A procurement agent goes live with read access to supplier data. Three months later, it can issue purchase orders. Nobody decided that. The integration team asked for write access to fix a sync problem. A finance analyst asked for PO creation to clear a backlog. Each request was reasonable, and someone approved it. Add them up and a piece of software is spending company money, and no executive ever signed off on that.

Ask your leadership team who authorized it and watch what happens.

Agents hold business authority

Most executive teams still treat AI agents as software. They budget for them like software, hand them to a platform owner, and expect security to keep them in bounds. That worked when AI answered questions. Agents now hold credentials, read enterprise data, call tools, and take actions that move money, change records, and reach customers. Every permission an agent holds is something the company has agreed can happen in its name.

Finance already has a name for that. It’s delegated authority, and you already write it down in a delegation of authority matrix. The matrix says who can sign a contract, approve a purchase, or issue a refund, and up to what dollar amount. Your CFO owns it, your board approves it, and your auditors test it. An agent that approves refunds or issues purchase orders is exercising exactly that authority. Almost none of them appear on the matrix.

Ask who owns an agent’s authority, and you get four partial answers. The CIO owns identity. The CTO owns the integrations. The chief AI officer sponsors the use case. The CISO answers for it when something goes wrong. Each answer is accurate, and the authority itself still has no owner.

The numbers say nobody has decided

In a Cloud Security Alliance and Oasis Security survey of 383 IT and security professionals, 78% said their organizations have no documented, formally adopted policy for creating or removing AI identities. In Gravitee’s 2026 State of AI Agent Security survey, about 14% said all of their agents went to production with full security and IT approval.

The cost of getting it wrong is real. IBM’s 2026 Cost of a Data Breach study found that one in four malicious breaches was AI-enabled, at an average cost of about $6 million. Among the AI-related breaches, 27% came in through compromised APIs, applications, and plug-ins, the same connection points agents run on.

Put agents on the matrix

Treat agent permissions the way you already treat signing authority for people. For every agent that can act, answer four questions.

1. Who authorizes its scope? The executive who owns the business process the agent works in. The CFO for a payables agent, the head of customer service for a refunds agent.
2. Which actions need a human to sign off? Set dollar and risk thresholds the same way the matrix does for employees.
3. Who watches what it does against those limits? The security team, with reporting the business owner reads.
4. Who can shut it off? One named person with standing authority to suspend or revoke access without calling a meeting. In most companies, that should be the CISO.

The CIO’s job is to make those answers enforceable through identity, scoped access, logging, and a revocation path that works when someone pulls it. Standards bodies are starting on the mechanics. NIST launched a project in February 2026 on identity and authorization for AI agents. None of that work will tell you which of your executives owns the decision.

Won’t this slow us down?

The fair objection is that approval rules will choke the value agents promise, and that it’s smarter to move fast and let vendors and standards bodies catch up. There’s something to that. Overbuilt controls stall useful automation, and an agent that summarizes meeting notes doesn’t need a board resolution.

The matrix already handles this. Small decisions get pushed down, and big ones go up. Once everyone knows where the thresholds sit, low-risk agents ship faster because nobody has to ask permission. What slows a company down is the incident that shows nobody ever owned the authority.

Where to start

At your next leadership or board meeting, take the agents with the most access and ask who authorized what each one can do. If nobody in the room can name a person, add those agents to the delegation of authority matrix and assign owners this quarter.

Plenty of companies don’t have a technology leader who can run that work from the inside. ZRG Fortium’s Technology Leadership as a Service (TLaaS℠) puts an experienced interim or fractional CIO, CTO, or CISO in place within weeks. That leader gets the agents on the matrix, and the work sharpens the spec for your permanent hire when it’s time to run the search.

‍