IVAI OS
Governed computer-use automation for enterprise systems.
IVAI OS is designed to run repeatable business workflows inside the applications an organisation already uses — under its own policies and permissions, with consequential actions authorised before they happen and confirmed after they happen.
IVAI OS is currently in private pilot development. Capabilities described on this website represent the intended product and validation programme unless explicitly identified as completed.
The system observes before it acts.
Read the state of the application through the most reliable channel available. Raw interface information resolves into structured state before anything is proposed — and content encountered in an interface is treated as untrusted data, never as instruction.
Which channel is used is set by policy, not chosen by a model.
A model may propose an action. It does not authorise one.
Every proposal is checked against policy, the acting identity's permissions, the scope of the workflow and the limits set for it. Actions with financial, external, irreversible or uncertain impact are designed to require a human approval bound to an identity, a scope and a set of values — not a standing permission.
One action. One path. Once.
Re-check that conditions still hold, then carry out one approved action — designed so that a retry after an interruption establishes what already happened before acting again. The same operation must not silently become duplicate business actions.
The model does not confirm its own result.
A confirmation screen is not evidence. Verification reads the resulting business state through a different route from the one used to act — independently of the component that acted. Success is reported only when that independent check passes.
Claimed → observed → verified. A run that cannot be confirmed stops, and says why.
Reviewable afterwards.
What was decided, authorised, executed and confirmed forms a structured record, so a completed run can be reviewed rather than reconstructed from memory. Your own visit has been producing one — ephemeral, in this tab only:
Powerful because the boundaries stay visible.
Long workflows are designed to be split into short segments, each with its own checkpoint, because long-horizon autonomy is where current computer-use systems are least dependable. Many bounded workflows — each with its own scope, authority, execution boundary, evidence and independent verification.
The difference is where authority sits.
The model is the system
- The model decides what to do, and then does it.
- Authority is implicit: whatever the session can reach, the agent can change.
- Success is whatever the model reports, judged from the screen it acted on.
- Recovery means trying again — which is how one action becomes two records.
- The audit trail is a transcript.
The boundary is the system
- A model proposes. It does not execute.
- Every proposal is checked against policy, permission and scope before anything happens.
- Results are confirmed from the application's own records, independently of the component that acted.
- Consequential steps are designed to establish what already happened before retrying.
- The record is structured for review.
A conceptual contrast between unrestricted and governed architectures — not a claim about how any particular agent system behaves.
Where it stands. Where it goes next.
The first local workflow baseline implements controlled invoice creation: enter one invoice from supplied line items, check the total against an independent expectation, submit it once, and confirm the result from the application's own records.
The current synthetic test environment includes an invoice simulator, a resettable local record store, deliberate fault cases and an independent verifier. Browser automation and a controlled validation programme against a licensed non-production enterprise application remain the next steps before any customer deployment.
Not one narrow bot. A governance architecture for consequential computer-use workflows.
Document to system entry
A structured document arrives, its fields are checked against master data, and a record is created in the system of record — read back and compared before the run is reported complete.
Cross-system reconciliation
The same fact is held in two systems and has to agree. Matching rules are applied, a correction is proposed, and undecided cases are escalated.
Rule-bound case processing
A queue of similar cases resolved by written policy. Standard cases complete; exceptions reach a person with the evidence already assembled.
Controlled master-data maintenance
Changes to reference data applied under approval, with before-and-after values captured and the effect confirmed from the system's own records.
No single component owns the chain.
Perception, decision proposal, authorisation, execution, verification and evidence are separated deliberately — and that separation is the product.
One named workflow, one accountable owner, one definition of success agreed in advance.
Approvals bound to an identity, a scope and a set of values — not a standing permission.
What was decided, authorised, executed and confirmed, structured for review.
Where confidence is too low, the system stops and hands control to a person with the reason attached.
IVAI OS
Start a conversation