77% of firms hit by ungoverned AI agents, survey finds

The survey

A study by Polaris, a data observability firm, surveyed 250 IT decision-makers and found that 77% had experienced negative consequences from AI agents operating without proper governance ^[1]. The most common problem was unexpected cloud costs, reported by 68% of respondents. Security incidents came second at 47%, followed by breaches of regulatory or internal compliance at 42% and negative impacts on customer experience at 38% ^[1].

The survey was published in early 2025, as organisations race to deploy AI agents that can act on their own: booking meetings, changing orders, querying databases and calling external services. Unlike a chatbot that merely answers, an agent does things. That makes it harder to oversee, and the survey suggests many firms are finding out the hard way.

What an AI agent is

Think of it like the difference between a calculator and a personal assistant. A calculator waits for you to tell it what to do. An assistant reads your calendar, notices you're double-booked and reschedules one meeting without asking. AI agents are the assistants: they interpret instructions, decide which tools to use and take actions in your systems.

The line between a helpful agent and an ungoverned one is oversight: can you see what it did, when it did it and on whose instruction? If the answer is no, you're running blind.

Why the costs run away

The cloud bill problem is easy to understand. An agent calling a large language model hundreds of times an hour can rack up charges fast, especially if it's using an expensive model for tasks a cheaper one could handle. The survey found 68% of respondents had been stung by unexpected costs ^[1].

But the costs are a symptom. The underlying issue is that nobody's watching how often the agent calls which service, or whether it's doing so efficiently. That's the governance gap.

The compliance and security side

Nearly half the respondents reported security incidents, and 42% said agents had breached regulatory or internal rules ^[1]. The AI Act's high-risk requirements don't yet apply to most of these systems, but when they do—from 2 December 2027 for Annex III systems—Article 12 will require automatic logging of events over the system's lifetime ^[2].

The law says what to log. It doesn't say how you prove the log hasn't been quietly edited after an incident. That's where the governance gap meets the evidence question: when something goes wrong, can you show what the agent did, or are you just hoping the log is intact?

What this means for you

If you're deploying AI agents, or thinking about it, here are three questions worth asking now:

• Can you see every action an agent takes, in enough detail to reconstruct a sequence when something goes wrong?
• Do you know who instructed the agent, which tools it called and what the results were?
• If a regulator, auditor or customer questions an agent's action six months from now, do you have a record that holds up?

The survey's findings are a warning: moving fast is one thing, but moving blind is expensive.

How a digital witness fits

Here, the thing to seal is each action the agent takes: the instruction it received, the tool it called and the result. Sealing takes a fingerprint of the record, not the record itself, so the action log never leaves your systems.

Say an agent changes a customer's order and the customer later says nobody asked it to. With each action sealed as it happens, you can show the instruction, the tool call and the timestamp in a way anyone you share it with can check. Change a single character in the log and the fingerprint won't match.

A digital witness doesn't stop an agent doing something daft, and it doesn't replace monitoring or access controls. What it does is give you a record that holds up when questions come later, which the survey suggests they will.

You can seal agent actions through WitniumChain's MCP server, which works with Claude and other agent frameworks. The content stays with you; the fingerprint goes on a chain where any later change would show, even one made by us. To check, anyone calculates the fingerprint again and compares it with the sealed one. The first 2,000 witnesses a month are free. Read the docs to see how it works.

How it would work

1. Your agent framework sends a fingerprint of each instruction and result to the MCP server as actions happen.
2. The server seals the fingerprint and returns a witness ID.
3. When someone questions an action, you share the log and the witness ID, and they check that the fingerprint matches.

What to watch

The AI Act's logging requirement in Article 12 applies to high-risk systems from 2 December 2027 for Annex III and from 2 August 2028 for Annex I ^[2]. Whether your agent qualifies as high-risk depends on what it does, but the evidence question applies either way: when something's questioned, can you prove what happened?

The survey's 77% figure is a snapshot of early deployments ^[1]. As agents become more common, expect the oversight gap to close or the consequences to grow. Probably both.