EPC Group issues governance standard for Microsoft-Decision-1

EPC Group published a four-control standard for enterprises planning around Microsoft-Decision-1, the new Microsoft decision-scoring model released Oct. 9. The guidance focuses on identity, data protection, approval rules and evidence before tenants let the model trigger actions in Microsoft 365 and related workflows.

Why it matters: - Microsoft-Decision-1 is cheap enough to sit in front of automated actions across Microsoft 365, Power Platform, Dynamics 365 and Azure workflows. - EPC Group says that shifts the risk from model cost to governance, especially when a score can trigger a business action. - Microsoft’s own guidance says consequential decisions about people need meaningful human review and should not rely on the model as the sole decision-maker.

What happened: - EPC Group published a four-control governance standard for enterprises evaluating Microsoft-Decision-1. - The standard was written by EPC Group founder and Chief AI Architect Errin O'Connor. - Microsoft released Microsoft-Decision-1 on Oct. 9, 2026 through Microsoft Foundry and OpenRouter. - EPC Group made the standard available free on its website.

The details: - Microsoft-Decision-1 returns calibrated probabilities for fixed options; it does not write text. - Microsoft lists support for up to 32,000 input tokens, a launch price of $0.042 per million input tokens, and free output tokens. - Microsoft says the model is based on a post-trained Qwen3.5-9B base, with other bases planned. - Microsoft reported top accuracy in its own 36-benchmark comparison of nearly 150,000 hidden questions; those results reflect Microsoft’s evaluation conditions. - EPC Group’s four controls cover identity, Microsoft Purview and data minimization, an explicit action policy, and an evidence record. - The identity control requires three least-privileged identities before a pilot starts: the workload that calls the scoring endpoint, the identity that retrieves source information and the agent or app that executes the action. - The Purview control requires minimized context and verification of sensitivity labels, data-loss-prevention policies and audit logging on the exact path the custom integration uses. - The action-policy control uses three workflow states: Proceed, Fail and Contested. - Proceed means mandatory checks passed and the action is in approved scope. - Fail means a mandatory check failed or the action is out of scope. - Contested means information is incomplete, judgment is uncertain or review is required. - EPC Group says a high score never maps directly to Proceed, a model outage never becomes permission, and a retry that returns a different score is treated as Contested. - The evidence record keeps request and transaction IDs, the approved question, options and policy version, the scores and threshold, model and deployment IDs, a secure reference to supporting context, and the reviewer’s decision and actual action. - The standard calls for reconstructing a completed pilot action from the record alone. - EPC Group’s TAR-8 Tenant AI Readiness framework applies the standard across eight tenant areas scored from zero for unknown to three for evidenced. - The eight areas are identity and authority to act; Purview and sensitive data; SharePoint and OneDrive permissions; Teams conversations and transcripts; Power BI and Microsoft Fabric; CRM and business applications; Azure infrastructure and external routes; and ownership and evidence. - SharePoint and OneDrive access must be resolved before retrieval and transmission. - Microsoft’s Restricted Content Discovery is treated as a discovery control, not an access boundary. - Teams transcripts and chat history must also be minimized. - In Power BI and Microsoft Fabric, the model can route questions or flag answers for review, but it does not settle conflicting business definitions. - In Dynamics 365, Dataverse and Power Platform, read scope and write authority are reviewed separately, and material fields are revalidated at commitment. - The standard distinguishes Microsoft’s Global Standard deployment from Data Zone Standard and says neither by itself creates a single-country guarantee. - OpenRouter is treated as a separate route that needs its own review of routing, credentials, terms and data handling. - The pilot process runs four weeks: inventory and scope in week one, labeled cases and observation mode in week two, control testing in week three, and a go, revise or stop decision in week four. - EPC Group says the economics should be measured in reviewer effort, error cost by category and the cost of controls.

Between the lines: - The standard is as much about workflow control as model performance. - EPC Group is signaling that enterprises should not assume Microsoft’s native protections carry over automatically into custom integrations. - The company is also drawing a hard line between scoring accuracy and permission to act. - That matters because many AI deployments fail at the handoff from prediction to execution, not at prediction itself. - EPC Group’s approach is designed for regulated tenants that need auditable decisions, not just better model output.

What's next: - EPC Group says tenant owners should verify changing Microsoft boundaries before they design around the model. - Those boundaries include Entra Agent ID changes for Copilot Studio agents, Restricted SharePoint Search limits, Copilot in SharePoint rollout timing, and Copilot Business billing changes. - For government-cloud customers, Microsoft-Decision-1 was not listed on Microsoft’s Azure Government model catalog as of Oct. 13, 2026. - GCC, GCC High and DoD tenants are expected to confirm availability, data boundary and terms before building around the model. - EPC Group says organizations can use the standard for a tenant review or a governed automation pilot.