Marilou Puissant

Building a Copilot Studio agent is easy. Getting that agent live, trusted by users, and actually embedded in daily operations? That's a different story. After deploying Copilot Studio solutions across multiple projects, we learned that the biggest challenge is rarely getting an agent to work. The hard part is making it predictable, explainable, maintainable, and deployable in an enterprise environment without everything turning into chaos after the pilot phase. This session is an honest look at what actually separates successful Copilot Studio implementations from the ones users quietly abandon. We’ll go beyond AI hype and focus on the lessons learned the hard way: why the most effective solutions are often not the most agentic ones, why unrestricted environments become impossible to govern, and how to embed this in existing processes to ensure adoption. The session covers: • Lessons learned: What broke, what surprised us, and the decisions we'd make differently. • Deterministic vs agentic: How to decide which parts of a workflow should stay deterministic, where AI truly adds value, and how both reinforce each other. • CI/CD and operational maturity: How to use solutions, environments, and deployment pipelines to scale beyond a single environment without losing oversight and maintainability. • Tools, subagents, and best practices: When subagents improve clarity versus when they only add complexity, how to extend agents through connectors and MCPs, and why topics are more reliable guardrails than prompt instructions alone. The session includes a live demo showing how flows, tools, and orchestration can work together in a production-ready Copilot Studio setup. By the end of this session, you will know how to structure agents that remain in control, how to deploy Copilot Studio solutions across environments without governance chaos, and how to balance agentic and deterministic design to build solutions that are both powerful and trustworthy.