OCAP's got four principles and the fourth one settles the argument. Ownership, Control, and Access are claims a tribe can assert on paper, in a contract, in a policy a council votes on. Possession is a physical fact about where the data sits and who can walk over and touch it. I've never seen a cloud agreement that changed a physical fact.
That's the whole problem with running tribal data through cloud AI, and it doesn't get solved by reading the terms of service more carefully.
OCAP and CARE Aren't the Same Kind of Document
The First Nations Information Governance Centre published the OCAP Principles, Ownership, Control, Access, and Possession, establishing that a tribe retains ownership and control of tribal data. The CARE Principles, Collective Benefit, Authority to Control, Responsibility, and Ethics, sit alongside them and govern how the data gets used once you've got it.
People treat these as one framework. They're not. CARE's an ethics document and it survives contact with a vendor just fine, because a vendor can be held to it contractually. OCAP's Possession principle is an infrastructure requirement wearing a governance document's clothes, and no contract satisfies it, because contracts govern what a party's permitted to do rather than what a party's physically able to do. You can't sign your way into holding something. Thirty-odd months of vendor compliance pages haven't changed that and won't.
The IHS Tribal Data Governance Guidelines from 2021 recommend local infrastructure for systems holding Protected Health Information, which reads like a technical suggestion until you notice it's the only recommendation that makes Possession operable. Tribal emergency management frameworks from the Bureau of Indian Affairs and tribal emergency management councils go further: incident data stays under tribal control during and after operations. That's a sentence written by people who've watched a federal response arrive and then watched the records leave with it.
Where the Work Really Is
Start with an EOC during an active event. It generates incident reports faster than anybody can read them, and the person drowning in that is usually one emergency manager with a radio in the other hand. That's a real AI problem. It's also a live record of what happened on tribal land during a declared emergency, which is exactly the category the frameworks say can't leave.
But a tribal government is a whole government, and the EOC is one office in it. Enrollment sits on lineage records that are nobody else's business, ever. The THPO handles Section 106 consultations and site data that has gotten people's ancestors dug up when it traveled. Social services runs ICWA cases where the file is a child. Natural resources and fisheries carry decades of ecological observation that outside institutions have been trying to get their hands on since long before anybody said "training data." Tribal court has its own code and its own docket. Housing runs NAHASDA reporting, TERO tracks who gets hired on tribal land, gaming compliance lives under Title 31 and the NIGC MICS, and fiscal reconciles 638 contracts against federal reporting calendars that don't agree with each other.
Health services are the same shape with worse consequences. A clinic pushing clinical documentation and record analysis through a cloud model is feeding PHI into a system whose HIPAA posture is Covered-Entity-adjacent at best, which is a polite way of saying nobody's established one.
Language and cultural preservation's the one where I'd ask people to slow down. There's real capability here. There's also a long, ugly history of Indigenous knowledge getting collected, centralized, and monetized by parties who contributed nothing to it. That work belongs on tribally owned infrastructure or it doesn't belong anywhere.
What I Won't Tell You Before I've Watched You Work
Here's where most vendors hand you a configuration. I'm not going to, and it isn't modesty.
I don't know what your enrollment officer does all day. I don't know which ICWA files take four hours because the state sends them in a format nobody can use, or which quarterly report your housing authority rebuilds by hand every single time, or what your court clerk keeps in a spreadsheet because the case management system was never configured for a tribal code. I don't know which of your departments has quietly built a workaround nobody wrote down, though I'd bet money at least three have.
Neither do you, entirely, and that's just how governments work, with no insult intended. That knowledge lives in the people who've done the job for fifteen or twenty years, and it barely lives in the documentation.
So the sequence runs backwards from how this usually gets sold. I show up, sit down next to those people, and shut up until I understand the work the way they run it instead of the way the SOP describes it. Discovery turns up the use cases already sitting in the building, most of which nobody's said out loud yet. Only then does anybody talk about what to buy, and what to buy is whatever the work turned out to need, with room to grow.
There's no catalogue here. There's no configuration I recommend to tribes generally, because "tribes generally" isn't a thing that has a workflow.
Then we install it together and I onboard your team, and that last part decides whether any of this was worth doing. Not a lunch-and-learn on prompting. Workflows and agentic orchestration built around what your government does all day, so the thing's still running a year after I've gone.
The Honest Comparison
Cloud AI's cheaper at the front. API fees for a clinic-sized workload run $500 to $1,000 a month, they never stop, and they climb. Ownership's a bigger commitment at the start and a smaller one every year after, and what it costs depends on what got built, which depends on the work, which is why it comes out of a scoped conversation instead of a page like this one.
The money's the least interesting part of this comparison anyway. A tribe that saves a little and gives up possession has made a bad trade at a discount.
What Getting Started Looks Like
Bring it to council as a sovereignty question, because that's what it is, and councils handle those better than they handle procurement questions. Find the department that owns the pain. Usually health, emergency management, or fiscal. Then get me in the building.
The infrastructure questions are real: power, a room that locks, where it sits on your network, whether it touches the internet at all. They get answered during that visit instead of guessed at from a blog post. If a satellite clinic needs to reach it, that's a VPN conversation and Tailscale or WireGuard handle it without exposing anything. If your compliance posture says the box never touches the internet, then it never touches the internet, period.
Your team runs it afterward through an interface built for people who aren't engineers, with the governance layer already wired in, so every action's identified, approved, and logged. That audit trail isn't a feature I'm upselling. For a sovereign government it's the difference between having your data and being able to prove what happened to it.
Possession, In Practice
A tribe's got the inherent right to govern its own information. Everybody agrees on that sentence, which is how you can tell it isn't the contested part.
The contested part is whether that right survives being handed to Microsoft, Google, or Amazon along with the data. It doesn't, because possession is a fact about a building, not a permission you grant.
Put the infrastructure on tribal land, under tribal control, run by tribal staff, and OCAP stops being a document a council adopted and starts being a description of where the servers are.
That's the whole argument. If you want to talk about what it'd look like in your government, start there, and I'll come to you.