Case study · Governed AI

AI-enabled operations, under governance.

LLM assistance integrated into enterprise operations through scoped tools, anonymized incident data and human approval — not an autonomous agent with production credentials.

Focus
Governed AI for operational support
Pattern
MCP tool integrations + RAG
Boundary
Read-only by default, approval for actions
Role
Automation expert
Support model
L1 / L2 operational support
  • MCP
  • RAG
  • FastAPI
  • Python
  • Microsoft Copilot
  • Power Automate
  • ServiceNow
  • Azure-hosted containers
  • HashiCorp Vault
  • CyberArk
  • LDAP

This is not autonomous AI. Tool access is explicitly scoped, actions that change state require human approval, and every retrieval and response is recorded.

The problem

Support teams spend much of an incident gathering context: which system, which change, which prior ticket, which runbook. That work is well suited to retrieval and summarisation. What is not acceptable in a regulated environment is handing a model broad access to production systems, sending incident data to a third party unfiltered, or acting on a recommendation nobody can trace.

Approach

Tools, not credentials

Capabilities are exposed as narrow MCP tools with defined inputs and outputs. The model never holds a credential.

Anonymization first

Incident data is stripped of identifying detail before it reaches a model.

Retrieval over approved sources

RAG is scoped to sanctioned internal documentation, not the open internet.

Ansible-backed actions

Where an action is permitted, it runs as reviewed Ansible automation rather than ad-hoc commands.

Human approval

Any state-changing operation waits for an operator decision.

Recorded end to end

Prompts, retrievals, tool calls and outcomes are logged for review.

Implementation

Services

  • FastAPI orchestration layer between assistants and tools
  • MCP tool definitions per capability
  • ServiceNow integration for incident context and records
  • Ansible-backed services for permitted operations
  • Azure-hosted containers for the service layer
  • Power Automate and Copilot workflows at the user surface

Security boundaries

  • Secrets from Vault and CyberArk, never in prompts or content
  • Identity resolved through LDAP
  • Read-only tools by default; write paths explicit and audited
  • Anonymization applied before model invocation
  • Scoped tool permissions per use case

What it changed

  • Context gathering during incidents became faster and more consistent
  • Operators kept the decision, with the reasoning visible to them
  • Sensitive incident detail stayed inside the anonymization boundary
  • Every AI-assisted action left a reviewable record

TODO — confirm which quantitative outcomes may be published for this engagement.

Related case studies

  • UN Automation Platform

    Multi-team AWX platform serving United Nations organizations.

    Read case study →
  • AWX Platform Modernization

    Legacy AWX rebuilt on the operator with GitOps configuration-as-code.

    Read case study →
  • Identity and Lifecycle Automation

    Joiner, mover and leaver automation with auditable records.

    Write-up in progress

Want AI in operations without losing control?

The design question is not which model. It is which tools it may call, on whose authority, and what evidence remains afterwards.