In short: At every size, the same rule holds: rent the context layer as it matures, and own the record of how you decide. What changes is what "own" means: a document and a habit at two people, a governed service at ten thousand.
Key points
- Models are rented, data connections use a shared protocol, and search, assistants and agent registries are competitive products.
- Before buying a tool that touches decisions, check it exports every record with history in a documented format.
- Also check the record survives replacing the vendor, and that you control retention and deletion record by record.
- Over-building early encodes guesses that later have to be unwound.
- Under-owning late leaves decision history spread across vendor systems that cannot be queried together or exported.
The buy-versus-build question for AI in 2026 is unusually clear, because the market has sorted itself. Models are rented. Connections to data use a shared protocol. Search, assistants, meeting tools and agent registries are competitive products. None of those is where an organisation's advantage lives.
The record of how it decides is. And at every size, the temptation is either to buy that too, inside a vendor's memory product, or to over-build it before it is needed.
At each rung
| Size | Buy or rent | Own | Don't build yet |
|---|---|---|---|
| Two | A model, transcription, a shared document | The founders' ledger and the weekly review | Knowledge base, taxonomy, integrations |
| Ten | An assistant that can read your documents | The ledger, readable by strangers; first definitions | A wiki organised by topic |
| Fifty | Meeting and writing tools that export | Glossary with owners; decision rights; one home for records | An ontology |
| Two hundred | Workflow tools that can carry a precedent slot | Decision classes, templates, handover packs, first calibration | A platform team |
| One thousand | Enterprise search, agent registry, identity and permissions | Trace specification; function-owned records; evidenced review | A single enterprise ontology |
| Ten thousand | Most of the context layer, at scale and under contract | Central invariants, export format, calibration, the learning layer | Anything that forces divisions into one model |
Three tests before buying anything that touches decisions
- Can every record it creates be exported, with history, in a documented format?
- Does the record survive if this vendor is replaced?
- Do you control retention and deletion, record by record?
If any answer is no, either negotiate it into the contract or keep the record elsewhere and let the tool read from it.
The cost of getting it wrong in each direction
Over-building early costs time, money and, more subtly, accuracy: structures designed before the organisation knows what it needs encode guesses that later have to be unwound.
Under-owning late costs more. An organisation that let its decision history accumulate inside several vendors' products discovers, at the first migration or audit, that its judgment is spread across systems it cannot query together and may not be able to export.
In practice
- Find your row in the table and check you are not building anything in the last column.
- Apply the three tests to every tool that writes decision records today.
- Write export and retention terms into the next renewal of each such tool.
What would prove this wrong. Organisations that rent their entire judgment record inside vendor products, never owning it, and suffer no loss of capability or leverage through migrations, audits and renewals.
Common questions
Should we build or buy AI tools for our company?
Rent the context layer, such as models, search, assistants and agent registries, as it matures. Own the record of how your organisation decides, because that is where its advantage lives.
What should I check before buying an AI tool that stores decisions?
Check that every record it creates can be exported with history in a documented format, that the record survives if the vendor is replaced, and that you control retention and deletion record by record. If any answer is no, negotiate it into the contract or keep the record elsewhere.
What should a two-person startup own versus rent for AI?
Rent a model, transcription and a shared document. Own the founders' ledger and the weekly review, and do not yet build a knowledge base, taxonomy or integrations.
What is the risk of keeping decision history inside vendor products?
At the first migration or audit, the organisation can discover its judgment is spread across systems it cannot query together and may not be able to export.