top of page

"AI should expand human potential without diminishing human control"   

Intelligence that keeps intelligence in check 

SilicoBalance is developing an enterprise software suite to help technology, security and operations teams govern what AI agents can do in connected business systems. It is designed to keep permissions tied to approved tasks, enforce limits before changes are made, and give operators evidence of the outcome. Our first local prototype is testing this approach through simulated transactions.

A

DEFINE WHAT EACH TASK ALLOWS

Every approved task has a defined perimeter. It specifies what the agent can do, which tools and records it can use, how much it can spend or consume, and when its access ends. The software checks the proposed action against that perimeter and the permissions in force at the time.

B

KEEP PERMISSIONS TIED TO THE TASK

As an agent takes on more work, each task keeps its own permissions. Changes to permissions or safeguards go through a separate authorisation process. The design allows operators to request withdrawal of access and distinguish confirmed outcomes from work that still needs checking. Withdrawal and recovery behaviour remain part of the prototype testing programme.

C

TRACE WHAT HAPPENED

SilicoBalance is designed to preserve a record of who authorised a task, what the agent proposed, which rules were applied and what the affected system recorded. That record shows which actions completed, which were blocked and which still need investigation.

D

CONNECT GOVERNANCE TO YOUR AI WORKFLOWS

SilicoBalance is designed as a governance layer for AI workflows. The suite brings together policy, action checks, operator oversight and audit evidence. The planned developer tools, including a software library, would help teams connect their agents and business systems. A proposed shared protocol would define how connected components exchange tasks, proposed actions, permission decisions and verified results. Enforcement depends on routing the relevant actions through SilicoBalance’s protected services.

E

KEEP PROPOSAL, PERMISSION AND VERIFICATION SEPARATE

The protected service checks that the proposed change matches the approved task and remains within the permissions and limits in force when the change is made. Reviewer feedback cannot expand those permissions or override the controls.

Controls the suite is designed to support

FEATURE 1

TASK PERMISSIONS

FEATURE 2

ACTION VALIDATION

FEATURE 3

CONTROLLED TOOL ACCESS

FEATURE 4

SHARED LIMITS

FEATURE 5

DUPLICATE PREVENTION

FEATURE 6

ACCESS WITHDRAWAL

HOW IT WORKS
 

  1. Approve a task and set its limits: Define the authorised task, the tools and records the agent may use, and any spending, resource or time limits.

  2. Submit and review the proposed action: The agent submits the exact proposed change and supporting evidence. A separate reviewer can assess it, but its approval cannot override the task’s permissions or limits.

  3. Enforce permissions at the point of action: The protected service checks that the change matches the approved task and remains within current permissions and limits before allowing it to proceed.

  4. Verify and record the outcome: A separate check reads the affected system directly to establish what changed. The evidence is linked to the approved task and the checks applied. Any outcome that cannot be confirmed is flagged 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

From Prototype to Independent Evaluation

We’re starting with a prototype that uses a simulated transaction. It gives us a way to test whether approved work can proceed, whether attempts outside the task are stopped, and whether the final result can be independently verified. The next phase would extend testing to failure and recovery scenarios and prepare the evidence for independent evaluation.

01

WORKING PROTOTYPE

02

FAILURE TESTING

03

PARTNER EVALUATION

04

LARGER-SCALE TESTING

bottom of page