01

Practice 04 · AI and Agentic Architecture

Agents That Work Through Governed Interfaces.

Enterprise agent architecture and agent-to-system integration, with the controls around them: tool and API governance, human-in-the-loop design, deterministic validation, read-before-write, execution governance and audit.

  • Agent architecture · Tool governance · Interoperability
  • Read before write · Approve by risk · Verify after
  • Deterministic validation · Audit · AI risk controls
02

What we deliver

Architecture First, Then Agents.

An agent that changes an enterprise system is an integration with judgement. We design it like one.

01

Enterprise Agent Architecture

Where agents sit in the landscape, which systems they may reach, through which interfaces, and who owns each boundary.

Deliverablesagent architecture, capability map, boundary design.

02

Agent-to-System Integration and Tool Governance

Tool catalogues over governed APIs, using open protocols such as MCP where they fit, with scopes, limits and versioning.

Deliverablestool catalogue, scopes and limits, versioning policy.

03

Human-in-the-Loop and Execution Governance

Actions classed by risk. Low-risk actions run within policy; high-risk actions wait for explicit approval from a named role.

Deliverablesrisk classes, approval matrix, review interface patterns.

04

Deterministic Validation

Business calculations in versioned, tested code. Schema, policy and limit checks before anything executes.

Deliverablesvalidation layer design, policy versions, test cases with exact answers.

05

Auditability and AI Risk Controls

Sources, calculations, approvals and resulting system state stored per action, mapped to the controls your auditors already use.

Deliverablesaudit design, control mapping, monitoring.

06

Cross-Agent Interoperability

Agents from different vendors working through shared contracts, with identity and delegation designed rather than assumed.

Deliverablesinteroperability contracts, identity and delegation design.

03

When this matters

When AI Starts Touching Systems of Record.

  • Agents are being given access to ERP, CRM or finance systems

  • Several AI vendors are being piloted at once

  • Internal audit or security has asked how AI actions are controlled

  • A pilot worked and now needs to run in production

  • Numbers in AI output need to be trusted

  • Nobody owns the architecture of AI in the enterprise

04

Checklist

The Governed Agent Control Checklist.

Ten questions to ask any AI vendor or internal team before an agent is allowed to write to an enterprise system.

05

Questions

Do you build models?

We apply established methods, from retrieval-augmented generation to optimisation and probabilistic models, where they are the right tool. Business calculations stay in deterministic code.

Which AI platforms do you work with?

We are vendor-neutral. The control model sits between any agent and your systems, so it holds whichever models or agent frameworks you choose.

Where can we see the control model working?

In Value Lens, our decision-support product, now in private beta on synthetic ERP-like data. We walk through its artefacts and review log in a discovery call.

07

Next step

Review Your AI Architecture Before It Reaches Production.

Start with an assessment of agent access paths, controls and evidence, or read our governance principles first.

A conversation with an architect, not a salesperson. Replies within one business day. NDA before discovery.