77% of Companies Have Already Paid a Price for an Ungoverned AI Agent

Most companies have now deployed an AI agent somewhere. A new survey suggests most of them have also already paid for one that nobody was properly watching.

On October 22, 2026, Redwood Software released its Enterprise benchmarks in AI and agentic orchestration 2026 report, based on a survey of 300 senior technology, risk and AI leaders collected by the independent research firm Leger Opinion. The headline: 77% have experienced a negative consequence from an ungoverned AI agent. The most common consequence wasn't a data breach or a rogue email. It was an unexpected cloud bill, followed by operational disruption and compliance issues.

A week earlier, IBM published a separate study of more than 3,200 senior executives across eight countries in Europe and the Middle East with a similar message. Agents are moving into production faster than the controls around them.

Redwood sells orchestration software, so it has an interest in this conclusion. But the two surveys, run by different firms on different populations, point in the same direction.

Deployment Is Outrunning Ownership

Redwood's numbers show how fast agents are spreading:

• 89% increased agentic AI deployment in the past year
• 54% describe their agentic adoption as scaling or mature
• 54% say they are fully prepared to orchestrate AI across existing systems

But the basics of oversight lag behind:

• Only 42% have a centralized registry of the agents they've deployed
• Only 51% have real-time monitoring of agent activity
• No single function owns agentic AI. Responsibility is split among IT operations, dedicated AI or automation teams and software engineering

Put simply, about half of the companies scaling agents can't produce a complete list of what agents they're running or watch what those agents are doing right now.

Incidents Are Already Reaching the Boardroom

IBM's study, published October 14, 2026, found that 80% of surveyed executives say an AI-related incident (an agent failure, data exposure, AI-driven decision error or security breach) reached their board or C-suite in the past year. Only 40% of those organizations changed their AI governance policy or framework in response.

The accountability picture is even fuzzier. Nearly all respondents (98%) have assigned AI risk to someone at board level. Yet when asked who is responsible for adverse outcomes from AI agents specifically, the most common answer, the CIO or CTO, was chosen by only 17% of respondents. Only 37% have a formal, board-reviewed estimate of the financial risk created by gaps in their AI governance, and only 37% say their governance is designed to scale with agentic AI.

Meanwhile, 35% are running agentic AI in production today, and three-quarters expect agents to have some decision-making autonomy by 2027.

Why the Bill Shows Up First

It makes sense that surprise cloud costs top Redwood's list. An agent that loops on a task, retries a failing step, calls a large model when a small one would do, or keeps running after its job is finished doesn't announce itself. It just consumes compute. Without a registry and live monitoring, the first sign is often the invoice.

Operational disruption comes next for a similar reason. Redwood found that 82% of respondents say existing systems are a bigger obstacle to scaling AI than a lack of strategy, and only 42% report a strong ability to connect agents to legacy or core systems. When an agent touches an older system that wasn't designed for it, small mistakes can ripple into real outages or bad records.

We've written before about why companies can't see what their AI agents are actually doing. These new surveys add the next chapter: the lack of visibility now has a measurable cost.

What Governed Agents Look Like in Practice

The respondents themselves are fairly clear about what would help. In Redwood's survey, 62% pointed to better observability or real-time visibility, and 58% cited proven integration with legacy and core systems. In IBM's survey, executives said they expect to need vendor liability terms for agent failures (33%), defined human escalation thresholds (32%), real-time monitoring and audit trails (30%) and the ability to immediately halt agent activity across the company (29%).

Translated into a practical checklist, every agent in production should have:

1. An entry in a registry. What it does, which systems it touches, who asked for it and when it was deployed.
2. A named owner. One person accountable for its behavior, not a shared team inbox.
3. Scoped access. Only the apps, data and credentials it needs for its job.
4. A spending limit. Budgets and alerts on model and compute usage, so a runaway loop gets caught in hours, not at month-end.
5. Clear approval thresholds. Actions above a defined risk level (payments, customer messages, record deletions) wait for a person.
6. A log of every action. So that when something goes wrong, the team can see what happened and why.
7. An off switch. The ability to pause one agent, or all of them, immediately.

None of this is exotic. It's the same discipline companies already apply to employees, contractors and software. The surveys suggest many organizations simply haven't extended it to agents yet.

Don't Forget Your Vendors' Agents

IBM's study flags one more gap that's easy to miss: 54% of respondents acknowledge gaps in how their governance covers AI systems used by key vendors and partners, and only 21% continuously monitor third-party AI risk. An agent running inside a vendor's product can touch your data just as easily as one you built. Ask vendors the same registry, access, logging and kill-switch questions you ask your own teams.

What This Means Going Forward

The good news in this data is that the problem is fixable. The companies paying the "ungoverned agent tax" mostly aren't suffering from bad strategy. They're missing basic operating controls: an inventory, an owner, a budget and a log.

As agents take on more autonomous work in 2027, those controls stop being nice to have. The companies that build them in now will be able to give agents more responsibility with confidence. The ones that don't will keep finding out about their agents from the invoice.

Frequently Asked Questions

What does "ungoverned AI agent" mean? An agent running without consistent oversight, such as no central record of it, no clear owner, no real-time monitoring or no enforced limits on what it can access or spend.

Who ran these surveys? Redwood Software's report surveyed 300 senior technology, risk and AI leaders via the research firm Leger Opinion. IBM's study surveyed more than 3,200 senior executives across eight countries in Europe and the Middle East.

What was the most common negative consequence? In Redwood's survey, unexpected cloud costs, followed by operational disruption and compliance issues.

Where should a company start? With an inventory. You can't set owners, budgets or approval rules for agents you don't know exist.