An AI agent is not just another tool. It is a digital worker that gets permissions, reads data, operates systems and spends money, with no direct manager. To get value from agents safely, an organization needs to know three things at any moment: which agents exist, what they are allowed to do and what they cost. Without those answers, both security and FinOps are guesswork.
From chat to agent: what changed
Two years ago, the question in management meetings was whether employees should be allowed to use ChatGPT. Today the question is different. An employee can build an agent in Copilot Studio in ten minutes, connect it to SharePoint and the CRM, and share it with the whole department. A developer runs a coding agent that reads the entire repository and executes commands on its own.
Behind every such agent there is an identity, permissions, connections to business systems and a bill that grows every hour. The key difference: a chat answers a question. An agent acts.
Question one: which agents do we have?
In most organizations I meet, nobody holds a single list of the agents, custom GPTs, automations and coding tools actually in use. They are spread across Microsoft, OpenAI, Anthropic, GitHub and low-code platforms, each in a different console and owned by a different team. What you cannot see, you cannot secure and you cannot budget.
Question two: what are they allowed to do?
An agent built "just to summarize emails" often receives read access to every mailbox. An agent shared "with everyone" exposes its knowledge sources to anyone who talks to it. On top of that sits prompt injection: a document or email with hidden instructions can make an agent act against the organization, using permissions it received legitimately.
The principle is not new. Least privilege still applies, but now it has to be applied to identities that are not people.
Question three: what do they cost, and to whom?
AI spend is no longer a single subscription. There are seats, credits, tokens and pay-as-you-go charges beyond the included quota, and every vendor measures differently. A seat nobody has used for two months, or an agent stuck in a loop burning tokens all night, will not show up in any report until the invoice arrives.
Knowing which team each cost belongs to matters just as much. Without that, there is no real conversation about return on investment.
Why security and cost are the same problem
A sudden jump in one agent's token usage can be a bug. It can also be abuse. Either way, it is money. An agent nobody knows about is both a security risk and an open budget line.
That is why I do not believe in one team checking permissions and another team checking invoices, each with its own tool. Both should work from the same picture: one inventory of agents and users, with permissions, usage and cost for each.
A practical checklist for AI agents
- One list of every agent, GPT and AI tool in the organization, including who created it and who owns it.
- Every agent shared beyond its own team has had its knowledge sources reviewed.
- No agent holds broader permissions than its task requires.
- API keys and agent connections are kept in a vault, not inside code or settings.
- The monthly cost of every AI tool is known per team, not only per invoice.
- Inactive seats are reviewed and released on a regular schedule.
- Unusual usage or behavior by an agent raises an alert instead of surfacing at month end.
Common questions
What is the difference between AI security and agentic security?
AI security deals mainly with what goes into and comes out of the model: data leaking through prompts and inappropriate content. Agentic security deals with what the agent does: which systems it reaches, which actions it takes and with which permissions.
Should we just block AI tools?
No. Blocking only pushes usage to places you cannot see. It is better to allow usage with visibility, policy and clear ownership.
How do we start FinOps for AI when every vendor measures differently?
Start with an inventory: who uses which tool under which licensing model. Then normalize to a calendar month and assign every cost to a team. Optimization comes only after the picture is reliable.
