top of page

SilicoBalance is developing an enterprise software suite to govern what AI agents can do in connected business systems. Its core controls are designed to keep permissions tied to approved tasks, check actions before changes are made, enforce shared limits and prevent duplicate changes. The design also covers access withdrawal and recovery, with outcome verification to distinguish confirmed results from work that still needs investigation.

TASK AND ACCESS CONTROL

Define the actions, tools, systems and records each approved task allows an agent to use. Keep access tied to that task, with clear time limits and separate authorisation for changes to permissions.

PROTECTED ACTION CHECKS

Check the exact proposed change against the approved task, current permissions and relevant system state. Enforce those checks when the change is made. An AI reviewer’s approval cannot expand permissions or override the protected controls.

SHARED LIMITS

Apply shared budgets and resource limits across agents working at the same time. Count outstanding and unresolved commitments against the available allowance, so simultaneous requests cannot each spend the same remaining budget.

DUPLICATE PREVENTION

Link repeated submissions and retries to the original action. The aim is to prevent a second change when the same request is submitted again, including when a response is delayed or lost.

ACCESS WITHDRAWAL AND RECOVERY

Request withdrawal of access when intervention is needed. Account for completed, pending and unresolved work, confirm that the relevant access has closed, and require fresh authorisation before affected work resumes.

INDEPENDENT OUTCOME VERIFICATION

Read the affected system directly, separately from the agent’s own report. Compare the recorded result with the approved task and preserve the supporting evidence. Flag any outcome that cannot be confirmed for investigation.

Background

These controls describe the suite we are developing. Our local prototype uses simulated transactions, with further failure, recovery and acceptance testing ahead. Enforcement depends on routing all relevant actions through SilicoBalance’s protected services.

Intelligence that keeps intelligence in check

bottom of page