Appendix ASovereignty Maturity Model
A five-level model for placing an institution's AI estate on the path from rented to owned. Each level is defined by what the institution can verify, not by what it has bought. Use it with Appendix B, which scores an institution against the questions here.
The five levels
| Level | Name | Defining property | Typical state |
|---|---|---|---|
| 1 | Rented | The vendor runs everything; the institution has accounts | Per-seat SaaS assistants; data in the vendor's cloud; one model; no source; no connectors the institution owns |
| 2 | Configured | The institution controls settings but not the system | Hosted platform with the institution's SSO and some connectors; model fixed; data copied to the vendor; export on request |
| 3 | Deployed | The institution runs the platform; the vendor supports it | Platform in the institution's cloud account from source under license; models routable; connectors through an institution-run broker; data read in place |
| 4 | Owned | The institution could run it without the vendor | Source built and deployed by the institution's own engineers; open-weight models on institution hardware for restricted data; connectors and skills in institution repositories; quarterly exit verification |
| 5 | Sovereign | Every property holds with the network cable unplugged | Air-gapped build proven; signed update bundles; all five properties of the framework verified, including local models, local embeddings and local identity |
What distinguishes each level
Level 1 to 2. The institution stops using personal accounts and consumer tools and adopts a platform with institutional identity and administration. Nothing structural has changed: the vendor still runs the system and holds the data, and the price is still per seat. Most institutions in 2026 are here, and many believe they are further along because the contract has a data-residency clause.
Level 2 to 3. The structural step. The platform is deployed into infrastructure the institution controls, from source the institution holds under license. The model becomes a setting. Connectors read systems of record in place through a broker the institution runs. This is the level at which the two problems from Chapter 1 are solved, and the level most of this book describes.
Level 3 to 4. The institution's own engineers can build and deploy the platform without the vendor's help, and have done so at least once. Restricted workloads run on open-weight models on the institution's own GPUs. Every connector, skill and agent configuration lives in repositories the institution owns. The four exit questions from Chapter 3 are verified on a schedule. The vendor relationship is now a choice renewed on merit.
Level 4 to 5. The institution has proven the air-gapped build: the platform starts and serves with no outbound network, with local models, local embeddings and local identity, and updates arrive as signed bundles. Few institutions need to operate at Level 5. Every institution at Level 4 benefits from being able to demonstrate it, because the demonstration proves there are no hidden dependencies.
Common misplacements
Three patterns recur when institutions self-assess.
A sovereign region is not Level 3. Running a vendor's hosted product in a hyperscaler's in-country region keeps the disks local. It does not give the institution the source, the model choice, or custody of the data path. That is Level 2 with a map.
Open source is not Level 4. A platform whose client is open source but whose backend is the vendor's hosted service cannot be run by the institution. Level 4 requires that the whole system, backend and infrastructure definitions included, builds and runs on the institution's infrastructure.
A pilot is not a level. An institution running a Level 3 pilot for one department while the rest of the estate is Level 1 is at Level 1. The model is applied to the estate the institution's people actually use.
Moving up one level in a quarter
Each transition has one decisive act.
- To Level 2: consolidate onto one platform with institutional sign-in and administration; retire personal accounts.
- To Level 3: deploy the platform from source into the institution's cloud account; connect the first system of record through the broker; make the model a setting and switch it once to prove it.
- To Level 4: have the institution's engineers perform a deployment end to end without vendor hands on keyboards; stand up the first local open-weight model for a restricted workload; move every connector and skill into institution repositories.
- To Level 5: build and start the platform on a disconnected network; load one model and one connector; verify the signed-bundle update path.
Appendix F sequences these into a ninety-day plan for an institution starting at Level 1 or 2 and aiming at Level 3.
Run the stack this book describes
ibl.ai is the agent operating system you run yourself: you own all the code and the data, in your cloud, on-premise or air-gapped, with any model, one identity across web, browser, desktop, mobile and terminal, and no per-seat pricing.