Enterprises can see their agents but cannot govern them, and vendors are selling the fix

The gap in numbers

By 2027, 74 percent of business leaders expect their businesses to use AI agents at least moderately, according to Talkdesk. Only 21 percent of organizations report having mature governance models for AI agents. That gap between adoption and control is the story of this week in enterprise AI: the agents arrived first and the governance is arriving second, sold as a product by the same vendors who sold the agents. m:…1478

The numbers behind the gap are uncomfortable. Talkdesk reports that 82 percent of organizations discovered shadow AI agents in the past year, and 65 percent reported at least one AI agent-related incident in the past year. Yet 68 percent of organizations report high confidence in their visibility into AI agents. Read together, those figures describe a familiar enterprise failure mode: confidence in what is visible, blindness to what is not, and incidents as the way the blind spot announces itself. m:…1478

Talkdesk traces the problem to how agents enter a company. AI agents enter enterprises through business applications, partners, internal teams and shadow AI, and vendor-specific dashboards lack end-to-end visibility for multi-vendor AI agent workflows. Its prescription is that governance must address identity, data access, tool use, delegation and business authority, and that governance needs to exist as a separate layer in the technology stack. Talkdesk has launched a public API to feed TRiSM systems with AI agent telemetry data, which is its version of that separate layer. m:…1478

Charters the agent cannot edit

ValidMind takes a different route to the same destination. Its Agent Authority enforces enterprise compliance charters on AI agents, and AI agents cannot edit or skip the Enterprise Charter. The company reports that regulators in Brussels, Washington and Hong Kong are issuing guidance for AI agents, and that the guidance includes a named owner for every agent, controls tied to agent purpose, human approval for high-impact actions, a way to stop the agent and a record of every action. m:…3051

The charter idea only matters if it bites on real decisions, and ValidMind gives three concrete examples. Agent Authority prevents agents from sending customer data to outside AI tools to comply with GDPR, requires human credit officer approval for loan declines to comply with fair lending regulations, and requires an approver and record for critical system changes to comply with DORA. The pattern is consistent: the agent proposes, the charter disposes, and the human approves the decisions that carry legal weight. m:…3051

Governance you can host yourself

CopilotKit's OpenBot for Business is a self-hostable workspace designed for enterprise AI agent governance. It runs on user-controlled infrastructure including AWS, Azure, Google Cloud or Kubernetes, supports single sign-on through identity providers like Google, Microsoft and Okta, and supports OpenAI, Claude, Gemini, Grok and open models like DeepSeek and Mistral. Agents built with LangChain, Google ADK or Agent Framework can join OpenBot as AI coworkers, according to the company. m:…6354

Two details stand out. First, administrators can define agent boundaries in plain English, which OpenBot converts into Google's Common Expression Language, so the policy is written by a person and enforced by a machine. Second, OpenBot for Business records every agent action and streams the logs to a SIEM system. It costs up to 29 dollars per active user per month on the pay-as-you-go plan, with a 5 percent charge on monthly model spend above 5,000 dollars per organization. The pricing matters because it tells you who the buyer is: this is infrastructure, not a feature. m:…6354

The hardest case: the law firm

Harvey shows what governance looks like when the stakes are highest. The legal AI platform enforces ethical walls on conversations, tables and drafts created from documents within its platform, and separates matter identity sync from access policy sync to allow independent scheduling. It uses Temporal as a durable workflow engine for bulk sync operations, and its largest workspace contains approximately 1.7 million matters. m:…8405

The engineering choices are all about failure modes. Harvey uses a dedicated search index built on turbopuffer to power the matter picker, and the search index contains searchable text but does not store wall data. Harvey encrypts customer data at rest using AES-256 and is contractually prohibited from training AI models on customer data. And if a wall fetch fails for a matter, Harvey retains the last stored policy. That last rule is the one to remember: when the control system cannot reach the policy, it fails closed. m:…8405

The retirement problem is where all of this converges. Only 21 percent of organizations have formal processes for decommissioning AI agents, and only 19 percent are highly confident they fully retire AI agents, according to Talkdesk. An agent that cannot be retired is an agent that keeps its access forever. The governance products launching this week all answer the same question, which is who owns the agent, what it may touch, and how it dies. Until a firm can answer all three, its dashboard is a map of a problem, not a control for it. m:…1478