AI agents are operating across enterprise environments, making the governance of their actions an immediate concern. Organizations must control what agents can do and stop them when they operate outside predetermined parameters. Enterprise communications is no exception.
Customer conversations, for example, are a governance challenge as AI can generate dozens of records, transcripts, summaries and CRM updates that could be stored across different systems, each governed by different policies.
Yet only half of organizations have implemented a structured AI security and compliance program to protect data and applications across workplace collaboration and contact centers, according to research from Metrigy.
Platform fragmentation is one of the many governance challenges facing the unified communications (UC) space as agents go deeper into the enterprise, said Prachi Nema, principal analyst for UC and collaboration at Omdia, a division of Informa TechTarget. Enterprises might run two to seven UC and collaboration tools, each with its own controls and retention policies, she added.
"Responsibility is also fragmented across collaboration, contact center, security, legal, compliance, records management and AI governance teams," Nema said.
When records outlive the call
David Smith, founder and principal analyst at InFlow Analysis, identified another risk in how AI changes established workflows. Before AI automated contact center processes, each handoff provided a chance for someone to review the work. Automating that process eliminates the manual review.
"When AI takes that handoff out, the company books the time saved, but the checking quietly stops happening too," Smith said. "That's a liability nobody is pricing yet."
Smith suggested a practical test for CISOs: Choose one customer conversation and trace everything it produces, including the recording, transcript, summary and CRM update. Then, check which records the retention policy covers.
"Retention policies were written for recordings. A call can be deleted right on schedule while the transcript and summary AI made from it live on somewhere else under a different policy," Smith said.
AI-generated content, meeting artifacts, agent activity, data residency and retention are all areas that require management. Governance for AI-enabled meetings must also address inconsistent recording practices and privacy risks, such as the need for policies to define storage, access and deletion requirements for recordings, transcriptions, summaries and other derived records, according to Gartner.
Organizations must also account for mobile and BYOD use, establishing which applications can access business conversations and what controls apply.
Shadow AI creates another sizable compliance gap. About 19% of organizations respond to security risks by blocking applications or disabling features such as recording, transcription and AI copilots, according to research from Metrigy. This often results in employees turning to unsanctoned tools, such as AI notetakers, outside IT control. Security teams need visibility into which AI tools employees use, clear rules for processing customer information and a method for reviewing exceptions.
Voice is now sensitive data, too
AI-generated voices have made it harder to establish who is calling, said Ran Inbar, CTO for UC at communications software vendor AudioCodes. Customers can no longer trust caller ID and logo verification on their own, he said, prompting them to seek additional ways to verify caller identity.
Within the enterprise, speaker identification and automated notetaking can rely on voice samples to recognize who is talking, which raises questions about where data is stored, who owns it and who deletes it when an employee leaves.
"If an attacker needs only a couple of seconds of someone's voice to impersonate them, a stored voice sample is a security risk as well as a privacy one," Inbar said.
Organizations need to inventory where voice samples are collected, restrict access, and define retention and deletion responsibilities. They should also assess applicable privacy requirements, such as GDPR.
Stronger caller verification is also part of the response, Inbar said. A bank, for example, can initiate calls through its mobile app after the customer has authenticated, tying the call to an established identity-verification process.
Holding AI agents accountable
Just as for each employee, every agent should have its own identity and an audit trail, Smith said. Organizations must limit agent permissions, retain records of their actions and establish when human approval is required. Security teams also need processes to investigate or stop unauthorized activity.
Nearly half of organizations have suspected or confirmed an AI agent had taken unauthorized actions in the last 12 months, according to a survey of more than 380 security and IT professionals conducted by AI security company Salt Security. Yet only 17.5% of respondents said they could automatically stop an agent from behaving unexpectedly.
When evaluating governance tools, CISOs should test whether products can apply policies across communications platforms, integrate with existing security and compliance systems, track AI-generated records and provide auditable evidence of enforcement, Smith said.
Vendors should further demonstrate these capabilities in the organization's own workflows, using representative scenarios to test policy enforcement, record retrieval, access controls and agent intervention. Governance maturity depends on whether these controls work consistently across the environment.
Organizations can assess their current position by checking whether these policies cover every communications platform, AI-generated record and agent action. They can then identify gaps in coverage and enforcement and prioritize closing them.
CISOs should work with communications, IT, legal, compliance and records teams to assign responsibility for policy decisions, technical enforcement and exceptions. Business and compliance owners should approve policy requirements, IT and security teams should implement and monitor controls, and designated owners should review exceptions.
Audit pass rates, policy exceptions, coverage of required records and mean time to remediate failures can help show whether controls work in practice and where exposure remains.
The aim is to govern the original conversation, the data derived from it and the agents that act on it across multiple platforms and teams, with controls to prevent agents from acting outside their authorized scope and mechanisms to intervene and stop them when they do.
Moshe Beauford is a writer with more than a decade of experience covering enterprise technology, including AI, unified communications and customer experience.
