BJM.
Back to Projects

Mid-Market SaaS Company

Agentic IT Onboarding Platform

Replaced a 3-day manual IT onboarding process with an AI-driven agentic workflow that provisions access, creates accounts, and generates documentation in under 15 minutes.

Claude API TypeScript Azure Functions Microsoft Graph ServiceNow API SharePoint

Key Outcomes

  • Reduced IT onboarding time from 3 days to under 15 minutes
  • Eliminated 95% of manual ticket creation
  • Full audit trail for SOC 2 compliance requirements
  • Handles 50+ onboardings per month autonomously

The Problem

Every new hire triggered a cascade of manual work: AD account creation, license assignment, SharePoint access provisioning, ServiceNow ticket chains, and documentation generation. Each step was manual, error-prone, and required coordination across three teams.

The Solution

An agentic workflow that treats the onboarding checklist as a planning problem, not a scripting problem.

The system receives a structured onboarding request (role, department, start date, manager) and autonomously:

  1. Plans the required provisioning steps
  2. Invokes the appropriate system APIs in the correct order
  3. Handles permission dependencies and retry logic
  4. Generates a completion report with full audit trail
  5. Escalates to humans only for decisions above its confidence threshold

Architecture

The orchestration layer uses a React-based admin dashboard backed by Azure Functions. Each integration (Graph API, ServiceNow, SharePoint) is wrapped as a typed tool with explicit schemas, making the agent’s behavior predictable and auditable.

The LLM component handles only the planning and decision-making — all actual system mutations happen through deterministic code paths, not LLM-generated code execution.

What Made It Production-Safe

Hard limits on what the agent can and cannot do were the most important design decision. The agent can assign licenses and create accounts, but cannot modify Global Admin roles, access financial systems, or take irreversible actions without human confirmation.