AI Agents in the Enterprise: Trust, Security and Adoption

The Enterprise Has Always Been a Driving Force Behind Technology Adoption

Some of us still remember a world without personal computers, the internet, or smartphones. Looking back, it is remarkable how quickly these technologies moved from being innovations accessible to relatively few people to becoming essential parts of our personal and professional lives.

One of the major drivers behind this transformation was enterprise adoption.

Personal computers became indispensable when organizations began integrating them into their daily operations. The internet evolved from a communication and research network into a fundamental infrastructure for global business. Smartphones transformed not only how people communicate but also how companies interact with employees, customers, and markets.

Of course, consumer adoption and technological innovation also played fundamental roles. But enterprise investment helped establish the infrastructure, business models, and economic sustainability required for these technologies to scale.

Today, I believe we are approaching another important turning point with artificial intelligence, particularly with the emergence of autonomous AI agents.

However, there is an important lesson from previous technological revolutions: the existence of an advanced technology does not automatically guarantee successful adoption.

What We Can Learn from Virtual and Augmented Reality

Having worked with immersive technologies, I have followed the evolution of virtual reality (VR) and augmented reality (AR) with particular interest.

Over the past decade, we witnessed enormous investment, technological improvements, and ambitious expectations surrounding immersive experiences.

The potential was, and remains, significant. From education and professional training to industrial simulation, healthcare, and remote collaboration, these technologies can provide experiences that conventional interfaces cannot easily reproduce.

Yet enterprise adoption has progressed more slowly than many expected.

In my view, the challenge was not simply technological maturity. In many cases, the business value was difficult to demonstrate, implementation costs remained significant, and organizations were not sufficiently prepared to integrate these solutions into existing workflows.

New hardware, specialized environments, user training, and changes in working practices created additional barriers.

This does not mean immersive technologies have failed. They continue to offer substantial value in specific applications.

But their evolution illustrates an important principle: technological potential and sustainable enterprise adoption are two very different things.

And I believe we should learn from the immersive tecnologies example when thinking about enterprise AI agents.

The AI Agents Paradox: Easier to Adopt, Harder to Govern

There is, however, a fundamental difference between immersive technologies and AI agents.

Unlike VR and AR, AI agents do not necessarily require new hardware, specialized devices, or major changes to existing user interfaces.

They can be integrated directly into the software environments that organizations already use: ERP platforms, CRM systems, financial applications, human resources systems, databases, and enterprise workflows.

This gives AI agents a potentially much faster path to enterprise adoption.

But it also introduces a new category of operational risk.

Traditional AI assistants primarily generate information, recommendations, or content. Autonomous agents can go further.

They can execute transactions, modify business records, communicate with external systems, initiate workflows, and coordinate actions with other agents.

We are moving from systems that assist human decisions to systems that can execute decisions on behalf of an organization.

AI agents may be easier to adopt than previous technologies, but they are potentially much harder to govern.

A mistake in a conventional information system may produce an incorrect recommendation. An autonomous agent with execution privileges could act on that recommendation before anyone notices the problem.

The consequences may include financial losses, operational disruptions, unauthorized access to information, or violations of internal policies.

This is why enterprise adoption cannot be driven by technological enthusiasm alone.

It must be supported by a clear understanding of business risks, responsibilities, and operational boundaries.

The Gap Between Experimentation and Enterprise Scale

The AI Agents 101 Atlas, presented by Dinis Guarda with contributions from Jack Widman and Phil Grady of Sheaf, highlights an important contradiction in the current market.

Drawing on McKinsey's State of AI 2025, the Atlas reports that 62% of surveyed organizations are experimenting with AI agents, while only 23% are scaling agentic systems somewhere in their operations.

This gap is significant.

It suggests that organizations are increasingly interested in the technology, but many have yet to establish the conditions required for reliable implementation at scale.

The Atlas also references Gartner's prediction that more than 40% of agentic AI projects will be cancelled by the end of 2027, with factors including rising costs, unclear business value, and inadequate risk controls.

Although this is a forecast rather than an observed outcome, it highlights a challenge that technology leaders should take seriously.

From my perspective, successful enterprise adoption should begin with a specific business problem rather than with the technology itself.

Organizations should identify processes that combine relatively low operational risk with clear, measurable business value.

This allows technical teams to demonstrate results, understand limitations, and gradually build management confidence.

At the same time, more complex applications can be explored in parallel through sandboxes and controlled environments.

A controlled implementation is not necessarily a slow implementation.

In fact, an uncontrolled deployment that results in a serious incident can delay adoption much more than a carefully planned implementation.

Trust takes time to build, but it can be lost through a single autonomous action.

Progressive Autonomy: Building Trust Through Experience

I do not believe organizations should approach AI autonomy as an all-or-nothing decision.

There is a significant difference between an agent that retrieves information, an agent that recommends an action, and an agent that independently executes a financial transaction.

These capabilities should not necessarily receive the same level of authorization.

A progressive approach can begin with read-only agents that analyze information without modifying business systems.

The next stage may introduce agents that prepare actions for human approval.

As organizations gain operational experience and establish reliable controls, certain agents can be authorized to execute predefined actions independently.

Higher-risk processes should remain subject to additional verification, monitoring, and intervention mechanisms.

The objective is not to remove humans from every decision.

It is to establish an appropriate relationship between autonomy, risk, and accountability.

An organization should grant autonomy based not only on an agent's historical performance but also on the potential consequences of failure, the quality of available information, and the ability to recover from mistakes.

This approach allows companies to innovate while maintaining control over critical operations.

NVIDIA OpenShell and Sheaf: Security Meets Verification

One of the most interesting topics in the recent webinar on the explosion of AI agents was the relationship between NVIDIA OpenShell and Sheaf.

These technologies address different, potentially complementary dimensions of agent security.

NVIDIA OpenShell focuses on controlling the execution environment of AI agents. Its approach establishes security boundaries around what agents can access and execute, including resources such as filesystems, networks, and processes.

Sheaf approaches the problem from another perspective: the mathematical consistency of decisions and information across agents.

This distinction is particularly relevant in enterprise environments.

An agent may have permission to execute a particular action without that action necessarily being appropriate.

Security policies can establish what an agent is authorized to do. Verification mechanisms can help determine whether its decisions satisfy defined consistency conditions.

However, neither mechanism should be considered a complete guarantee of operational correctness.

Enterprises still need identity management, authorization controls, monitoring, auditing, and human oversight.

In my view, the challenge is not necessarily to replace existing enterprise security architectures.

It is to extend them to accommodate a new generation of autonomous software actors.

Every AI Agent Needs a Verifiable Identity

Another important consideration is the identity of autonomous agents.

Enterprise systems have traditionally been designed around human users, service accounts, and applications operating under predefined permissions.

Autonomous AI agents introduce new questions.

Should an agent have its own digital identity? Should it operate exclusively under the identity of the person who requested an action? How can organizations distinguish between a human decision and an autonomous execution?

I believe agents should have identifiable technical identities, even when acting on behalf of human users.

This allows organizations to maintain a clear distinction between the person authorizing an action and the software component executing it.

For autonomous agents interacting with external organizations, this becomes even more important.

Imagine an agent negotiating with a supplier, initiating a payment, or communicating with another company's agent.

The receiving organization needs to establish which agent is involved, who operates it, what authority it has been granted, and whether that authority remains valid.

The concept of an Agent Passport, explored in the AI Agents 101 Atlas, offers a useful framework for this discussion.

Such a framework can include agent identity, operator information, delegated authority, permission scope, provenance, audit records, and revocation mechanisms.

This is not simply a technical concern.

It is fundamental to accountability in an increasingly autonomous digital economy.

Organizations can delegate authority to AI agents, but they cannot simply delegate away accountability.

The Difference Between a Verified Decision and a Correct Outcome

There is one technical challenge that I believe deserves particular attention.

Imagine an autonomous agent operating within a financial institution.

The agent analyzes a transaction, checks the relevant business policies, and verifies that the proposed action satisfies the required conditions.

Everything appears correct.

However, before the action is executed, another agent modifies the underlying data or changes the business context.

The original decision may have been valid when it was verified, but it may no longer be valid at the moment of execution.

This is not merely a theoretical possibility.

In distributed enterprise environments, information changes continuously, multiple applications operate simultaneously, and autonomous agents may interact with the same business processes.

Computer science already recognizes a related problem: time-of-check to time-of-use, commonly referred to as TOCTOU.

The challenge is ensuring that the conditions verified before an action remain valid when the action actually takes place.

Possible safeguards include revalidating critical conditions immediately before execution, using transactional controls, detecting changes in underlying data, and requiring human approval when uncertainty exceeds acceptable limits.

This also illustrates an important distinction between decision consistency and operational correctness.

A mathematically consistent decision is not necessarily based on complete or accurate information.

And even a reasonable decision can produce an unfavorable outcome when circumstances change unexpectedly.

Verification can reduce specific categories of risk, but it cannot eliminate uncertainty from real-world business operations.

This is why mathematical verification, execution security, and organizational governance should be considered complementary rather than interchangeable.

The Future of Enterprise AI Is Trusted Autonomy

I remain optimistic about the future of AI agents.

Their ability to integrate with existing business systems, automate complex workflows, and coordinate activities across applications could significantly transform how organizations operate.

But the next stage of enterprise AI will not be defined simply by the number of agents deployed or the volume of tasks automated.

It will be defined by the ability of organizations to demonstrate that autonomous systems create measurable value while operating within acceptable levels of risk.

Technology providers will continue improving agent capabilities, execution environments, verification mechanisms, and interoperability.

But enterprises must also evolve.

They need to rethink how authority is delegated, how digital identities are managed, how automated decisions are audited, and how human responsibility is preserved.

The real challenge is not choosing between innovation and control.

It is designing systems where both can coexist.

The future of enterprise AI is not uncontrolled autonomy. It is trusted autonomy.

And I believe the organizations that understand this distinction will be better positioned to transform AI agents from promising experiments into reliable components of their business operations.