Finding the AI Apps Your Teams Already Built
How to build an AI inventory of the apps, chat assistants and agents your teams already built: six places to look, what to record, and which ones to test first.
You can’t answer ten questions about an app nobody has told you exists. Our AI risk assessment template starts with “who owns it?” and that assumes someone knows “it” is there. In the four client cases behind The AI Risk We Keep Finding Is the App You Built, none of the apps was on an inventory when it became critical. This post is the step before the template: how to find what your teams already built.
What is an AI inventory, and how do you build one?
An AI inventory is a list of every AI system your organization runs or depends on, each with an owner, a purpose, the data it touches, its connections and the date it was last reviewed. Most organizations start with the AI tools employees use. The harder half is the apps employees built: portals, chat assistants, automations and agents that nobody bought, so procurement never saw them. You find those by following the traces building leaves behind. Check the money (AI and hosting subscriptions), the identities (OAuth grants, API keys, service accounts), the names (DNS and certificate logs), the traffic (servers calling model APIs) and the code (model SDKs and agent configs). Then ask people directly. Record each find, rank by data and exposure, and test the riskiest first.
Tools your people use, apps your people built
These are two different inventory problems, and they need different methods.
The first is the AI tools employees use: ChatGPT in a browser tab, a writing assistant, an AI feature switched on inside a SaaS product you already pay for. We covered that in AI Went From Chatbots to Agents. Your Security Didn’t. The risk there is mostly data going out.
The second is software employees built with AI that other people now depend on. That includes:
- Web apps and portals built in Lovable, Replit, Cursor or a chat window, and hosted somewhere a card could pay for.
- Chat assistants wired into a website or portal through a model API, often by following a tutorial.
- No-code automations in tools like Zapier, Make or Power Automate that pass business data through a model.
- Custom assistants (custom GPTs, Copilot agents, Gemini Gems) connected to shared drives, mailboxes or internal systems.
- Agents that act with someone’s credentials: reading mail, filing tickets, reviewing and merging code.
- MCP servers and plugins someone installed so an AI tool could reach a database, a repository or a SaaS account.
The second group is the one that ends up holding customer data and running real operations. It’s also the one a SaaS discovery tool is least likely to see, because a lot of it runs outside any SaaS catalog: a small app on a hosting platform, calling a model API with a key someone pasted into a settings page.
Six places building leaves traces
Nobody builds an app without paying for something, signing in to something, naming something or sending traffic somewhere. Each sweep below catches a different slice. None of them catches everything, which is why you run all six.
1. Follow the money
Ask finance for twelve months of expense reports and corporate card transactions, and search for three kinds of vendor:
- Model providers: OpenAI, Anthropic, Google AI, Mistral, Azure OpenAI and AWS Bedrock usage on cloud bills.
- AI app builders and coding tools: Lovable, Replit, Bolt, v0, Cursor, Windsurf and similar.
- Hosting and backends: Vercel, Netlify, Render, Railway, Supabase, Firebase, plus any domain registrar.
A recurring $20 charge for a coding tool tells you someone is building. A recurring hosting charge tells you something is running. Usage-based model API charges that grow month over month usually mean an app with real users behind it.
2. Follow the identities
Every app that reads company data has to sign in to it somehow. Your identity provider records most of that.
- OAuth grants. In Microsoft Entra ID, the enterprise applications and consent records show which apps users have authorized and with what permissions. Google Workspace has the same view under app access control. Look for apps with mail, file or calendar scopes that you don’t recognize, especially ones a single user consented to.
- Model API keys and AI resources in your cloud consoles. An Azure OpenAI resource, a Bedrock model enabled in an AWS account or a Vertex AI project in Google Cloud each mean someone is building.
- Service accounts and tokens created in the last year, including GitHub personal access tokens and GitHub App installations in your organization. Agents and automations need credentials, and builders usually make their own.
3. Follow the names
Apps need addresses. Pull your DNS zone files and look for subdomains nobody can explain, and search certificate transparency logs, the public, append-only record of TLS certificates as they’re issued, for every certificate issued to your domains. A search site such as crt.sh makes that a one-line query. A host named portal-test or ai-intake that appeared last spring is worth a question. Also search the common hosting platforms’ default domains for your company or product names, since a lot of AI-built apps never get a custom domain at all.
4. Follow the traffic
A laptop calling a model API is a person using AI. A server calling a model API is an app. Search your firewall, proxy and DNS logs for model endpoints such as api.openai.com, api.anthropic.com, *.openai.azure.com and generativelanguage.googleapis.com, then split the results by source. Anything coming from a server, a container or a cloud workload is an app until proven otherwise. If you run a secure web gateway or CASB, its catalog gives you a head start. Microsoft Defender for Cloud Apps, for example, has a Generative AI category in its cloud app catalog.
5. Follow the code
Search your repositories, and any you can reach on GitHub or GitLab under your organization, for:
- model SDK imports and calls (
openai,anthropic,@google/genai,langchain); - API key patterns, which your secret scanner may already be flagging;
- agent and MCP configuration files, such as
mcp.json,.cursor/directories and agent instruction files. Anmcp.jsonlists the servers an AI tool connects to, which makes it a map of that tool’s reach.
Code search only finds code that made it into a repository. When the answer to “where does the code live?” turns out to be “in the AI tool’s project,” record that too. It’s an answer to question 2 in the template, and not a good one.
6. Ask people
The sweeps find traces. People tell you what the traces are. Send every department head one question:
What did your team build, or connect to AI, this year, and what does it touch?
Two things make the answers honest. First, say plainly that the goal is to support what people built, not to shut it down. Teams that expect a ban hide things. Second, ask about connections as well as apps. “We connected our CRM to an assistant” is an answer people give freely, and it’s often the riskier one.
Compare the answers with what the sweeps found. Anything a team mentioned that no sweep caught tells you where your visibility has a gap. Anything a sweep found that no team mentioned needs a follow-up conversation.
What to record for each one
An inventory entry needs enough detail to decide what to do next. For each app, record:
| Field | What to capture |
|---|---|
| Name and purpose | What it does, in one sentence |
| Owner | A named person, plus a backup |
| Users | Staff only, customers, or the public |
| Data | What it reads and stores, and its classification |
| Connections | Systems, accounts, API keys and tokens it uses |
| Hosting | Where it runs, and who has admin access there |
| How it was built | Tool or platform, and where the code lives |
| Logging | Whether its activity reaches your central logs |
| Last review | Date of the last security review or test |
These fields are deliberately the first questions in the AI risk assessment template: ownership, how it was built, logging, access, data and connections. The inventory tells you an app exists and gives you a rough read. The template, filled out by its owner, gives you the full one.
Expect blanks. A blank owner is the most common, and it’s a finding in its own right. Assign one before you do anything else.
Which ones to look at first
You’ll find more apps than you can review at once. Rank them by three things:
- Who can reach it. Anything on the public internet, or open to customers, comes first. The plumbing company’s chatbot in our cornerstone post was a public chat form in front of an agent with access to every customer account.
- What data it holds. Regulated or customer data (financial records, health information, personal data) moves an app up. The financial services portal in the same post held client account details with no central logging.
- What it can do. Read access is one thing. An agent or automation that can write, send, merge, pay or delete moves up the list, and one that does it with a person’s credentials moves higher still. See what to ask before you hand an AI agent the keys.
An internet-facing app holding customer data with write access to internal systems goes to the top and gets tested now. An internal read-only summarizer for public documents can wait for the next review cycle.
Keep it current
An inventory starts going stale the day you finish it. Teams keep building, and the next app won’t announce itself either. Three habits keep it current:
- Register at the point of building. Put a short intake step wherever building starts: requesting a model API key, a hosting account or an AI platform license. Ask for the owner, the purpose and the data, and the app is in the inventory before it holds anything sensitive.
- Rerun the sweeps on a schedule. Quarterly works for most organizations. Expense data, OAuth grants and certificate logs are cheap to re-pull once you’ve done it the first time.
- Rerun them after events. A new AI platform rollout, an acquisition, or a framework vulnerability that sends you hunting for every affected app. When React2Shell and PHP2Shell dropped, the teams that patched fastest already knew where their apps were.
This is also what the frameworks ask for. The NIST AI Risk Management Framework (version 1.0, which NIST is now revising) says it plainly in GOVERN 1.6: “Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities.” Federal agencies have to keep an annual AI use case inventory under OMB memo M-25-21. You don’t need to adopt either framework to adopt the habit.
Start with one sweep this week
You don’t need a platform to start. Pull last year’s card transactions and search for the vendors above. Pull your OAuth grants. Send the one-question email. Most organizations find their first unknown app inside a week, and the second and third come quickly once people see that reporting one doesn’t get it shut down.
If you’d rather have it done across the organization, discovery is the first phase of our AI Security Assessment: we run these sweeps, interview teams, put an owner on everything we find and rank it against the NIST AI RMF and OWASP AI guidance. Then the apps that matter get tested, through an AI Application Security Review or AI agent security testing. For the rest of the series, start with The AI Risk We Keep Finding Is the App You Built.