And the record of it, held by the business, not the vendor.
Regulated businesses have paid for AI and cannot get it past their own security review, because nobody can prove what the AI is permitted to do. Lokorium is how they move.
LOKORIUM · the independent referee between an AI decision and its consequence
The regulated market is splitting two ways against the same wall. One is quietly exposed. One is quietly frozen. Neither of them can prove what the AI is allowed to do.
A regulated consumer lender has put AI right across the business. It is getting real value. In its own words, its boundaries are stated positions rather than things it can prove are enforced, and nothing inspects the data leaving. It is not looking to start. It is looking to close a gap it already knows is there.
State · Exposed
A second regulated business has the money and the mandate. The programme dies in the security review because nobody can answer the two questions that matter: what is the AI allowed to do, and how would you know if it did something else. Value sits unrealised while a competitor moves.
State · Frozen
One is exposed. One is frozen. Both are waiting on the same missing thing.
Every time the AI tries to do something that matters, send data somewhere, run code against a system, move money, make a promise, Lokorium decides yes or no against the rules the business set, at the moment it tries to act. Not a filter after the fact.
The artifact is a permit. One signed, bounded, expiring permit describes one consequence, with eight parts. Change any of them and the permit no longer matches.
01 · Who
agent id · signed session
02 · Why
stated business intent
03 · What
exact action, one consequence
04 · Where
named target system
05 · Limits
amount · scope · rate
06 · When
valid from · valid until
07 · Freshness
issued at · nonce
08 · Proof
signature · rule hash
Without Lokorium, a gateway can say no and the target can still say yes through another route or credential. With Lokorium in place, the target only accepts requests that carry a valid permit. A different route does not create authority. The bypass reaches the target and dies there.
Honest note on altitude: enforcement is strongest at target-native deployment. Where the target is instrumented to check the permit, the refusal holds. Where it is not, we’ll tell you plainly what the enforcement rests on, per workflow.
Illustrative · scroll to follow the full path
Illustrative · scroll to follow the full path
When a regulator or a board asks how you know, the answer is a record the business holds, spanning every vendor. Not the cloud’s logs. Not the model vendor’s logs. Both are interested parties.
Two artifacts. A chain of five linked records per request. And a ladder that grades the evidence honestly, because a log is not proof.
01
Proposal
the intended action
02
Policy decision
rule matched
03
Signed authority
the permit
04
Execution gate
target check result
05
Outcome receipt
what actually happened
Illustrative · records signed and chained · scroll to follow the chain
Evidence ladder
The chain is graded. A decision log is the floor. A corroborated outcome, matched against the customer’s own system of record, is the ceiling. We do not pretend a log is proof. We aim, per workflow, for the highest rung the target and the estate allow.
Illustrative · graded evidence, per workflow
The panel below is a concept, not a live product demo. Two example sessions: a legitimate request, then a bypass that tries another route and dies at the target because it has no valid permit.
A legitimate request
A permit is issued, bound to this action. The target accepts it and executes.
Position, not guarantee
A business uses one provider to reason, another for identity, another for the estate. None of them will ever sit in judgement over the others. A control sold by one vendor exists to sell you more of that vendor.
Lokorium has no incentive to favour any one vendor. That is the structural position we design toward. Structural, and intended. Not a stamped guarantee.
Every system in the business is reached by paths that never pass through one vendor’s edge. So the thing that must refuse an action is the system it would affect. Lokorium is designed to sit at that point, across every vendor the business uses.
When a regulator asks how you know, the answer is a chain the business holds, spanning every vendor. Not the cloud’s logs. Not the model vendor’s logs. Both are interested parties. The chain is graded, from a decision log up to an outcome corroborated against the customer’s own system of record.
Not every business needs Lokorium. If your programme has no rules an auditor would ever ask about, it does not. If it does, this is for you. For each sector, the one action that would be career ending if the AI got it wrong.
Career ending
Moving customer funds or data outside a named, current authority.
Career ending
Filing or disclosing without a partner sign off, or breaching legal privilege.
Career ending
Reading, writing to or acting on the wrong patient record.
Career ending
Paying, or refusing, a claim outside underwriting authority.
Career ending
Issuing a case decision without the delegated authority to make it.
The permit layer follows an arc. Today it lives in enterprise data and action paths. As agents take on more, it moves with them. As robots arrive, it moves again. One layer, three eras.
Lokorium governs what AI is allowed to do with sensitive data and actions inside regulated businesses.
The same permit layer extends to agents acting across systems on the business’s behalf.
Embodied AI. Robots acting in the real world. The question is identical: is this allowed to happen, before it happens.
Bring an adult a glass of water.
Permit matches. Target accepts.
Bring a child a glass of alcohol.
No matching permit. Target refuses.
Software today, or a robot arm tomorrow. The question is the same in every era: is this allowed to happen, before it happens.
A horizon, not today’s product. Today Lokorium works on the first of the three.
Tell us where you are stuck, or where you have gone live and cannot yet prove what the AI is allowed to do. We will come back with what a briefing would cover for your workflow, and what we would need from you to make it useful.
Direct email: enquiries@lokorium.io