Governed AI in the Flow of Work

From question to governed answer
in under 60 seconds

See how a governed AI agent reads enterprise data, answers in Microsoft Teams, and produces a complete audit trail — all with a real M365 identity. Not a concept. Not a prototype. Working software.

< 10s
Response time
5
Governance layers
8+
Audit events per interaction
Real-time
Approval gates
Case Study #1

Governed AI in Microsoft Teams

A real user asks a real agent a real question — and every step is identity-verified, policy-checked, and hash-chained. Here's exactly what happens.

T+0:00M365 Identity Pipeline

Agent Provisioned

Sage BA provisioned with a real Microsoft 365 identity — UPN, mailbox, Teams presence, license, and group memberships — via an 8-step governed provisioning pipeline.

User Principal Name (UPN) created in Entra ID
Exchange Online mailbox provisioned
Teams presence activated
M365 license assigned
Security group memberships configured
Provisioning audit event hash-chained
T+0:01Teams Chat

User Messages Agent

A user opens a 1:1 chat with Sage BA in Microsoft Teams and sends: "What are the key priorities this quarter?"

Message received via Microsoft Graph subscription
Chat context loaded (conversation ID, user identity)
Inbound message logged as audit event
T+0:02Agent Resolver

Identity Resolved

The system resolves the M365 object ID to the enrolled agent identity. Verified the agent is active (not paused or quarantined) and authorized for this tenant.

M365 object ID → agent identity lookup
Agent lifecycle state verified: ACTIVE
Tenant enrollment confirmed
T+0:03Policy Engine

Policy Evaluated

Policy engine classifies the action as ANSWER_QUESTION_INTERNAL, assigns risk level LOW, and returns decision: AUTO — no human approval required.

Action type: ANSWER_QUESTION_INTERNAL
Risk classification: LOW
Decision: AUTO (approved automatically)
Policy evaluation logged as audit event
T+0:04Runtime Safety

Kill Switch Checked

Runtime confirms no active kill switch at any of the four scopes: Global, Department, Agent, or Run. Execution is cleared to proceed.

Global scope: CLEAR
Department scope: CLEAR
Agent scope: CLEAR
Run scope: CLEAR
T+0:05Governed Runtime

LLM Generates Response

Azure OpenAI (GPT-4o) generates the answer within the governed runtime. Input and output are bounded by the policy context. LLM call logged.

Prompt constructed with policy boundaries
Azure OpenAI GPT-4o invocation
Response generated within governed context
LLM audit event logged (model, tokens, latency)
T+0:06Hash-Chained Audit

Audit Event Written

SHA-256 hash-chained audit event appended to the immutable log — capturing the action, actor, tenant, timestamp, and cryptographic link to the previous event.

SHA-256 hash computed (action + actor + tenant + timestamp + prev_hash)
Append-only log entry written
Hash chain integrity maintained
Event filterable by: Request, Policy, LLM, Teams
T+0:08Teams Reply

Response Delivered

Agent replies in Teams using its own M365 identity via Microsoft Graph API. The user sees the response in their Teams chat — from a real, governed identity.

Response sent via Graph Chat API
Message attributed to Sage BA identity
Delivery confirmation logged
Total: 8+ hash-chained audit events for one interaction

What makes this different

The agent has a real M365 identity — not a webhook, not a bot framework token. It appears in Entra ID, has a mailbox, shows presence in Teams. Every response passes through a 5-layer governance stack (identity → policy → kill switch → audit → delivery) and produces 8+ hash-chained audit events for a single interaction.

Case Study #2

SharePoint Excel reading — governed file access in the flow of work

A user shares a spreadsheet link in Teams. The agent reads real data, computes real numbers, and returns real answers — every file access logged and policy-checked.

T+0:00

User Shares Excel Link

User pastes a SharePoint Excel file link into the Teams chat with Sage BA.

SharePoint URL detected in message
File enrichment pipeline initiated
T+0:01

URL Detected & Parsed

Agent detects the SharePoint URL pattern and initiates the file enrichment pipeline. The file path, site ID, and drive item are resolved.

SharePoint site ID resolved
Drive item identified
File metadata retrieved (name, size, modified date)
T+0:02

Policy: Auto-Approved

Policy engine classifies the action as READ_DOC_SUMMARY — risk: LOW — decision: AUTO. File access is approved without human intervention.

Action type: READ_DOC_SUMMARY
Risk classification: LOW
Decision: AUTO (approved automatically)
T+0:03

File Accessed via Graph

Agent accesses the Excel file via Microsoft Graph Workbook API using a delegated OAuth token — not a shared service account. Scoped to minimum required permissions.

Delegated OAuth token used (not service account)
Microsoft Graph Workbook API call
Worksheet data retrieved
T+0:04

File Access Logged

File access logged as a hash-chained audit event (type: File Read) — capturing filename, connector type, file hash, and the policy decision that authorized it.

Audit event type: FILE_READ
Filename, connector, file hash recorded
Policy authorization reference linked
SHA-256 hash chain maintained
T+0:06

Data Analyzed

Agent reads the Transactions sheet, calculates totals across categories, identifies top products by revenue — all within the governed runtime.

Transactions worksheet parsed
Category totals calculated
Top products identified by revenue
LLM generates natural language summary
T+0:08

Response with Real Data

Agent replies in Teams with actual numbers from the spreadsheet — not a template, not a pre-loaded answer. Real data, governed access, complete audit trail.

Response contains real computed values from the file
All numbers sourced from the actual Excel data
Evidence bundle generated automatically
T+0:08

Evidence Bundle Created

Complete evidence bundle generated: inputs, outputs, policy decisions, file access records, and SHA-256 checksums — exportable JSON, ready for audit.

Inputs: user message, file URL, file metadata
Outputs: agent response, computed values
Policy: action type, risk level, decision
File access: path, connector, hash, timestamp
Bundle checksum: SHA-256 integrity verification

Real data. Real governance.

The agent accessed a real Excel file, read real data, and returned real numbers — not a template, not a pre-loaded answer. Every file access was logged, policy-checked, and hash-chained. The evidence bundle includes inputs, outputs, policy decisions, and file access records — exportable JSON with SHA-256 integrity verification.

Governance Architecture

Five independent governance layers

Every agent interaction passes through all five layers. No shortcuts. No overrides. Each layer produces its own audit events.

Immutable Audit Trail

Every action hash-chained with SHA-256 — append-only, tamper-evident, cryptographically linked. One-click chain verification returns CHAIN INTACT or flags tampering.

  • SHA-256 hash chain (each event links to previous)
  • Filterable: Requests, Policy, Approvals, File reads, LLM, Teams/Email, Auth
  • One-click integrity verification
  • Exportable for external audit

Kill Switch

Instantly halt agent operations at four scopes: Global (all agents), Department, individual Agent, or single Run. Each activation is itself an audited event.

  • 4 scopes: Global, Department, Agent, Run
  • Instant activation — agents stop immediately
  • Every activation/deactivation is hash-chained
  • No silent overrides — full audit trail

Policy Engine

10 action types classified by risk. Four risk levels (Low, Medium, High, Critical) mapped to three decisions: Auto-approved, Requires Approval, or Blocked. Configurable per tenant.

  • 10 classified action types
  • 4 risk levels: Low, Medium, High, Critical
  • 3 decisions: Auto, Approval Required, Blocked
  • Per-tenant configuration

Approval Gates

Risk-based holds for sensitive actions. Approve or reject from the Operator Console with full context. Time-limited capability leases. Every decision audited.

  • Risk-based approval holds
  • Approve / reject with full context
  • Time-limited capability leases
  • Every approval decision hash-chained

Evidence Packs

Exportable JSON bundles per agent run — inputs, outputs, policy decisions, file access records, approval chains. SHA-256 checksum for integrity. Hand it to a regulator, not a screenshot.

  • Per-run JSON export
  • Inputs, outputs, policy, file access, approvals
  • SHA-256 checksum for integrity
  • Structured for regulatory submission

Role-Based Access Control

7 defined roles with module-level access control. Operators, admins, and auditors see different views. Role switches are logged. No privilege escalation without audit trail.

  • 7 roles with granular module access
  • Role switches are audited events
  • Operator Console scoped by role
  • Least-privilege by default
M365 Identity

Not a chatbot.
A governed employee.

SageOS agents are not webhook endpoints or bot framework registrations. Each agent is provisioned with a real Microsoft 365 identity — visible in Entra ID, manageable with your existing IT tools, and governed by the same policies as your human workforce.

Real M365 Identity

UPN, mailbox, Teams presence — appears in Entra ID

8-Step Provisioning Pipeline

Governed creation: identity, license, groups, mailbox, audit

Agent Lifecycle Management

Active, Paused, Quarantined states — controllable at any time

Delegated OAuth Tokens

No shared service accounts — scoped, auditable, revocable

IT-Manageable

Use your existing Entra ID, Intune, and M365 admin tools

What's next

Meeting CTO Agent

Coming Soon

Speech-to-text in Teams meetings. Agent joins calls, transcribes, extracts action items, and routes approvals — all governed.

Multi-Agent Orchestration

In Development

Multiple specialized agents collaborating through the orchestrator — with cross-agent audit trails and approval chains.

Integrations Catalog

Roadmap

Governed connectors for Jira, Salesforce, ServiceNow, and more — each with its own policy rules and audit events.

Ready to govern your AI workforce?

Every claim on this page is backed by working software. Book a demo and we'll walk you through every governance layer live.