Agents you can trust in production.
Copilot Studio, Microsoft Foundry and Microsoft Agent Framework, built to hold up with real customers.
An agent in production at SSEN, AI-run Playwright testing at Principality, and GitHub Copilot and Claude Code across a 15-person Microsoft team.
- Azure OpenAI and RAG
- Model evaluation: OpenAI, Claude, Gemini
- MCP
- Threat modelling
- DPIA
- Responsible AI
- Production hardening
Retrieve, augment, generate, review.
An AI feedback analysis architecture from my Handbook’s case study: every channel in, Azure OpenAI grounded on Dataverse, and a person in the loop for revisions.
Answers that come from your data.
Grounded on Dataverse and your documents, with citations, and checks that every cited answer is also correct.
Narrow tools, clear limits.
A small tool surface through MCP, sign-in and entitlement checks before it acts, and threat models for the agent, the model and the network, as at SSEN.
The eval suite is the contract.
Write the evals before the build and run them on every change. Releases ship when the suite passes.
Every step traced, every token costed.
Every tool and model call traced, with cost per outcome. At SSEN, about 16p per item made the business case plain.
A person decides where it matters.
Durable approval with checkpoints and an audit trail, so a workflow can pause for a decision and resume safely.
Three Microsoft tiers. Use the lowest that fits.
- Copilot Studio. Grounded business agents, built fast and published with sign-in.
- Microsoft Foundry. Multimodal assessment with structured outputs, evaluation and tracing.
- Microsoft Agent Framework. Code-first, multi-agent workflows with shared state and durable approval.
Most use cases never need the top tier. Knowing which do is what my new book is about.
Build Production-Ready AI Agents89%
Efficiency gain reported by SSEN, with 90% accuracy against human review
A feedback agent that won Bronze for Best Use of AI.
- Azure OpenAI with RAG on Dynamics 365.
- Load-tested for 100,000 items a month, with human review of special cases.
- Governed by a DPIA, consent and threat models.
Reference builds you can see: Claims 365, the system my new book builds on all three tiers, and Project EVE.
More case studiesThree ways to take an agent to production.
Agent readiness review
Your current agents and data against what production needs, with a board-ready report.
Ask about thisLead AI architect
I hold the agent architecture inside your team while you build and ship.
Ask about thisAdvisory days
Design reviews and hard questions, for teams building agents. Booked by the day.
Ask about this