A standard by True North Agentics
Most AI advice skips the only question that matters when you hold other people's data: who can see it, and what is the AI allowed to do on its own? The Controlled-AI Framework answers that in four controls — the standard we build every workflow against, and the checklist you can hold any AI tool or partner to.
The four controls
Any one of these on its own isn't enough. Controlled AI is the combination — data, access, action, and proof, all accounted for.
For hosted AI, verify the exact commercial product, settings, feedback choices, and contract. For self-hosted AI, data can remain within the configured environment only when the deployment, connectors, logging, telemetry, and backups are all scoped that way.
Sensitive or high-impact actions always require named human approval. Bounded low-risk actions may run automatically only when explicitly approved, tested, logged, and reversible.
Roles and permissions keep each client's data separated and reachable only by the right people. The AI connects to exactly what a task needs and nothing more — least access by default.
Define which inputs, outputs, decisions, approvals, and system changes must be logged, then test that the configured record is complete enough for the workflow's review and retention requirements.
Take these four controls into any AI conversation — vendor, consultant, or in-house. If a tool can't answer all four plainly and in writing, it isn't ready for your clients' data.
Built for regulated, trust-heavy work — accounting and bookkeeping firms, finance and insurance teams — where a data leak or an unapproved action isn't an inconvenience, it's the whole reputation.
Common questions
The Controlled-AI Framework is True North Agentics' four-control standard: verify provider and deployment data boundaries; require named human approval for sensitive or high-impact actions; grant least access; and define, test, and review the required audit trail. Bounded low-risk actions may run automatically only when explicitly approved, tested, logged, and reversible.
Security isn't one setting — it's where the data lives, who can reach it, what's allowed to happen, and whether you can prove it afterward. Miss any one control and it isn't controlled AI. Use all four to screen any AI tool, consultant, or in-house build before it touches sensitive data.
The reason to name this a framework instead of a feature is simple: firms get burned when they check one box and assume the rest. A product with a default training exclusion can still be configured incorrectly or take an action no one approved. A system with human approval can still over-share access across clients. Real control means all four at once — where the data lives, who can reach it, what's allowed to happen, and whether you can prove it after the fact.
True North Agentics builds every workflow against these four controls, and we'd rather you ask hard questions than take our word for it. Bring the framework to any vendor or consultant you're considering — including us.
Curious whether a specific workflow in your firm can be automated inside these four controls? A free async fit check gives you a straight answer — including an honest "not worth it" if that's the case.
Start the async fit checkControlled AI