Webinar recap: Practical agentic enterprise governance advice from AI-forward companies

Key takeaways from the agentic enterprise governance webinar

• The biggest practical surprise is not the raw number of agents but how many have no clear owner. Orphaned AI agents, often sitting on long-lived service accounts or OAuth grants, create accountability gaps that many organizations treat as a reason to pause.
• Agents arrive mainly as features inside platforms companies already pay for or as rapid builds by business users. Effective agent inventory management is harder than traditional discovery and is the foundation for agentic enterprise governance.
• Two identity models are emerging. One ties agents to human owners so joiner/mover/leaver processes still work. The other gives agents persistent identities of their own, which matters for agent-to-agent governance.
• Traditional domains (SaaS, IAM, security, FinOps) each see only a slice of agent activity. A single control-plane view with named ownership is still missing.
• Speed is winning. Teams route around slow processes. Progress comes from making the governed path easier and, where possible, delivering value back to users.
• A workable sequence is See → Decide → Act → Prove. Jumping to policy or AI agent FinOps before visibility and ownership usually fails.
• Next 90 days: complete agent inventory management and put a named owner on every agent that can touch company data.

Agentic enterprise governance webinar recap

On November 10, 2026, BetterCloud hosted a panel discussion titled “Beyond AI Adoption: Governing the Agentic Enterprise.” Moderated by Hayley Mathis, the panel featured:

• Prashant Kulkarni, VP and AI Architect, Security Partner Co-Innovation at Google Cloud
• Mike Lee, VP of Data + AI at Taylor Farms
• Jack Murphy, Head of Information Technology and Security at Flipdish
• Raj Kunnath, President at BetterCloud

Together, they discussed how organizations are moving from static AI tools to autonomous agents, addressing critical governance challenges such as:

• Orphaned AI agents
• Agent inventory management
• Identity models and agent-to-agent governance
• AI agent FinOps

From AI tools to agents

A year ago, most AI governance conversations centered on tool discovery and basic observability. Organizations focused on tracking employee tool usage, protecting sensitive documents, and enforcing compliance.

Simply put, the dominant pattern was request and response: a person asked a question and received an answer. Humans stayed firmly in the loop.

Our panelists agree that we’re quickly moving past this model. Agents now read company data, operate across multiple systems in a single workflow, and take actions using real credentials. The relevant questions have shifted to what:

• An agent can reach
• Decisions it makes and
• Actions it actually takes 

This change defines the practical scope of agentic enterprise governance.

Raj Kunnath described the difference in operational terms. Earlier AI tools required a human to initiate every interaction. Now, agents are given agency to act. They can run with varying degrees of autonomy inside guardrails that are still being defined, and their effective scope often expands as users find additional ways to apply them. He rightly characterized them as a new class of employee, as a “virtual employee.”

Prashant offered a complementary analogy that resonates as well: agents resemble eager virtual interns. They are highly motivated to complete the assigned task, yet they have limited judgment about what is safe, appropriate, or within intended bounds. 

So what’s the key implication for every organization? There’s a gradual movement from humans in the loop, to humans on the loop, and in some cases toward humans out of the loop for routine work. 

Managing this transition is the central challenge for agentic enterprise governance.

Orphaned AI agents and the ownership gap

Across BetterCloud customer assessments, the primary surprise isn’t the sheer number of agents, but how many lack a clear owner. Raj noted that roughly 30% of the discovered agent landscape is completely unmapped to any human supervisor.

These orphaned AI agents commonly operate through:

• API keys
• Service accounts
• OAuth grants

All of these were approved at an earlier point and have not been revisited.

The absence of an AI agent owner creates an accountability gap

Without a defined responsible person, issues arise when the AI agent drifts from its original purpose, acquires additional access, or needs to be shut down. A growing number of IT teams treat the lack of a named owner as a reason to pause the agent rather than as a documentation detail to address. This stance is becoming a basic control within agentic enterprise governance.

Ownership also matters for agent lifecycle events

When the person who originally created or sponsored an agent leaves the company or changes roles, this unowned agent can continue running indefinitely. Establishing early ownership makes it possible to apply existing onboarding, mover, and offboarding logic required with any other user lifecycle management.

Agent inventory management in practice

Our panelists emphasized that agent inventory management is significantly more complex than traditional SaaS or cloud discovery, as agents originate across multiple entry points and rapidly evolve. Mike Lee and Jack Murphy shared how this plays out in real-world environments.

Agents used across the enterprise 

At a large produce company, Taylor Farms, people both inside and outside IT build and use AI agents. How? Business stakeholders closest to operational problems use AI-assisted coding tools to create proof-of-concept solutions inside internal sandboxes. 

Once value is demonstrated, the work moves through an assessment process and ITSM before broader deployment is considered. 

To increase visibility, the company created an internal discovery tool, referred to as “pixel.” Using this tool, employees can check whether a similar agent already exists before submitting a new proposal. An internal AI academy provides a modest incentive: teams that complete training gain access to centrally managed resources.

Agency visibility remains incomplete

Many agents appear as features that are simply switched on inside platforms the company already pays for, such as productivity suites or business applications. 

Mike from Taylor Farms described the overall situation with an apt analogy: AI agents are like multiple trains leaving different stations at the same time. Realistically, IT’s role is to guide them toward a shared destination rather than to stop departures.

At Flipdish, an Irish company serving the restaurant sector, the pattern is similar. The organization has:

• Deployed browser extensions
• Built internal discovery tools
• Established an AI operations team that helps citizen developers use approved templates and processes

Despite these efforts, many agents only become visible when something fails or when a team requests help. 

One of the reasons why is that many AI capabilities simply appear as a new feature inside products already in use. Jack’s team at Flipdish responded by building some of its own Model Context Protocol (MCP) implementations and associated logging to improve the ability to attribute actions to a specific agent. 

Still, there are open questions around maintenance. Is an AI agent updated when underlying APIs or interfaces change? Who, if anyone, is accountable for these AI agents months after creation?

Effective agent inventory management is continuous

Both panelists emphasize that inventory management must cover custom-built agents as well as built-in platform features. It also needs to include the credentials that give each agent its actual reach. Without a solid foundation here, broader agentic enterprise governance efforts remain incomplete.

Identity models and agent-to-agent governance

The agentic enterprise governance panel outlined two approaches to identity, each with different consequences for agent-to-agent governance.

1. The human-owner model

Favored by BetterCloud’s Raj and Flipdish’s Jack, they prefer to associate each agent with a specific person. 

This approach allows organizations to reuse existing onboarding, role change, and offboarding processes. When an employee changes roles or leaves, the agents linked to that person can be reviewed or retired through familiar workflows. 

Attribution also improves, because logs that might otherwise show only a user account or a generic integration can be tied more clearly to the agent acting on that person’s behalf.

2. The independent-identity model

Prashant at Google Cloud pursues this approach. The independent-identity model gives agents first-class identities of their own. These identities are implemented with SPIFFE (Secure Production Identity Framework For Everyone) certificates—identity documents issued under this standard—and persist for the life of the agent. 

For those who may be unfamiliar with SPIFFE, it’s an open standard for giving workloads (applications, services, or now AI agents) a strong, cryptographic identity that is independent of the underlying infrastructure or network. A SPIFFE certificate is usually a short-lived X.509 certificate that proves “this workload is who it claims to be.”

This is what allows the “first-class identity for agents” approach that the panel discussed. It supports agent-to-agent communication, anomaly detection, and more precise accountability.

Anomaly detection then monitors the trajectory of the agent’s actions. If the path unexpectedly diverges—for example, shifting from an anticipated sequence to an unplanned branch—a kill switch can interrupt execution. Sandboxing provides an additional constraint: the agent submits a request, the sandbox performs the query or computation against the data, and only the result is returned. In this case, the agent never sees the raw data.

This model specifically focuses on agent-to-agent governance.

Even under this more autonomous approach, human accountability never disappears. 

The person or business unit that deploys the agent remains responsible for defining its purpose, conducting testing and red-teaming, and ensuring a kill switch is available. In this case, embedded preventive controls in the workflow mean you don’t solely rely on after-the-fact detection.

Agent Ownership Model	Human-Owner	Independent Identity
How it works	Agent is tied to a specific person	Agent has its own persistent identity
Lifecycle handling	Reuses lifecycle management processes	Identity persists for agent lifetime
Accountability	Clear human owner	Deploying person/team still accountable
Best for	Permission management & offboarding	Agent-to-agent / swarm-style work
Key controls	Ownership and access reviews	Agent gateway, SPIFFE, kill switch, sandbox
Key risk if neglected	Orphaned agents when people leave	Unclear accountability or over-permissioning

It’s not an either-or decision when it comes to agent identities. Organizations can use both: SPIFFE for strong agent identity, and BetterCloud for the ownership, inventory, and cross-system governance layer.

AI agent FinOps and current cost realities

The panel also tackled AI agent FinOps, noting that managing agent-related costs remains an immature practice for most organizations and is currently secondary to solving discovery and ownership.

Right now, token and model costs are still tracked primarily with homegrown tools. At Flipdish where the goal is to prevent rapidly increased consumption, internal systems allocate spend by person and by the type of project or group involved. 

The panel also brought up another common AI agent FinOps problem: Too many users, especially citizen developers, tend to select the newest or most expensive model by default. 

Both Raj and Prashant chimed in that many workflows do not require the highest-tier model for every step. A lighter and less expensive model is often sufficient for execution once a planning or reasoning stage has been completed.

Mike noted that even when known cloud and AI expenditures are visible inside central budgets and separate ledger accounts, untracked usage continues because:

• Controlling AI agents using service accounts can interrupt legitimate work
• Chargeback or per-agent cost attribution is still hard to achieve

As a result, meaningful AI agent FinOps depends on agent inventory management and ownership assignment. Without a reliable inventory and clear accountability, there’s no stable foundation to control costs.

A practical sequence for control

Raj proposed a straightforward sequence that aligned with the operational experience shared by the other panelists: See, Decide, Act, and Prove.

The sequence is a continuous exercise. After all, AI agents and agent-to-agent interactions continue regardless of quarterly review cycles. 

Jack from Flipdish emphasized an additional practical requirement: the process must remain fast enough that teams do not simply bypass it. Agents tend to quickly evolve after their initial deployment, so registration and updates need to keep pace without creating too much friction. Executing See/Decide/Act/Prove consistently turns isolated controls into functioning and effective agentic enterprise governance.

90-day priorities and longer-term direction

The near-term recommendation across the panel was consistent. Organizations should inventory every agent, automation, or AI integration capable of accessing company data and assign a named owner to each one. 

This establishes the necessary baseline for accountability and for any subsequent controls. Education, hands-on testing of models and platforms, and the introduction of basic guardrails can follow once visibility and ownership exist.

Looking further ahead, agent onboarding and offboarding within existing identity systems should become as routine as human user lifecycle management. 

Whether an organization chooses to keep agents tightly linked to human owners or adopts more independent identities with gateway and sandbox controls, a functioning control plane is still required. 

That control plane needs to answer:

• What a given AI agent is permitted to do
• Who is accountable for that AI agent
• How an AI agent can be stopped without waiting for a periodic review.

Data cited during the session highlighted projections that a substantial share of agentic initiatives could fail due to runaway costs, unclear business value, or weak risk controls—underscoring the urgency of establishing strong operational foundations early.

To underscore the importance, panelists cited recent analyst research from Gartner. They project that over 40% of agentic AI projects will be canceled by the end of 2027 due to cost, unclear value, or weak risk controls, even as 33% of enterprise applications are expected to include agentic AI by 2028. Some mentioned a 2025 McKinsey study that says 62% of organizations are experimenting with agents and 23% are scaling at least one. 

Your next step in agentic enterprise governance

The BetterCloud webinar panel had one clear agentic enterprise governance message: start with AI agent visibility and ownership before trying to deploy broader controls. 

For assessing your own environment, begin with a focused AI agent inventory. Identify every agent that can access company data and assign a named owner. This single step addresses the panelists’ most common gap and creates the foundation for everything else.

Ready to learn more? Watch the webinar recording: Beyond AI Adoption: Governing the Agentic Enterprise for the full discussion or get the BetterCloud AI Governance and SaaS management platform demo.