From generative AI to agents: why enterprise AI governance must evolve

August 2026 | SPOTLIGHT | RISK MANAGEMENT

Financier Worldwide Magazine

August 2026 Issue

Generative artificial intelligence (genAI) has rapidly moved from experimentation to mainstream business adoption.

Across many organisations, AI is no longer limited to drafting emails or summarising documents. Rather, it is increasingly embedded into customer service, software development, regulatory review, knowledge management, cyber security operations and business process automation.

The next stage of this evolution is even more significant: AI agents.

Unlike traditional genAI tools that primarily produce content, AI agents can retrieve information, interact with applications, execute workflows and make decisions within predefined boundaries. In practical terms, they are beginning to function as digital workers operating alongside human employees. This shift fundamentally changes the enterprise risk landscape.

When organisations first adopted generative AI, the primary concerns centred on data leakage, inaccurate outputs and employee misuse. Those risks remain relevant, but autonomous agents introduce a new set of governance questions. What systems can the agent access? What actions can it perform? How much autonomy should it receive? And ultimately, who is accountable when an AI-driven action creates unintended consequences?

These questions extend far beyond cyber security. They sit at the intersection of governance, privacy, compliance, legal risk, operational resilience and business accountability.

The governance landscape is maturing

Regulators and standards bodies are increasingly focusing on operational accountability rather than high-level AI principles.

The European Union’s AI Act represents a significant milestone in this evolution. While much public discussion has focused on AI providers, the regulation also places responsibilities on organisations deploying AI systems within business operations. For multinational enterprises, the implications are global. AI-enabled services, third-party platforms and cross-border data flows rarely align neatly with national boundaries.

At the same time, industry frameworks such as the National Institute of Standards and Technology’s AI Risk Management Framework and the International Organization for Standardization International Electrotechnical Commission 42001 provide organisations with practical approaches for governing AI-related risks.

Their importance lies not in compliance alone but in their ability to translate abstract governance principles into measurable controls, defined responsibilities and auditable processes. For enterprise leaders, the message is increasingly clear: AI governance is becoming a core business capability rather than a voluntary exercise.

Visibility remains the first challenge

Despite growing investment in AI governance, many organisations struggle with a fundamental problem: they do not know where AI is being used.

Employees adopt public AI tools for productivity. Business units purchase AI-enabled software as a service platforms. Development teams integrate large language models into applications. Vendors continuously introduce AI functionality into existing products. As a result, AI often spreads throughout the organisation faster than governance processes can adapt.

This phenomenon, often described as ‘shadow AI’, creates significant challenges. Sensitive information may be processed through unapproved tools. Business decisions may become dependent on third-party models. And new operational risks may emerge without formal review.

Before organisations can govern AI effectively, they must establish visibility. An enterprise AI inventory should identify where AI is being used, who owns the solution, what data is processed, which systems are connected, and how outputs influence business decisions. Without this foundation, meaningful risk assessment becomes difficult.

Visibility is not merely an administrative exercise. It is the prerequisite for effective governance.

Not all AI requires the same controls

One of the most common governance mistakes is treating all AI systems as if they present identical risks. In reality, risk varies significantly depending on how AI is used.

A writing assistant that helps employees prepare internal communications presents a very different risk profile from an autonomous agent capable of updating customer records, initiating financial transactions or influencing clinical decisions.

Effective governance therefore requires a risk-based approach. Organisations should evaluate AI systems based on several dimensions, including data sensitivity, level of autonomy, business impact, regulatory exposure and dependence on external providers.

Systems processing regulated data or supporting critical decisions naturally require stronger controls than tools used for low-risk productivity activities.

The rise of AI agents makes this distinction particularly important. Once a system can take actions rather than simply generate recommendations, governance expectations must evolve accordingly.

Governing AI agents as digital workers

Many organisations continue to view AI agents as software applications. A more useful perspective is to view them as digital workers.

Like employees, AI agents are assigned responsibilities, granted access to systems and expected to perform tasks on behalf of the organisation. Unlike employees, however, they can operate continuously, scale rapidly and interact with multiple systems simultaneously. This makes governance essential.

Every AI agent should have a clearly defined business purpose, designated owner, approved operating scope and documented access permissions. Organisations should resist the temptation to deploy broadly empowered agents simply because technology makes it possible.

Access rights should remain tightly aligned with business purpose. For example, an agent designed to summarise internal policies should not possess the ability to send external communications. A customer service agent may assist with information retrieval and response generation but should not independently modify customer records without appropriate controls.

A security operations agent may support investigation and triage while enforcement actions remain subject to predefined approval requirements. The principle is straightforward: authority should never exceed purpose.

Identity and access management matters more than ever

As AI agents become operational participants in enterprise environments, they must be governed as non-human identities.

Each agent should have a unique identity, clearly assigned ownership, controlled credentials and role-based permissions. Shared accounts, embedded credentials and unrestricted access create unnecessary risk and undermine accountability.

Organisations frequently focus on model governance while paying insufficient attention to identity governance. In practice, excessive access privileges often represent a greater risk than model behaviour itself.

An AI agent with broad access across enterprise systems effectively becomes a privileged digital user. Without appropriate oversight, the resulting exposure can exceed that associated with many human accounts.

Applying established identity and access management principles to AI agents is therefore not optional – it is foundational.

Auditability and human oversight

As AI systems become more deeply embedded in business operations, organisations must maintain the ability to understand how decisions are made and actions are executed. This requires appropriate logging, monitoring and auditability.

Records should capture key information such as instructions received, systems accessed, actions performed, approvals obtained and outcomes generated. These records support incident investigations, regulatory inquiries, internal audits and accountability reviews.

Equally important is the role of human oversight. Organisations should distinguish between AI systems that recommend actions and those authorised to execute actions. High-impact activities involving legal commitments, financial transactions, personal information, security enforcement or customer outcomes should generally remain subject to human review and approval.

This approach allows organisations to capture the efficiency benefits of automation while preserving accountability for consequential decisions.

Preparing for failure before it happens

Responsible AI governance assumes that failures will occur. This does not imply a lack of confidence in the technology. Rather, it reflects a mature understanding of operational risk.

Organisations routinely design resilience mechanisms for infrastructure, cyber security incidents and business continuity events. AI systems deserve the same level of preparation.

Practical safeguards may include transaction limits, approval thresholds, escalation procedures, rollback capabilities and emergency shutdown mechanisms. The specific controls will vary according to the use case, but the underlying principle remains consistent: organisations must retain the ability to intervene when autonomous behaviour deviates from expected outcomes.

The goal is not to eliminate autonomy. The goal is to ensure that autonomy remains controllable.

Governance is a business capability

Perhaps the most important shift is cultural rather than technical. Many organisations still treat AI governance primarily as a compliance requirement. This perspective is increasingly outdated.

Effective governance does not slow innovation. It enables innovation. Organisations with weak governance often face one of two outcomes. They either move cautiously because every AI initiative appears risky, or they move aggressively and accumulate unmanaged exposure.

Organisations with mature governance frameworks operate differently. They understand where AI is used, how risks are evaluated, who owns key decisions and what controls are required before deployment. As a result, they can make faster and more confident decisions.

The next phase of enterprise AI will not be defined by how many models are deployed or how many agents are created. It will be defined by whether organisations can govern those capabilities responsibly, transparently and sustainably.

For chief information security officers, privacy leaders, compliance professionals and business executives, this represents a significant opportunity. Governance is no longer simply about reducing risk. It is about creating the trust required for AI to deliver business value at scale.

In the years ahead, trust – not technology alone – will become the defining factor in successful AI adoption.

Great Gu is global chief information security officer at GenScript. He can be contacted by email: bobbycoming@outlook.com.

Great Gu

GenScript