- Unowned intelligence is AI capability a company relies on but does not control: prompts in a vendor's console, workflows inside a SaaS tool, logs and feedback that never leave the provider.
- Foundation models are a commodity input that improves and gets cheaper every year. The lasting value is in your context, your evals, your workflows and your data about what worked.
- Ownership means those assets live in your repositories and your infrastructure, in portable formats, with your team able to change them.
- A simple test: if the vendor disappeared or doubled its price tomorrow, how much of your AI capability would you keep?
This is part 3 of Enterprise AI. Parts 1 and 2 covered cost and context. This part is about a quieter problem that decides who benefits from all the AI work a company does: who owns it.
01What is unowned intelligence?
It is AI capability your company depends on but does not control. It accumulates in ordinary ways:
- Prompts written and tuned in a provider's web console or a SaaS tool's settings page, with no version history in your repository.
- Workflows built inside a vendor's automation builder that only runs on that vendor.
- Knowledge uploaded into a product's proprietary index, which cannot be exported with its structure and permissions intact.
- Usage data and feedback (which answers people accepted, edited or rejected) that stay in the vendor's analytics, if they are kept at all.
- Evaluation data, the hard-won set of examples that defines “good”, scattered across spreadsheets or never written down.
Each one is a reasonable shortcut at the time. Together they mean the company is renting its AI capability while believing it is building one.
02Why does it matter?
Because of where the value actually sits.
Foundation models are a fast-moving commodity input. New models arrive every few months; prices per token keep falling; today's frontier is next year's mid-tier. Tying your capability to one model or one vendor's interface means tying it to something that is about to change.
What does not depreciate is everything you build around the model:
| Asset | Why it compounds |
|---|---|
| Context: connectors, indexes, business rules | Encodes how your business works |
| Evals: examples of correct behaviour | Lets you switch models safely and measure progress |
| Workflows: where AI sits in real work | Determines adoption (see put the model inside the workflow) |
| Feedback: accepts, edits, rejections | The best training and eval data you will ever have |
| Logs and traces | Cost, quality and audit history |
If those live with a vendor, the vendor captures the compounding. If they live with you, every model improvement becomes your improvement.
03What does AI ownership look like?
Ownership is practical, not ideological. It means:
- Prompts and tool definitions in your repository, versioned, reviewed in pull requests, tested in CI.
- Evals in your repository, run against any model you are considering.
- A model-agnostic seam in your code, so swapping providers is configuration, not a rewrite (see avoid vendor lock-in).
- Your own logs and traces, in your observability stack, with your retention rules.
- Context in your infrastructure: your index, your connectors, your permission model.
- Feedback captured by your product, stored as your data.
task: invoice.extract
owner: finance-platform
prompt: prompts/invoice_extract.v7.md # versioned in git
schema: schemas/invoice.json
eval: evals/invoice_extract.jsonl # 180 labelled real invoices
quality_bar: { field_accuracy: 0.97, totals_exact: 1.0 }
routing: { primary: small-structured, escalate: mid-tier }
retention: { inputs: 30d, outputs: 1y }A registry entry like this is boring, and that is the point. Everything that defines the capability is a file your team owns.
04Does this mean never buying from AI vendors?
No. Buy models: training frontier models is not your business. Buy good infrastructure where it fits. Use enterprise chat tools for people who need an assistant for open-ended work.
The line to draw is simple: rent the models, own the intelligence. Anything that encodes how your business works, or how you measure whether AI is doing it well, should be yours.
If the vendor disappeared tomorrow, what would you still have? That answer is the AI capability you actually own.
05Where does this fit?
Ownership spans L5 product and workflow, where the capability meets real work, and L0 infrastructure, where logs, traces and deployments live. It is the organising idea behind the owned intelligence layer.
Previous: Out of context. Next: Data compliance with LLMs.