Ask your IT team how many AI tools your company uses. You’ll probably get a rough estimate and a few nervous laughs. An AI asset inventory fixes that. So, what is an AI asset inventory? It’s a single list of every AI asset in use, from AI apps, agents and models to the data, platforms and computing resources behind them.
Most companies can’t see their AI assets clearly yet. In the Flexera 2026 State of ITAM Report, only 31% of respondents said they had accurate visibility into AI software. And 59% said their wasted AI spend had grown over the past year. Security teams see the same gap from another angle.
In this guide, we’ll start with what counts as an AI asset and how it differs from the software you already track. Then we’ll show you how to build an AI asset inventory: the fields each record needs, how to track each type of AI and where to find the AI nobody told you about. We’ll finish with how to turn your AI asset inventory into controls and keep it up to date.
What is an AI asset?
An AI asset is a model or anything that calls a model, hosts one or bills you for AI use. That’s a broad definition, and it should be. If your AI asset inventory only covers the AI someone bought, it’ll miss a lot of what’s running.
Most AI assets fall into six areas:
| What it is | What it includes |
| AI applications | Chat assistants, coding assistants, meeting and writing tools, AI features inside software as a service (SaaS) apps |
| AI agents | Agents that vendors ship inside apps you already use, plus agents your teams build in low-code builders, sales and service platforms, workflow tools or custom code |
| AI models | Foundation models you reach through an API, open-weight models, fine-tuned versions, machine learning models your data teams train in-house |
| AI data clouds | AI functions, agents and model serving inside data platforms such as Snowflake and Databricks |
| AI compute | GPU and other accelerator instances, inference endpoints, Kubernetes clusters that serve models and laptops that run local AI models |
| AI platforms | Managed services such as Amazon Bedrock, Microsoft Foundry and Google’s Gemini Enterprise Agent Platform |
Those six areas cover the AI assets you buy and run. Four more types of records tie them together, and they’re where AI governance tends to break:
- Identities and credentials: the user accounts, service accounts, API keys, OAuth tokens and agent identities that make AI calls
- Tools and connectors: Model Context Protocol (MCP) servers, plugins, browser extensions and other links that let AI read or change your systems
- Data sources: retrieval-augmented generation (RAG) document sets, vector databases, training data and stored prompts and responses
- Contracts and terms: pricing units, commitments, data-use clauses and renewal dates
If something holds a credential, touches company data and can call a model, it belongs in your AI asset inventory, even if nobody paid for it. So why can’t your existing software inventory just absorb all of this?
The difference between AI assets and traditional software assets
That’s not a knock on IT asset management (ITAM). ITAM teams have handled shadow IT and SaaS sprawl before, and AI asset management can use the same playbook: find it, normalize it, assign an owner and optimize it. What’s different is the AI asset itself. AI breaks the usual software model in four ways.
You pay by the token, not the seat
A standard software license counts seats, users or devices. Many AI vendors bill by use instead, or on top of seats: per token, credit, action or workflow run. So with usage-based pricing, costs keep climbing as long as people and agents keep working. OpenAI, Anthropic and Google, for example, use token-based pricing for API use, so AI token costs rise with every input and output token. Salesforce bills Agentforce actions in Flex Credits. Microsoft charges Copilot Studio agents in Copilot Credits. One employee can use five different AI tools and generate five different invoices, each in a different unit. No wonder token bills are becoming the new cloud bill.
Your heaviest AI user might not be human
Agents, workflows and service accounts, often running on shared API keys, drive a growing share of AI usage. And these non-human identities are multiplying fast. Palo Alto Networks’ 2026 identity security research found an average of 109 machine identities for every human. Of those, 79 are AI agent identities. Gravitee surveyed more than 900 executives and technical practitioners for its State of AI Agent Security 2026 report. It found that 45.6% of teams still rely on shared API keys for agent-to-agent authentication. A shared key tells you a cost happened. It can’t tell you who caused it.
AI shows up in software you already own
Vendors keep adding embedded AI features to products you already license, and some switch them on by default. Your contract stays the same, but your data may now flow to a model. The National Institute of Standards and Technology (NIST) saw this coming. The NIST Generative AI Profile (NIST AI 600-1) suggests a written policy on which built-in AI features, if any, you’ll leave out of your AI asset inventory. In other words, decide what’s in scope on purpose.
AI models and APIs change fast
AI providers release new model versions, retire old ones and shut down whole APIs on their own timelines. Take the Assistants API shutdown. OpenAI turned that API off on August 26, 2026, a year after it announced the change. On October 23, 2026, it also retired GPT-4, GPT-4 Turbo and o1 from its API, three of this year’s OpenAI model deprecations. Any code still calling a retired model or API gets an error instead of an answer. If your AI asset inventory can’t tell you within minutes which apps and agents depend on it, a support ticket will.
There’s one more difference, and it’s the biggest. Unused software licenses mostly waste money. Unmanaged AI assets can also expose sensitive data and take actions on their own. That’s why regulators, auditors and boards now want to see the list.
Why you need an AI asset inventory
Three pressures make this urgent: rising AI bills, growing security risks and new rules. They beg the question: who’s driving your AI costs?
With usage-based pricing, your AI bill moves every time someone, or something, sends a request. If you can’t connect that usage to a specific AI asset and its owner, you can’t explain the bill or spot waste. And most companies aren’t there yet. The same Flexera research found that only 29% of organizations measure the value of their AI software.
Your AI asset inventory gives finance and engineering a starting point for questions like these:
- Which models and apps drive the highest costs?
- Which agents burn tokens or credits without delivering enough value?
- Are teams paying for overlapping tools?
- Which workloads could run on a smaller, cheaper model?
- Who owns that sudden spike in usage?
What should an AI asset inventory include?
Start with fields that answer six basic questions:
- What is the asset
- Who owns it
- What does it do
- What can it access
- How much does it cost
- How risky is it?
Then add detail as risk and complexity grow. A writing assistant doesn’t need the same record as an agent that can change production data. Use these 10 fields as your AI asset inventory checklist:
| To track | Details needed | Why? |
| Asset identity | Unique ID, name, category, vendor, product, edition and version | Prevents duplicates and helps you match data from different sources |
| Ownership | Business owner, technical owner, responsible team and cost center | Gives finance, IT and security a person or team to contact |
| Purpose | What it’s for, who uses it and any decisions about people it helps make | Shows why the asset exists and what could go wrong |
| Model dependency | Model provider, model ID, version or alias, endpoint and hosting method | Shows which apps depend on a model and what breaks when it changes |
| Data handling | Data categories, sensitivity, source, destination, retention period, processing region and vendor terms | Helps you assess privacy, confidentiality, data residency and intellectual property risks |
| Identity and access | User identities, service accounts, app identities, credentials, permissions and connected systems | Shows who or what can reach the system and its resources |
| Agent permissions | Available tools, read and write access, transaction rights, approval gates and shutdown method | Separates a read-only assistant from an agent that can change business records |
| Risk and compliance | Risk tier, approval status, applicable rules, assessment results, open findings and incidents | Shows what you’ve reviewed and what’s still open |
| Usage and cost | Billing unit (seats, tokens, credits, actions, workflow runs or compute hours), commitments, renewal date, monthly spend and budget alerts | Supports AI cost attribution, forecasting and optimization |
| Lifecycle | Stage (requested, pilot, live, retiring or retired), last review date and any known end-of-support dates | Keeps outdated records from passing as active systems |
That might seem like a lot, but every field answers a question someone will ask about cost or risk. Copy the table into a spreadsheet, and you’ve got a starter AI asset inventory template. When it comes to generative AI, NIST Generative AI Profile recommends a few additional fields. Keep track of where the system’s data originated. Record any known issues that have been reported against the system. Note who is responsible for overseeing it. And record any special rights connected to intellectual property or sensitive data.
How to track different forms of AI
Apps are the most visible AI assets, but even here, a lot slips through.
Approved apps and personal accounts
Approved AI apps are the easy part. You bought them, so they already show up in your SaaS management tool, your single sign-on (SSO) logs and your contracts. Note which plan you’re on, too. Personal, team and enterprise versions of the same tool often handle and store your data differently.
Then come the apps nobody bought: unsanctioned AI tools that people sign up for on their own. Netskope’s Cloud and Threat Report 2026 found that 47% of people who use generative AI at work still use personal AI apps, down from 78% a year earlier. That’s progress, but a lot of company data still flows through accounts you don’t control. An AI tool inventory built only from purchase records misses all of it. So track that usage by app and department in your AI asset inventory, even if you can’t see individual prompts.
Embedded AI in tools you already pay for
Embedded AI is trickier, because vendors keep adding new AI features in SaaS apps you already use. A meeting app starts writing summaries. A help desk adds an agent. A design tool adds image generation. Each new feature can change where your data goes and what the vendor can do with it.
So run a short AI review for every SaaS app that adds an AI feature, and log the answers in your AI asset inventory. Ask:
- Which AI features does the product offer?
- Which ones are on, and can admins turn them off?
- Which model or provider powers each feature, if the vendor says?
- What data does the feature process or send elsewhere?
- Can the vendor use that data to train or improve models, and on what terms?
- Does the feature add usage charges, change permissions or need its own approval?
You may not be able to identify the underlying model for every vendor-managed feature. Record the details the vendor discloses and mark the rest as unknown. Do not invent model versions or assume that all features use the same provider.
Data sources
AI data sources include internal documents, operational databases, vector stores, public data, training datasets and prompt and response logs. Track each important source and link it to the models, apps and agents that use it. For each one, record:
- The data owner and source system
- Its classification and sensitivity
- Who and what can access it
- Any personal, regulated or confidential data it holds
- Retention rules and deletion steps
- Approved processing regions and destinations
- Where it flows, such as retrieval indexes, training datasets or AI outputs
- Vendor and contract restrictions
Managed AI platforms
Managed AI services can expose models, tools, safety settings, regions and access policies through a single account or project. Examples include Amazon Bedrock, Microsoft Foundry and Google’s managed AI services.
For each one, record the account, subscription or project, the models it can call, its regions, owners, permissions and service policies. Link it to the model deployments and workloads it runs, wherever the platform exposes those details.
A platform account is not a substitute for an inventory of its contents. Knowing that a team uses an AI service does not tell you which models it has enabled, which agents depend on them or which data those workloads can access.
So by now, you have likely noticed the pattern. Every AI asset record needs an owner, and ownership runs through identity.
How to turn your AI asset inventory into AI governance
Raw records turn into useful AI governance in three steps: clean them up, sort them by risk and connect them to controls.
Step 1: Clean up the names and versions
The same tool can show up under five different names. It’s a web domain in proxy logs and a merchant name on a card statement. It’s also an app ID in SSO, a line item on an invoice and a library in code. Models are even messier, since aliases, dated versions and cloud-specific IDs can all point to the same model. So match every AI asset record to one standard product name, edition and model version before you count anything. That’s ITAM data normalization, and it’s the step most homegrown AI inventories skip.
Step 2: Sort by risk, not by vendor
Give each AI asset a risk tier based on what it can access and what it can do. That’s where AI risk management gets practical. For your AI risk classification, four AI risk tiers work for most teams:
- Prohibited: uses you’ve banned, such as pasting regulated data into consumer AI tools
- High impact: AI that helps make decisions about people, such as hiring, credit or insurance, or that can change data in core business systems
- Standard: everyday productivity tools that only use approved types of data
- Low: tools that only use public data, can’t act on their own and don’t connect to sensitive systems
Step 3: Connect each record to a control
An AI asset inventory earns its keep when it changes what you can enforce. Here’s how common findings map to controls:
| Findings | Control |
| Unapproved application | Apply the organization’s access or application policy after validating the finding |
| Agent with excessive permissions | Restrict its identity, tools and data access |
| Usage without clear ownership | Require workload registration and correct identity mapping |
| Unexpected token or credit consumption | Set budgets, rate limits and spending alerts where supported |
| Unreviewed model or connector version | Require security review and approved versioning before rollout |
| Sensitive data sent to an external service | Apply data access policies, approved routing and appropriate data loss prevention controls |
| Upcoming contract renewal | Review usage, ownership, terms and alternatives before renewal |
Each control works at a specific layer. SSO can manage access to supported apps, but it can’t block every direct API request. An AI gateway can enforce policies for traffic that flows through it, but not for a model endpoint that bypasses it. Record those gaps in your inventory. Don’t assume you control everything from one place.
Your first 90 days: an AI asset inventory plan
You don’t need a perfect inventory before you start. Build a usable baseline and improve it as you learn more.
| Timeline | What to do | How to do it |
| Weeks 1 to 2 | Define the scope |
|
| Weeks 3 to 6 | Collect source data |
|
| Weeks 7 to 9 | Normalize and assign ownership |
|
| Weeks 10 to 12 | Apply risk and cost controls |
|
Eight AI asset inventory mistakes
1) Treating the inventory as a one-time project
AI products, model versions, integrations and prices keep changing. An AI asset inventory built for one audit goes stale within weeks. Feed it from live sources and reconcile it on a schedule.
2) Listing approved tools and applications only
A purchase list shows what the company intended to use. An AI asset inventory must also account for unmanaged tools, embedded features and systems built by individual teams.
3) Tracking applications but not their model dependencies
When an endpoint retires or a model changes, the application list alone cannot identify affected workloads. Track model dependencies and retain version history where available.
4) Ignoring AI agents and their permissions
An agent’s risk depends on the tools, data and actions it can reach. Record its identity, connected services, approval gates and ability to send data out.
5) Treating shared credentials as ownership
An API key proves someone used a credential. It doesn’t tell you which person or workload did. Use separate identities wherever you can, and keep explicit mappings.
6) Separating cost inventory from security inventory
If finance and security keep separate lists, each team misses half the picture. Connect costs, owners, data access and permissions through shared asset IDs.
7) Banning AI instead of offering a better option
Blocking all AI pushes people toward personal AI apps. Offer approved tools, clear data rules and an easy intake process, then match restrictions to the risk.
8) Designing an inventory that nobody can maintain
A massive form with dozens of required fields stalls adoption. Start with the fields you need for identity, ownership, dependencies, cost and risk. Add detail only for AI assets that need a closer look.
How Flexera helps you build and manage an AI asset inventory
Flexera One Token Cost Control covers usage-based AI that vendors bill you directly, such as model APIs, AI apps and coding tools. It pulls cost and usage data from each connected vendor’s API and turns tokens, credits and workflow runs into one view of spend. Then it ties each cost to the user, agent, app or workflow behind it.
Not sure where your team stands? Take our AI Cost Management maturity assessment.
Pramit Marattha
Pramit Marattha is a technical content writer and strategist with 5+ years of experience covering AI, data engineering, data cloud platforms and open source technologies. He turns complex technical concepts into clear, useful content for engineers and developers.