The Dynamics AI Agent Lie: It’s Not Acceleration, It’s Architectural Erosion This episode isn’t about whether Dynamics 365 Copilot works—it does. It’s about what it quietly dissolves. We explore how agentic assistance accelerates throughput while eroding the architectural joints that carry governance, accountability, and intent. Not through failures or breaches, but through drift: controls still exist, dashboards stay green, and meaning slips away. What We Cover

• Acceleration vs. Erosion: Why speed isn’t neutral—and how increased throughput stresses the places where policy meets behavior.
• Agents as Control-Plane Participants: Copilot isn’t an in-app helper; it’s a distributed decision engine spanning Dynamics, Graph, Power Automate, Outlook, and Teams.
• Mediation Replaces Validation: How summaries, confidence bands, and narratives reframe what humans actually review.
• Composite Identity & RBAC by Composition: Why least privilege passes reviews while effective authority expands across orchestrated pathways.
• Non-Determinism on Deterministic Rails: How probabilistic planning breaks regression testing and replay.
• Blast Radius Growth: Helpful actions propagate across surfaces, widening incident scope.
• Audit Without Causality: You can see what happened, not why—because the decision trace lives outside your logs.

The Four Scenarios That Quietly Reshape Control

1. Invoice Approval — Validation becomes mediation; approval quality tracks narrative quality, not signal quality.
2. Credit Hold Release — Deliberate exceptions become suggestible defaults; seasonality and partial histories collapse into a click.
3. Procurement Vendor Selection — “Neutral” recommendations privilege data density and integration, calcifying supplier mix.
4. Customer Service Resolution — Ambiguous authority by design; benevolence defaults leak value under queue pressure.

The Mechanics Behind the Drift

• MCP & Orchestration: View models expose affordances; planners compose legitimate actions into emergent pathways.
• Human-like Tooling (Server-Side): Robust navigation without a client increases confidence—and hides discarded branches.
• Deterministic Cores, Probabilistic Paths: The function is stable; the path to it isn’t.

Controls That Fray—and What to Do Instead

• Why DLP, Conditional Access, Least Privilege, ALM, and SoD struggle against composition and synthesis.
• What survives: intent enforced as design—decision traces, step-up on sensitive tool invocation, ALM parity for prompts/tool maps/models, and SoD across observe-recommend-execute.

The One Test to Run Next Week Pick one real, dollar-impacting Copilot-influenced decision and score it on five flags: Composite Identity Unknown, Lineage Absent, Non-Determinism, Unbounded Blast Radius, Accountability Diffused. Two or more flags isn’t a bad record—it’s your baseline. Executive Takeaway Speed improves medians while widening tails. The debt shows up as variance you don’t price, blast radius you don’t bound, and explainability gaps you don’t track. Pay a little friction now—gates, traces, step-ups—or pay later in archaeology. Remember This

• If intent isn’t enforceable in code, it won’t hold in production.
• If you can’t reproduce a decision, you can’t defend it.
• If your logs don’t capture causality, you don’t have accountability.
• Exceptions are entropy; budget them.
• Paper controls can’t govern compiled behavior.

Resources & Checklist: Link in the notes.
Subscribe for more calm, clinical breakdowns of enterprise AI—without hype.

Become a supporter of this podcast: Back to Episodes