Enterprises are giving AI agents more authority faster than they're building the systems needed to track them and control their actions.
That gap is becoming more visible as companies move beyond experimenting with agents and start letting them act across enterprise systems. The challenge isn't only determining what an agent should be allowed to do. Companies first need to know which agents are operating, what they can access and whether they can stop them when something goes wrong.
For some businesses, even that basic inventory remains difficult. Security leaders told Computer Weekly this week that companies are struggling to keep track of a growing number of AI agents. That becomes more difficult as those agents gain access to company data and applications and begin taking action rather than simply providing information.
New IBM research points to a similar gap. As agents move from answering questions to making decisions and acting, businesses are struggling to keep governance and accountability in step with their growing capabilities.
Related:Meta joins enterprise agent race with Muse Wonder, facing a trust gap
InformationWeek reported that identity is part of the answer to this pervasive problem, but it only addresses one part of it. An enterprise may know which agent is operating and what identity it's using. That doesn't necessarily tell the company whether the action the agent is attempting to take is appropriate.
That distinction matters as agents become more autonomous. Traditional identity and access systems were largely designed to determine whether a user or an application should be granted access to a system. Agents add another layer by making decisions and taking a series of actions after access is granted.
So, enterprises need more than a way to authenticate agents. They need visibility into where agents are operating, what they're doing and what authority they have. They also need controls that can limit or stop an agent when its behavior moves beyond what the company intended.
The potential security consequences add another concern. A Workato report found that 36% of leaders would try to handle an agent-driven security incident quietly rather than disclose it, suggesting companies may be confronting agent-related incidents before their oversight practices are fully established.
The next phase of enterprise agent adoption may depend less on how many agents companies can deploy and more on whether they can actually manage the ones they already have.
Giving an AI agent authority is relatively easy. Keeping track of that authority once the agent starts using it could prove much harder.
Related:Prompt: AI agents are running into the limits of access
Also in AI news this week:
Humanoid bottlenecks emerge amid US-China tensions: Rising production of humanoid robots is putting pressure on suppliers, as geopolitical tensions complicate Western manufacturers' access to alternatives to Chinese components.
House proposal would clarify liability rules for AI agents: A proposed House bill would hold developers liable when their AI agents harm third parties with conduct that would be unlawful if performed by a person.
Enterprises are growing the use of AI – but proving what it earns is the harder part: Companies including Santander, Axa and ABN Amro are grappling with how to demonstrate measurable business value as AI deployments expand.
Microsoft pushes custom AI chips with Azure Maia 300: The tech giant unveiled its third-generation Maia AI chip as it expands its custom silicon strategy and puts Maia 200 to work powering Copilot.
DeepSeek and what enterprises should learn from AI agent exploits: Another case involving an out-of-control AI model highlights why enterprises need explicit instructions and supervision as models and agents gain more autonomy.
Robots can handle a third of physical jobs - but not yet cost-effectively, Anthropic says: The generative AI vendor estimates robots could technically do these jobs, but economics remain a major barrier to widespread deployment.
Related:Google’s workplace agent is out, but questions abound
Reliability standards for AI data centers, crypto miners move closer to adoption: Proposed reliability rules for large computing loads of 50 megawatts or more are advancing toward possible approval by the North American Electric Reliability Corporation and federal regulators.
Liz Hughes is an award-winning editor and writer covering AI and emerging technology and the former editor of AI Business and IoT World Today.