How to Inventory AI in 90 Days: A Practical Discovery Playbook

Every AI governance program has the same prerequisite: you need to know what AI systems your organization is running before you can govern them. Without an inventory, you cannot assess risk, assign ownership, apply policy, or demonstrate compliance.

The problem is that building an AI inventory from scratch feels like an impossible project. AI capabilities are embedded in existing software. Employees use dozens of tools not sanctioned by IT. Browser-based AI tools leave no installation footprint. The inventory is never complete because the landscape keeps changing.

This playbook gives you a 90-day process to build a defensible, actionable AI inventory — not a perfect one, but a working one that can be maintained and updated.

Why 90 Days?

Ninety days is enough time to apply multiple discovery methods, validate findings, and produce an inventory that is useful for governance purposes. It is fast enough to be practical given the urgency of the risk — 98% of organizations already have shadow AI usage — and structured enough to avoid the "AI inventory project" that sits in planning for a year.

The output is not a spreadsheet gathering dust. It is a living document tied to your governance program.

Phase 1: Passive Discovery (Days 1-30)

Start with data your organization is already collecting.

DNS and Proxy Log Analysis

Your network traffic logs contain the record of every domain your organization has connected to. Known AI service endpoints — api.openai.com, claude.ai, api.anthropic.com, copilot.microsoft.com, bard.google.com, gemini.google.com, and dozens of others — appear in those logs.

What to do:

• Pull 90 days of DNS query logs and proxy logs
• Filter for known AI service domains (build a list from major providers plus any tools mentioned in employee feedback)
• Aggregate by user or workstation and by volume
• Flag high-volume users and high-frequency tools for follow-up

This step typically surfaces the majority of AI usage, since most AI tools are web-based. The limitation is that it identifies domains, not use cases — you will know that a user is connecting to claude.ai but not what they are doing there.

Software and Browser Extension Inventory

For managed endpoints, run an inventory of:

• Installed applications with "AI," "assistant," or "copilot" in the name
• Browser extensions — AI-powered writing assistants, summarizers, and productivity tools are extremely common and often not visible to IT unless actively monitored

Intune, Microsoft Endpoint Configuration Manager, or equivalent MDM tools can produce this inventory for managed devices. Note that unmanaged devices (personal laptops used for work) will not appear in this inventory — an important gap to document.

Expense Report Review

AI tool subscriptions often appear in expense reports as individual software purchases. A query against expense data for payments to known AI vendors (OpenAI, Anthropic, Midjourney, Jasper, etc.) surfaces individually-purchased tools that did not go through IT procurement.

This also identifies employees who are paying for productivity tools that IT has not approved — often because they have legitimate needs that the approved toolset does not meet, and the request process is too slow.

SaaS Procurement Review

AI capabilities are increasingly embedded in existing SaaS platforms: Salesforce Einstein, HubSpot AI, ServiceNow AI, Workday AI, and many others. These are not "shadow AI" — they are features of tools already in use — but they are AI systems that require governance. Your procurement data identifies which SaaS vendors have AI capabilities that may be enabled.

Phase 2: Active Discovery (Days 31-60)

Passive discovery finds what is visible in logs. Active discovery finds what employees know.

Anonymous Employee Survey

A short, anonymous survey with specific questions surfaces tool usage that technical discovery misses:

• What AI tools do you use for work, inside or outside company-approved tools?
• What tasks do you most frequently use AI for?
• Have you ever input customer data, financial data, or company strategy into an AI tool?
• What AI tools would help you work more effectively that you do not currently have access to?

The last question is the most valuable for governance design — it tells you where demand exceeds supply in your approved toolset, which is where shadow AI pressure is greatest.

Survey results are qualitative signal, not a complete inventory. Use them to guide follow-up interviews and to validate or expand technical discovery findings.

Department-Level Interviews

Work with HR, Finance, Legal, Engineering, and Sales to interview representatives from each team:

• What AI tools are team members using?
• What workflows have they built that incorporate AI?
• What AI capabilities are embedded in the tools the team uses?
• What are the data inputs to those workflows?

Department interviews surface context that logs cannot provide: workflows, use cases, and data flows. They also build relationships with department leads that support the governance implementation in Phase 3.

IT Help Desk Ticket Review

Review the last 90 days of help desk tickets for any mention of AI tools: requests for access, questions about approved tools, complaints about blocked sites. This is a lightweight source of signal that costs almost nothing to collect.

Phase 3: Inventory Build and Risk Scoring (Days 61-90)

Consolidate findings from both phases into a structured inventory.

Inventory Schema

Each AI system in the inventory should capture:

• Tool name and version/tier
• Vendor and data processing agreement status (DPA in place / not in place / pending)
• Deployment scope (number of users, departments)
• Data classification (what categories of data have been or may be input)
• Use cases (documented workflows where the tool is used)
• Governance status (approved / conditionally approved / under review / prohibited)
• Owner (the person responsible for this tool's governance)
• Risk score (based on data classification + DPA status + deployment scope)

Risk Scoring

A simple risk matrix:

• High risk: Handles confidential or regulated data; no DPA; wide deployment
• Medium risk: Handles internal data; no DPA or limited DPA; moderate deployment
• Low risk: Handles public data only; DPA in place; limited deployment

High-risk tools require immediate action: either a DPA is established and data classification rules are documented, or the tool is blocked with a documented rationale and an approved alternative offered.

Governance Mapping

For each tool in the inventory, map it to your AI AUP: which tier does it fall into? What data classification rules apply? Does it require any additional controls?

Tools not currently on the approved list that have legitimate use cases and can meet data protection requirements should be added through your approval process — not left in limbo, where employees will continue using them without any governance.

The 90-day deadline creates pressure to actually finish. An inventory that takes 18 months to build is an inventory that does not exist when you need it.

Want support building your AI inventory and turning it into a governance program? Talk to JP Stratton.