Governed AI, in Plain English
A short glossary of the terms that come up when AI has to be safe for sensitive data and real decisions. Each entry leads with why it matters, then how it works. Each one links to the part of the platform where it lives.
Attribute-Based Access Control (ABAC)
Attribute-based access control answers a sharper question than the usual "who logged in?" It asks "who is allowed to see this, run this, and do this, right now?" and answers it on every request. That precision is the benefit: you can express a rule a fixed list of roles simply can't, such as "allow this user to query financial documents only if they're in Finance, the document is in their business unit, and it isn't a draft." Mechanically, each decision is computed from attributes (who the user is, what they're touching, what action they're attempting, and the context) rather than a static role label. One authority makes the call, so there's no second permission system to drift out of sync.
See it in Governance & Control →
Tamper-Evident / Attested Audit
Almost every system claims an audit trail; few have one that would survive a dispute. A tamper-evident (or attested) audit trail is the difference between a record you wrote down and a record you can prove hasn't been altered. The benefit: a regulator, auditor, or customer can verify it without trusting your database administrator. Mechanically it is append-only by design, structured so any change would be detectable, and complete at the point of action: who did what, to what, when, and on whose authority. For AI that influences hiring, credit, procurement, or legal outcomes, that shift from logged to provable is fast becoming the expectation.
See it in Governance & Control →
Governed Knowledge Plane
Finding a passage is now a commodity; the hard part is turning scattered, sensitive, ever-changing information into knowledge a person or an agent can actually rely on. A governed knowledge plane is that trustworthy layer. It gives you answers you can prove and act on, instead of text you have to double-check. Mechanically it means retrieval that respects your access rules at every hop, answers that cite the sources they're drawn from, knowledge that stays current as permissions and content change, and the ability to follow relationships between documents without leaking across boundaries. It's the difference between "search that found something" and "an answer you can build on."
See it in Deep Data Intelligence →
Agent Identity & Controlled Delegation
When AI starts to act, the key safety question is no longer "can it do the work?" but "on whose authority, and how far can it reach?" Giving every agent its own identity means you can always answer "which agent did this, on whose behalf, and with what authority?" The benefit is containment and accountability. Controlled delegation is the mechanism: an agent holds its own attributes and lifecycle, and authority narrows as work is passed from a person to an agent and on to a sub-agent, never widens. An agent several steps deep in a task can do almost nothing it wasn't explicitly granted, so a tricked or misbehaving agent's blast radius stays small.
See it in the Agent Workforce →
Local / Self-Hosted LLM (Large Language Model)
A large language model (LLM) is the AI that reads and writes text. Most platforms send your prompts (which contain your data) to an outside model provider; a local or self-hosted LLM runs the model inside your own environment instead. The benefit is control: no prompt or data leaves your perimeter, and no external AI provider is ever called — which for the most sensitive work, or an air-gapped network, is the only arrangement that works at all. Mechanically you choose the model and where it runs rather than being locked to one vendor's, and both your data and your inference stay inside your infrastructure boundary. It's how you get modern AI without exporting your crown jewels to run it.
See it in Deployment & Data Control →
Data Sovereignty
Data sovereignty is the principle that your data — and increasingly the AI processing it — stays under your jurisdiction and control, subject to your rules rather than someone else's default. The benefit is straightforward: regulated, sensitive, or commercially critical information doesn't end up in shared infrastructure you can't see into or govern. Mechanically it's delivered by running the platform where your obligations require: private cloud, on-premise, even air-gapped. Full tenant isolation means no other customer's workload shares your data environment, and data-residency and access controls are enforced at the architecture level. Sovereignty turns "trust us with your data" into "your data never left your control in the first place."
See it in Enterprise Security →
AI + Cloud Cost Governance
"What is our AI actually costing us?" is a question most stacks answer with a monthly surprise. AI + cloud cost governance replaces that with a precise, trustworthy number: spend you can see, attribute, forecast, and cut. Mechanically it keeps a per-call ledger of AI usage, recorded to a fraction of a cent and attributed by team, user, model, and project, right down to a single conversation. It then joins that with the cloud spend underneath it for one honest picture. Because the collection is read-only and can run inside your own account, it's also the lowest-risk place to start: prove value on a problem you already have budget for, then expand.