AI agents need a management layer before they need more autonomy. Fortune reported on September 16 that Cisco, Intuit, Workday, and ServiceNow are all building ways to monitor, govern, and contain agents as they take on more work inside companies. Buying agents brings management decisions: who manages them, what they are allowed to do, and how the business proves what happened after they act.
The reported systems offer examples of how to record agent activity and enforce permissions. Their adoption figures alone do not establish business outcomes.
Capability needs operating controls
Alongside tests of coding, search, and other capabilities, test the controls around an agent’s access and actions.
Fortune describes several approaches. Cisco launched an internal platform called MyAgent in August and gave roughly 90,000 employees access. The company says it reached 50% daily adoption within two weeks. Employees can create their own agents, but Cisco requires approval from a centralized team before those agents become authorized. Around 700 have been approved.
Cisco’s daily adoption figure measures usage; it does not establish productivity gains.
Cisco’s executive vice president of operations, Thimaya Subaiya, told Fortune the company wants to avoid “agent sprawl” and will not authorize third-party agents across random pockets of the organization. Whether you agree with that level of centralization or not, the management principle is right. Someone has to know which agents exist, what they touch, and whether they are still doing the job they were created to do.
Agents turn governance into daily operations
Acceptable-use and security policies need an enforcement path in the tools agents use.
An activity record should show information requested, tools called, and records changed. For example, if an agent sends a sales manager a recommendation based on customer data at 2:14 p.m., a reviewer should be able to trace the source access and the message.
Fortune reported that when Intuit designed its generative AI operating system, GenOS, the company also drew a security, risk, and fraud layer into the architecture from the beginning. Every AI request is tracked and every response is recorded. Intuit CTO Alex Balazs put it plainly: “You don’t want to try to retrofit the ability to enforce security and responsible AI foundations after the fact.”
Test the record layer early. If activity is already distributed across departments and vendor accounts, the inventory will need to reconcile those records and identify gaps.
Record the agent’s identity and activity
Workday calls its approach an “agent system of record.”
Agents are becoming non-human identities performing work that the business has to account for. If they have access, tasks, permissions, outputs, and consequences, they need records.
Record the agents that exist, their owners, permitted systems and tasks, and actions awaiting approval. The activity record also needs a way to surface a run that failed without producing an alert.
If your organization lacks that record, it is not ready to scale agents. Resolve gaps in the record before expanding access.
PwC’s Joe Atkinson told Fortune that “the agent made me do it” will not be a moral or legal defense. Treat that statement as an accountability warning; specific legal obligations need their own review.
Begin with an inventory
List every AI agent, assistant, automation, and workflow currently touching business work. Include unofficial tools as well as approved deployments. For each one, capture the owner, use case, systems touched, data touched, approval requirements, failure mode, and whether the output is reviewed before it affects a customer, employee, vendor, or financial record.
Fortune also cited new Collibra research showing that seven in ten data management, privacy, and AI decision-makers say poor or unaligned data foundations are the root cause when AI pilots hit roadblocks. In large organizations with $100 million or more in revenue, Fortune reported that 96% tie AI project failure to poor data foundations, and 64% say they need manual review before autonomous agent outputs go live.
The immediate work is to build better data foundations, clear ownership, review paths, records, and permissions that make agent activity visible.
Use the inventory to select a workflow for a control test. Check whether its owner can inspect a recent run, identify the approval boundary, and stop or recover the work when it goes wrong.