is a choice. Using a large outside model is a deliberate decision, recorded and budgeted.

Why small models change the economics

Most AI work in a business is routine: classifying a ticket, drafting a standard reply. Small AI models trained for one job can do much of it. By Kintoro's estimates, 1,000 agent actions cost about $1 on a small trained model, about $13 on large general models alone, and about $2.30 on a mix of 80 per cent small and 20 per cent large.¹ The trained small models are still being built at piMonk and today's apps use a mix. But routine work on small models you control costs less and exposes less.

That is why piMonk runs on servers we own and operate. Every use of a large outside model is budgeted and on the record. piSA handles outreach without any outside model seeing a customer's prospect data.

The case against

Large general models are better at hard tasks. Running servers is hard. And sovereignty can become an excuse for isolation.

So we argue for choice, not abstinence: send the hard minority of tasks to the best model, knowingly and on the record. Running the servers should be the vendor's job, not every customer's. Sovereignty means choice, not refusal.

Honesty means stating limits. We do not promise a hosting location or a choice of region; we tell you exactly where your data is hosted before you sign. In-country hosting or customer servers are options to discuss, and piSentinel already runs on customer servers under the customer's keys.

Beyond the company

Governments across Africa, the Middle East and Asia are asking where their citizens' data lives. If the learning from their businesses flows only into systems owned elsewhere, those economies will rent back their own knowledge at a price set by someone else.

What this means for a mid-market leader

Where does my data go, and which models see it? A vendor should answer app by app.

If I leave, what do I take? Only records, or also the rules and memory that made the system useful?

Which tasks need a large general model, and at what cost? If the answer is "all of them", ask how the price will hold.

The thinking can be rented, and there is no shame in that. The learning is the accumulated intelligence of your business. Keep it.