Quick answer: This is the written version of the talk I gave at MAIA 2026, the international conference on Management in the Age of AI held by the Institute of Management and Quality Sciences in Wroclaw. The argument is simple. When you rent intelligence from a handful of very large providers, you also rent the part of your company that makes judgments, and accountability leaks out of the building with it. Small models you own and run on your own hardware give that accountability back, and for most business work they are good enough.
Key takeaways
- Renting is not neutral: the reasoning behind your decisions lives on someone else's machine, under terms they can change.
- Three things leave quietly: the audit trail, continuity when a model is retuned or retired, and control over the data relationship.
- Small models are not worse at everything: they are worse at some things, and most business work is narrow and repeatable.
- Owning has a real price: you take on operations, you accept a capability ceiling, and you need someone who finds this interesting.
- It is a management decision: the question is where accountability sits, not which vendor scores highest on a benchmark.
Renting intelligence means renting judgment
Most companies adopting AI right now are renting it. You send your context to an endpoint, a model you cannot inspect returns an answer, and that answer shapes a decision that someone inside your company then owns. The convenience is real, and I understand completely why it wins. What I want to point at is what moves out of the building along with the compute.
A model is not a spreadsheet formula. It takes part in judgment. When it drafts the reply to a customer, scores a supplier, or decides which application gets read first, it is doing work a person used to be answerable for. If that work happens somewhere you cannot see, on a version you did not choose, under terms that can change with a release note, then your accountability has quietly become a promise made by a vendor.
Three things you hand over without noticing
The first is the audit trail. When a regulator, a customer, or your own board asks why a decision was made, "the model said so" is not an answer. Owning the model does not hand you a good explanation by itself, but it does give you the weights, the prompt, and the exact version that produced the output, which is where any real explanation has to begin.
The second is continuity. Rented models are deprecated, retuned, and silently improved. A process you validated in March can behave differently in September without anyone telling you. If your quality assurance rests on a model you do not control, your quality assurance has an expiry date you cannot read.
The third is the data relationship. Even under good terms, you are routing your customers, your suppliers, and your unreleased work through infrastructure you do not own. For a small European company working under the GDPR, that is not an abstract worry. It is paperwork, it is risk, and it is a conversation you would rather not have to have.
But the small models are worse
At some things, yes. That is the honest starting point and I have no interest in pretending otherwise. A small model running on a desk in Munich will not out reason a frontier system on a hard, novel problem.
The useful question is a different one. What are you actually asking it to do? In our company the answer is mostly narrow and repeatable: classify an incoming email, pull structured fields out of a supplier document, turn a text into an embedding so it can be found again, describe an image, draft a first version of something a person is going to edit anyway. On work of that shape, the distance between a frontier model and a good open weight small model is far smaller than the marketing suggests, and most of the time it is invisible.
What owning actually looks like
Ours is not a data centre. It is three Macs sitting in an office. One machine holds the embeddings and the search index, another does the vision work, a third runs the chat model that most of our internal tools call, and a small voice service reads text aloud in my own cloned voice. The models are open weight and small enough to live in memory on consumer hardware.
The thing I would tell any founder considering this is that the win was not the cost, although the cost is lower. The win is that nothing leaves, nothing changes underneath us, and when something is wrong we can go and look at it.
The honest cost of owning
You take on operations. Models fall over, memory runs out, a machine needs restarting at an inconvenient hour, and there is no support line to call. You accept a capability ceiling and you have to design your work around it rather than wish it away. And you need at least one person who finds this interesting in its own right, because otherwise it rots.
I would not recommend this to a company that wants AI to be invisible infrastructure it never has to think about. I would recommend it to a company that has decided its own judgment is part of what it sells.
Why this belongs in a management conversation
The reason I made this argument to a management conference rather than a technical one is that the choice is not really about model quality. It is about where accountability sits. Every organisation has to be able to say who is answerable for a decision, and when the reasoning behind decisions moves onto a rented system, that line gets blurry. Blurry lines have a habit of becoming somebody's problem later, usually at the worst possible moment.
Own it, do not rent it. Not because owning is fashionable, but because you cannot delegate accountability to a vendor, however good the vendor happens to be.
This is the written version of the talk delivered on 13 July 2026 at MAIA 2026, Management in the Age of AI, hosted by the Institute of Management and Quality Sciences in Wroclaw, Poland. The certificate of presentation and the session details sit on the speaker page, where you can also ask for the slides.