Init Intelligence Atlas

Init Intelligence Engineering Build Playbook for AI ITSM

3 min read

Init Intelligence Engineering Build Playbook for AI ITSM

Principle

Learn only what lets Init Intelligence build the next slice. The engineering path should produce product artifacts, not abstract expertise.

The first useful prototype should be: AI access/onboarding operator with context graph, approval policy, deterministic execution, and audit trail.

Build Sequence

1. Request Object

Build a normalized request object before building agents:

  • requester,
  • target user,
  • service/app/device,
  • requested outcome,
  • required approvals,
  • policy checks,
  • proposed actions,
  • execution trace,
  • system-of-record sync.

Read: agent-first-itsm, outcome-automation-vs-step-automation

2. Access / Onboarding Workflow

Build one bounded workflow: grant app access, onboard a user, or offboard a user.

The workflow should include:

  • Slack/Teams-style intake,
  • manager or policy approval,
  • IdP/group action stub,
  • ticket/evidence writeback,
  • audit trace.

Read: initlabs-engineering-enterprise-it-integration-substrate-2026-04

3. Context Graph v0

Use Postgres tables or simple graph edges at first. Model:

  • users,
  • managers,
  • groups,
  • apps,
  • devices,
  • tickets,
  • ownership,
  • policies,
  • source and timestamp per fact.

Read: context-graph, integration-and-context-layer-for-ai-itsm

4. Deterministic Execution

Pick a durable execution substrate. For founder speed, start with Trigger.dev, Inngest, or Hatchet. Study Temporal as the reference model.

Every step should be:

  • typed,
  • idempotent where possible,
  • logged,
  • retryable,
  • approval-aware,
  • safe to resume.

Read: deterministic-agent-runtime

5. Agent Planning

Add the agent after the workflow exists. The agent should classify, gather context, draft a plan, ask missing questions, and propose tool calls.

Do not let the first agent execute broad arbitrary actions. Keep agent autonomy inside a narrow tool list.

Read: initlabs-engineering-agent-frameworks-observability-2026-04

6. Tool Governance

Build a thin tool gateway:

  • tool registry,
  • input/output schemas,
  • credential reference,
  • required scopes,
  • policy check,
  • approval rule,
  • audit output.

Study Arcade, Composio, OpenFGA, OPA, and Cedar.

Read: agent-tool-governance

7. Customer Trace

Make every workflow produce a trace that a customer can trust:

  • request summary,
  • context used,
  • policy decision,
  • approver,
  • tool call,
  • external result,
  • evidence written,
  • human override if any.

This trace is the bridge between managed blackbox delivery and enterprise trust. ^[inferred]

Non-Negotiable Engineering Ideas

  • Ticket is audit record, not workflow engine.
  • Context graph is runtime substrate, not diagram.
  • Living playbooks capture process logic, not just retrieval text.
  • Agents propose; deterministic runtime executes.
  • Every privileged action needs policy, approval, credential scope, and trace.
  • System integrations are the product surface.
  • Readiness debt is product opportunity.

What To Study While Building

First Product Slice

The best first slice is secure access + onboarding/offboarding:

  • It is high-frequency.
  • It crosses HRIS, IdP, SaaS, MDM, Slack/Teams, ticketing, and compliance.
  • It has clear approval and audit requirements.
  • It naturally expands into devices, security hygiene, compliance evidence, and broader back-office workflows.

Graph View

Backlinks