Shadow AI agents are not shadow AI tools: how to build an AI agent inventory.
A person using an unapproved assistant is one problem. An agent acting alone, with nobody's name on it, is another. The inventory is the first control.
- Topic
- Shadow AI
- Product
- Novie Control
- Format
- Guide
- Reading time
- 4 min
Shadow AI agents are AI systems that take actions inside your business without anyone having signed them off. They differ from shadow AI tools, where a person uses an unapproved assistant, because the agent acts on its own and keeps acting after the person has gone home. An AI agent inventory is the register of every agent running, who owns it, what it can touch and what it costs. It is the first control that makes agents governable.
Shadow IT, shadow AI, shadow AI agents
Shadow IT is software used without IT approval: a file-sharing account, a messaging app, a personal cloud drive. Shadow AI is the use of AI tools without approval or oversight: a personal chatbot account, a browser extension, an AI feature switched on inside a licensed product. Shadow AI agents are the third step. An agent has a goal, tools it can call, and permission to act, and it was approved by the person who built it, on the day they built it.
- What it is. Shadow IT: unapproved software. Shadow AI tools: unapproved AI assistants used by people. Shadow AI agents: unapproved AI systems that act on their own.
- Who acts. Shadow IT and shadow AI tools: a person, each time. Shadow AI agents: the agent, repeatedly.
- Main exposure. Shadow IT: data stored outside control. Shadow AI tools: data processed outside control. Shadow AI agents: actions taken outside control.
- Approval. Shadow IT and shadow AI tools: usually none. Shadow AI agents: the builder, once, with no scope.
- What decays. Shadow IT and shadow AI tools: little. Shadow AI agents: the approval, as models and tools change.
The difference that matters is the last one. A tool does what a person asks. An agent does what it was set up to do, and when the model behind it is upgraded or its tool surface grows, it can do more than anyone approved without anyone changing it.
Why this is now a board question
In 2026, IBM's Cost of a Data Breach Report found that shadow AI was involved in 43 percent of breaches, up from 20 percent the year before, across 602 breached organisations. The average shadow AI incident cost $5.39 million, and more than two thirds of organisations had no governance process in place to limit shadow AI. Only 32 percent of breached organisations had AI policies at all, down from 37 percent.
Agents are arriving ahead of the controls. In its Q2 2026 Pulse Survey, KPMG found 53 percent of large US enterprises had AI agents deployed and 18 percent were orchestrating multiple agents across workflows, double the previous quarter. Deloitte's 2026 State of AI in the Enterprise found 74 percent expect to be using agents by 2027 and 21 percent have a mature governance model for them.
Novamind's view: shadow AI is an inventory problem before it is a policy problem. A policy tells people what they should not do. It has no way of noticing when an agent does it anyway.
How to build an AI agent inventory
An inventory is a register, kept current, with one row per agent. Eight fields do the work.
- Name and owner. The agent, and the person accountable for it. A team name is not an owner.
- Purpose. What business outcome it exists to produce, in one sentence.
- Model and tools. Which model it runs on, and which tools or APIs it can call.
- Systems touched. Every system it can read from or write to.
- Data reached. The classes of data it can see, including personal information.
- Approval. Who approved it, when, and the scope they approved. If the scope is blank, the approval is a signature on a demo.
- Cost. What it costs per run and per month, attributed to a team and project.
- Stop. How it is paused, and who can do it.
Where to find the agents you did not deploy
Nobody keeps a list of the agents nobody approved, so the inventory is assembled from the places they leave marks.
- Procurement and finance. Vendor contracts with AI features, card statements with model-provider charges, cloud bills with inference lines.
- Identity. Service accounts and API keys created in the last twelve months, especially ones with no human named on them.
- Vendors you already pay. Ask each which AI features they have switched on inside the product you license, and which of those can take actions.
- The teams. Ask engineering, operations, marketing and finance separately what runs on their behalf. The gaps between the answers are the finding.
Keeping it current
An inventory that is right on the day it is built decays at the rate agents change. Two habits keep it alive. First, every new agent enters through the same governed path, so its row is written when it is built rather than discovered later. Second, when an agent's model, tools or reach change, its approval comes up for renewal and its row is reviewed. The approval is a statement about a capability at a point in time; store it that way.
How Novie Control does it
Novie Control is live. Every agent carries a named owner, an approval trail and a full audit trail, so the inventory for governed agents is a live view rather than a spreadsheet. Every AI dollar is traced to the team, the project and the owner at the moment it is spent, and caps hold before spend is committed, with a named human on every increase.
The agents outside that path still need the manual work above. The point of the platform is to make the governed path the easy one, so the next agent lands in the inventory on the day it is built. Novie OS, launching soon, is where that next agent is described, approved and built.
Sources
- IBM, Cost of a Data Breach Report 2026 (29 July 2026), via Cybersecurity Dive
- IBM, Cost of a Data Breach Report 2026 (29 July 2026), via Kiteworks
- KPMG, AI Quarterly Pulse Survey Q2 2026 (24 June 2026)
- Deloitte, Business and IT leaders report AI agents are scaling faster than their guardrails (24 April 2026)