Agentic AI & Copilot Studio Architect @ Capgemini · Production agents for regulated Financial Services

Seventeen years in enterprise delivery, the last several spent getting AI agents past the pilot stage on the Microsoft stack. My working bias, in short: orchestration over prompting, evaluation in the pipeline rather than after it, and human-in-the-loop placed where the risk actually sits instead of wherever the UI had room for a button.

Most of what I build is client work and stays with the client. What I can publish goes to the Microsoft guidance repos and to the newsletter, both below.

Submitted to microsoft/PowerPnPGuidanceHub:

Audit, eDiscovery and Regulator-Ready Evidence for Copilot Studio Agents (Purview-first) A Purview-first evidence pattern for financial services, where auditability is a primary requirement rather than an afterthought. Separates agent execution from compliance operations: audit events from Copilot Studio, Power Automate and Dataverse flow into Microsoft Purview Audit, with optional SIEM/SOC integration, while Purview eDiscovery handles investigations, holds and regulatory production. Covers policy enforcement, approval workflows, least-privilege access, environment governance and evidence retention, aligned to the Power Platform Well-Architected Framework.

Agent Operations Center for Copilot Studio and Power Platform An operations-first pattern for production agents that take real business actions — the part that gets skipped between a working demo and something a support team can run. Separates agent runtime from operational support: Copilot Studio for orchestration, Power Automate for execution, Dataverse for state, a model-driven app as the operations surface, Application Insights and Log Analytics for observability, Power BI for reporting, Teams for human escalation. Covers monitoring, recovery, retry handling, escalation and SLA tracking.

Grounded: AI & Architecture — Real delivery. Real trade-offs.

A LinkedIn newsletter on enterprise AI architecture. No vendor keynote summaries; each issue takes one decision I have had to make on a real engagement and shows the trade-off, including where I got it wrong.

Seven issues so far, newest first:

• 07 · Copilot Studio Changed. Still Need Foundry? — whether the Copilot Studio/Foundry boundary still holds now the GitHub Copilot harness has gone GA. Three architecture choices, five questions for your next review.
• 06 · The MCP Spec Changes on 28 July — the largest revision since launch, and what it breaks for anyone already in production. A 15-minute architecture review to run before approving an MCP server.
• 05 · The Day the Agent Asked the Wrong Question — why reliable enterprise AI depends on context engineering rather than better prompts. The problem was not the answer; it was the context behind the question.
• 04 · Human-in-the-Loop Is Not a Button. It's a Design Decision — an approval step by default versus oversight by design: decision boundaries per action type, defined exception paths, escalation as a first-class pattern.
• 03 · Capable by Design. Provable by Design. — the question that changed the architecture. Copilot Studio got us started; Foundry made it provable — step-level trace, evaluation pipelines, per-domain risk tiering.
• 02 · Copilot Needed Orchestration, Not Just Prompts — when prompting alone failed and the architecture had to evolve: multi-step agent flows, tool-use optimisation, real validation and audit.
• 01 · From Canvas Apps to Code Apps — the eight seconds that changed my architecture decisions. Delegation limits, business logic trapped in the UI layer, and what replaced them.

Practical guidance for teams putting Copilot Studio and Power Platform agents into production — checklists, decision criteria and configuration patterns, with client detail stripped out.

In progress, in this order:

1. Pre-production review gate for Copilot agents — the checks to clear before an agent goes live: decision boundaries per action type, defined exception paths, recovery and idempotency under retry, evidence and retention. The criteria behind each check, and what actually satisfies it.
2. MCP server review checklist — what to establish before approving a third-party MCP server in an enterprise: capability surface, permission scope, data exposure, auth hardening, lifecycle and spec-version compatibility.
3. AI Act Article 50 transparency pack — mapping the EU AI Act's disclosure and AI-content-labelling duties onto concrete Copilot Studio and Power Platform configuration.
4. Lotus Notes DXL extractor — parses Domino XML exports of NSF databases into structured, inspectable output: forms, views, fields, computed formulas and document data. The extraction layer only, so a migration team can see what is actually in a database before deciding what to do with it.

Expert — Agentic AI Business Solutions Architect · Azure Solutions Architect · Power Platform Solutions Architect

Associate — Azure AI Engineer · AI Agent Builder · Power Platform Developer · Power Automate RPA Developer · Fabric Analytics Engineer · Power BI Data Analyst · Teams Application Developer

Specialty — Azure IoT Developer

Other — Frontier Transformation Engineer - TITAN (Microsoft) · AI Transformation Leader (Microsoft) · Claude Certified Developer, Certified Associate and Code Partner (Anthropic) · GitHub Copilot Core Skills · Four Copilot Dev Camp badges including MCP Integrator

Reach me on LinkedIn — happy to talk through anything above.