Enterprise framework

Enterprise Framework for MCP and Agentic AI Governance

Practical guidance for safe enterprise AI adoption, operational control, platform enablement, and assurance where AI agents connect to tools, data sources and business workflows.

Principles

Core governance principles

These principles define the minimum operating posture for enterprise MCP and agentic AI adoption.

01

Bounded purpose

Every agent has a named business purpose, approved user group, permitted systems and prohibited actions.

02

Read-only first

New agents begin with retrieval, summarisation or drafting before write actions are considered.

03

Action tiering

Tool calls are classified as read, write, destructive, privileged or high-impact before release.

04

Human control

High-impact and irreversible actions require human approval, rollback and accountable ownership.

Architecture

MCP enterprise architecture and discovery

The enterprise pattern separates the user experience, AI orchestration, MCP protocol connection, governance gateway, approved MCP servers and downstream systems.

MCP overview: Model Context Protocol is an open interoperability protocol that lets AI applications connect to external tools, resources and prompt templates through MCP servers. In an enterprise, MCP should be treated as the connection standard, while the governance gateway and platform controls decide which connections are allowed, monitored and evidenced.

Reference architecture pattern

All agent actions pass through a governed platform layer before reaching enterprise systems. The gateway is the enterprise control point for identity, policy, approval, monitoring and evidence.

Gateway firstCatalogue controlledRead-only firstEvidence by default
Experience layerWhere the user works
User

Employee or process

Starts a request from chat, Copilot, portal, workflow or business application.

Host

AI client / agent host

Manages the AI experience and determines which approved agent capabilities are available.

Governed AI platform layerCentral operating control
MCP client

Protocol session

Connects the AI host to a specific MCP server using the MCP protocol.

Gateway

Enterprise MCP Gateway

Enforces identity, action tier, approvals, routing, rate limits, monitoring and audit logging.

Catalogue

Agent and server catalogue

Records approved agents, MCP servers, tools, data classes, owners and release status.

MCP capability layerApproved connector boundary
Server

Approved MCP servers

Managed connectors exposing approved tools, resources and prompt templates.

Tools

Actions

Search, create ticket, update record, trigger workflow or call an approved API.

Resources

Context

Documents, records, knowledge entries or datasets retrieved within access boundaries.

Prompts

Patterns

Reusable prompt templates that standardise approved interaction flows.

Enterprise systems layerSystems of record and action

Collaboration and knowledge

Microsoft 365, SharePoint, Confluence, knowledge bases and policy stores.

Business platforms

CRM, ERP, HR, finance, customer service and operational workflow platforms.

Engineering and operations

Git repositories, DevOps tools, ITSM platforms, monitoring and cloud APIs.

IdentityUser context and agent identity
PolicyAllowed server, tool and action tier
ApprovalHuman control for high-impact actions
MonitoringRuntime logs, anomalies and incidents
EvidenceTraceable release and audit record
Architecture rule 1 — No direct system accessProduction agents do not call enterprise systems directly. They call approved MCP servers through the enterprise MCP gateway or equivalent governed platform layer.
Architecture rule 2 — Catalogue before connectionAn MCP server, tool or prompt template is not available to agents until it is registered with owner, purpose, action tier, data class and release status.
Architecture rule 3 — Evidence by designEvery agent action must produce a traceable evidence record showing identity, server, tool, action tier, approval status, result and monitoring outcome.
User requestintent and identity
Agent hostapproved experience
MCP clientprotocol session
Gatewaypolicy and access
MCP servertools/resources/prompts
Evidencelogs and review

MCP component overview

Host

The AI application or agent environment that presents the experience to users and decides which approved capabilities are available.

Client

The protocol component inside the host that opens and manages the MCP connection to one server at a time.

Server

The governed connector that exposes approved capabilities to the agent. It sits in front of the actual business system.

Tools, resources and prompts

Tools allow actions, resources provide context, and prompts provide reusable interaction templates. Each requires ownership, classification and release control.

Enterprise MCP discovery setup

1

Discover servers

Maintain an enterprise MCP catalogue covering internal, vendor and approved open-source servers.

2

Discover capabilities

Record each server's tools, resources, prompts, action tier, data class and approved caller list.

3

Discover authorization

For protected remote MCP servers, align clients with the correct authorization server and access policy.

4

Register through gateway

Route production MCP traffic through the gateway so identity, policy, HITL, rate limits and logs are enforced.

5

Review lifecycle

Retire or block servers without owner, logs, current review, approved purpose or passing release evidence.

Enterprise scenarios

Apply the framework

Select a scenario to view the governance question, default decision pattern, platform action and role involvement.

Risk and controls

Control model

Controls are applied to the agent, the MCP server, the tool action and the resulting evidence record.

Control domainEnterprise requirementImplementation patternEvidence artefact
Identity and accessEach agent and user context is attributable.Managed identity, OAuth, RBAC, least privilege, no shared service accounts.Access policy, identity record, approval history.
Tool action tieringActions are classified before release.Read, write, destructive, privileged and high-impact tier model.Tool tier dictionary, release gate record.
Prompt injection protectionUntrusted content is not treated as trusted instruction.Input separation, adversarial tests, content filtering, restricted tool invocation.Test pack, failed case log, remediation record.
Human controlHigh-impact actions are approved and reversible where possible.Human-in-the-loop, rollback path, break-glass ownership.Approval record, rollback procedure, incident owner.
Monitoring and auditAgent actions are observable and reviewable.Structured logs, correlation IDs, SIEM integration, anomaly alerts.Audit log, dashboard, exception report.

Action tier dictionary

TierExamplesDefault decision pattern
ReadSearch knowledge base, fetch policy, list incidents, query approved view.Allowed after identity, access and logging controls are confirmed.
WriteCreate ticket, draft case note, update non-sensitive field.Allowed after pilot if scope is narrow, logged and reversible.
DestructiveDelete record, close case, remove file, cancel workflow.Human approval required; rollback and audit record mandatory.
PrivilegedChange access, run admin action, alter configuration.Privileged access controls and explicit approval required.
High-impactAction affecting employment, finance, customer entitlement, safety or legal position.Human decision ownership required; agent may support but not independently decide.
Incident response

Agent incident response model

Agentic systems require an operational response path because failures may involve tool actions, data exposure, incorrect records, workflow disruption or prompt injection in production.

1. Detect

Trigger from SIEM alert, gateway anomaly, user report, support ticket, cost spike, failed evaluation or unexpected tool-call pattern.

2. Contain

Disable the agent, block the MCP tool, revoke token, pause the gateway route, or switch the agent to read-only mode.

3. Preserve evidence

Capture correlation ID, prompt context, tool call, user identity, MCP server, response, timestamp and affected system record.

4. Assess impact

Classify data exposure, wrong action, customer impact, operational disruption, compliance impact and rollback feasibility.

5. Remediate

Reverse records where possible, notify accountable owners, patch prompts/tools, update access policy and retest failed scenarios.

6. Reopen or retire

Return to production only after release gates pass again. Retire the agent or tool where residual risk is not acceptable.

Incident typeMinimum responseRequired evidence
Wrong record updatedDisable write tool, identify affected records, rollback or correct, notify business owner.Tool call log, record diff, approval record, rollback evidence.
Prompt injection suspectedFreeze agent route, preserve prompt and retrieved content, run adversarial retest before release.Prompt context, retrieved source, failed test case, remediation record.
Workflow crash or runaway call patternTrigger circuit breaker, pause gateway route, review rate limits and retry behaviour.Call volume, error rate, timeline, system impact, corrected threshold.
Data exposureStop affected retrieval path, classify exposed data, involve privacy/security owner.User identity, resource accessed, classification, recipients, containment action.
Governance pattern

Accountability model

Governance is distributed across accountable owners, platform teams and assurance functions. The platform layer provides consistent controls.

AI Experience & Platform Enabler Team

Owns the governed AI platform layer: agent catalogue, approved MCP/connectors, prompt patterns, release gates, operational telemetry, support model and evidence packs.

Business owner

Owns the process outcome, business risk, user group, acceptable action scope and adoption outcome.

Security, privacy and data partners

Validate access, data handling, prompt injection controls, monitoring, DLP and privacy requirements.

Architecture and operations

Maintain reference patterns, platform standards, incident path, support runbooks and lifecycle operations.

Evidence and assurance

Evidence model

The evidence model converts governance into artefacts that can be reviewed by risk, audit, architecture and operations.

Evidence areaRequired artefactsOwner
Agent inventoryAgent name, owner, purpose, user group, systems connected, action tiers.AI Experience & Platform Enabler Team
ArchitectureReference diagram, data flow, MCP server list, integration dependency map.Architecture lead
ControlsAccess policy, approval gate, test result, monitoring rule and rollback path.Platform, security and operations
OperationsSupport model, incident workflow, known issues, service ownership, lifecycle status.Service operations
FinOpsUsage dashboard, quota settings, budget alerts, licence assignment, API and connector consumption.Platform and FinOps owners
Review recordRelease decision, exceptions, residual risks and next review date.Governance forum
FinOps and usage governance

Cost and capacity control

Multi-vendor agentic AI introduces licence, token, API, connector and infrastructure cost. Cost governance should be part of the platform layer, not an afterthought.

Usage visibility

  • Track agent sessions, model calls, tool calls, connector usage and failed calls.
  • Attribute consumption to product, team, environment and business owner.
  • Separate pilot, production and experimental usage.

Cost controls

  • Set quotas and budget alerts for API, token and connector usage.
  • Apply rate limits for expensive or high-volume tools.
  • Require approval for agents that call paid external APIs or high-cost models.

Value review

  • Review active use, repeat use, support demand, quality failures and avoided manual effort.
  • Retire unused agents and duplicated connectors.
  • Prefer shared MCP servers and reusable prompt/tool patterns over one-off builds.
Decision framework

Release gates and default decision path

The release decision is based on capability, action tier, data exposure, operational support and evidence. A gate that fails blocks release until remediated or formally accepted by the accountable risk owner.

1

Classify the use case

Purpose, users, data, systems, autonomy level, business impact and accountable owner.

2

Classify the action tier

Read, write, destructive, privileged or high-impact. The highest tier drives the release gate.

3

Apply release gates

Access, injection, schema, monitoring, approval, support, rollback and FinOps checks.

4

Operate with evidence

Catalogue status, logs, support pathway, incident handling, cost tracking and lifecycle review.

Minimum release gates

  • Named business owner and platform owner recorded.
  • Agent and MCP server registered in the catalogue.
  • Data class and action tier recorded for every tool.
  • Access policy validated against least privilege.
  • Support category and incident owner available before go-live.

Pass/fail thresholds

  • No high-severity prompt injection or tool misuse test remains open.
  • All tool schemas validate expected, boundary and invalid inputs.
  • All destructive, privileged and high-impact actions require human approval.
  • Monitoring confirms correlation ID, user identity, agent identity and tool name for each action.
  • Rollback or containment path is documented for every write-capable tool.

Action-tier release rule

  • Read: release after access, logging and data-boundary checks pass.
  • Write: release after pilot, reversibility and support checks pass.
  • Destructive: release only with approval workflow, rollback and break-glass owner.
  • Privileged: release only with privileged access controls and additional security review.
  • High-impact: release only with accountable executive approval and monitoring review.

Non-release conditions

  • Unregistered MCP server or unmanaged connector.
  • Shared service account with no attributable identity.
  • External or user-generated content used without injection testing.
  • No incident response path for production failure.
  • No method to monitor usage, cost, errors or exceptions.
Operating model

Roadmap and team responsibility

Implementation begins with discovery, then platform setup, read-only pilots and scaled operations.

1. DiscoveryIndicative duration: 2–4 weeks
  • Inventory agents, AI tools and MCP servers.
  • Identify data classes, action tiers and business owners.
  • Confirm platform owner and governance forum.
2. Platform setupIndicative duration: 4–8 weeks
  • Establish catalogue, approved connector path and release gates.
  • Define action tier dictionary and evidence model.
  • Integrate monitoring and support categories.
3. Read-only pilotIndicative duration: 4–6 weeks
  • Start with retrieval, summarisation and drafting scenarios.
  • Test prompt injection, tool boundaries and access controls.
  • Capture adoption, issues and evidence.
4. Scaled operationsIndicative duration: ongoing, reviewed quarterly
  • Introduce controlled write actions with approval gates.
  • Publish patterns, runbooks and platform dashboard.
  • Review exceptions, incidents and lifecycle status.

Where the team comes in

The AI Experience & Platform Enabler Team owns the reusable platform layer and co-ordinates the release path. Security and data partners validate controls. Business owners approve the purpose and risk. Operations run support and incident workflows. Architecture maintains reference patterns.


Steady operating rhythm: platform backlog, fortnightly release review, monthly risk and controls forum, monthly usage and cost review, quarterly evidence and roadmap review.

Australian AI alignment

Australian governance alignment

The framework maps to Australian responsible AI practices and agentic AI security guidance through accountable ownership, risk controls, testing, monitoring and human control.

Australian guidance areaFramework alignmentOperational artefact
Decide who is accountableNamed agent owner, business owner and platform owner.Agent catalogue record.
Understand impactsUse case classification, data class and action tier.Impact review summary.
Measure and manage risksControl gates, access policy, prompt injection testing and monitoring.Risk and controls record.
Share essential informationUser disclosure, support route and operational documentation.Service page and support article.
Test and monitorRelease tests, telemetry, audit logs and anomaly alerts.Test result and dashboard.
Maintain human controlApproval gates for destructive, privileged and high-impact actions.Approval and rollback record.
Reference patterns

Implementation reference patterns

These patterns define how enterprise teams should implement and govern MCP-connected agents.

Pattern 1: Catalogue before connection

No agent or MCP server should be available for enterprise use until it has a catalogue record, owner, purpose, data class and action tier.

Pattern 2: Read before write

Read-only use cases establish data access, quality, monitoring and support pathways before write actions are enabled.

Pattern 3: Gateway before system access

Agent traffic should pass through a governed gateway or control layer before reaching MCP servers and enterprise systems.

Pattern 4: Test adversarial content

Prompt injection and tool misuse tests are required for agents that read user-generated or external content.

Pattern 5: Evidence by design

Logs, approvals, test outcomes and exceptions are captured during normal operation, not reconstructed later.

Pattern 6: Vendor-neutral platform layer

The same catalogue, control gate and evidence model applies across Microsoft, ServiceNow, Salesforce, GitHub, Atlassian, Snowflake and internal systems.

Pattern 7: Incident path before production

Every production agent requires a defined containment, rollback, evidence preservation and re-release path.

Pattern 8: Cost control by design

API, token, licence and connector costs are monitored and attributed as part of normal platform operation.

Source links

Model Context Protocol specification

Authoritative protocol reference for MCP servers, clients, tools, resources, prompts and authorization.

Open official MCP specification

MCP introduction

Official overview describing MCP as an open standard for connecting AI applications to external systems.

Open MCP introduction

MCP authorization

Official authorization guidance, including protected resource metadata used by clients to discover authorization servers.

Open MCP authorization specification

Australian Guidance for AI Adoption

Australian Government guidance covering accountability, impact, risk, information sharing, testing, monitoring and human control.

Open Australian AI adoption guidance

ACSC / Five Eyes agentic AI guidance

Guidance for careful adoption of agentic AI services, including security and operational risk considerations.

Open agentic AI guidance

Microsoft 365 Agent Registry

Microsoft documentation for agent visibility and lifecycle management in the Microsoft 365 admin centre.

Open Microsoft Agent Registry