For the CIO and Chief Architect
AI becomes part of the architecture: owned boundaries, known access paths and one control model across vendors.
01
AI and governance
Enterprise AI Should Not Bypass Architecture.
Agents that read and change enterprise systems are integrations with judgement. They need the same contracts, controls and evidence as any other integration, and a few more. These are the principles we design to.
02
Principles
Seven Principles for Governed Enterprise AI.
Operate through governed interfaces.
AI reaches enterprise systems through defined tools and APIs with scopes and limits. Not through shared credentials, screen automation or direct database access.
Keep business calculations deterministic.
Prices, margins, quantities, tax and balances are computed in versioned, tested code. The model can explain a result; it does not produce it.
Read before write.
An agent reads the current state of the system before it proposes a change, and the proposal is expressed as a difference from that state.
Validate before execution.
Every proposed action passes schema, policy and limit checks before anything runs. A failed check stops the action and says why.
Approve by risk.
Low-risk actions run within policy. High-risk actions wait for explicit approval from a named role, and the approval is recorded with the action.
Verify independently.
After execution, a separate check reads the system back and compares the result with the intent. Sending a command is not the same as success.
Keep evidence and state auditable.
Sources, calculations, approvals and the resulting system state are stored per action, in a form internal audit and security can read.
03
Architecture
The Layers Every Agent Action Passes Through.
The controls sit between agents and systems of record, so they hold whichever models, agent frameworks or vendors you choose.
- Agents and assistantsAny model, any vendor
- Tool and API governanceScopes · limits · versions
- Validation and policySchema · policy · approval by risk
- Execution and verificationRead before write · verify after
- Systems of recordERP · CRM · data · legacy
- Evidence storeSources · decisions · state
04
In practice
What Changes When These Principles Are Applied.
For Security and Internal Audit
Every agent action can be traced to its sources, its checks, its approver and the state it left behind.
For Finance and Operations
Numbers come from tested code under versioned policies. AI explains them, and a person decides what to do.
05
Questions
Does this slow AI adoption down?
It makes production possible. Pilots without these controls tend to stop at the point where an agent needs to write to a system of record.
Is this tied to a particular AI vendor?
No. The principles apply to any model or agent framework, and the control layer sits outside them.
Can you review what we already have?
Yes. We assess existing agents and pilots against these principles and deliver a gap list with the changes needed to run in production.
06
Next step
Review Your AI Architecture Before It Reaches Production.
We assess agent access paths, controls and evidence against these principles and give you a prioritised gap list.
A conversation with an architect, not a salesperson. Replies within one business day. NDA before discovery.