From Barriers to Breakthroughs: Governing AI Agents for Safe, Scalable Adoption

Artificial intelligence has been reshaping industries for years, and a new chapter is already underway: agentic systems. These systems do more than classifying or responding - they take initiative, make decisions, and learn continuously with minimal human input.

According to Berkeley’s California Management Review, these systems are no longer on the horizon; they are maturing rapidly and becoming central to strategic advantage. The numbers underscore the urgency: the global AI market is projected to grow from nearly $300 billion in 2025 to more than $1.7 trillion by 2032, a compound annual growth rate of 29.2%. Adoption is no longer theoretical - it’s happening now. The question is how to make it resilient, scalable, and valuable.

The shift from automation to autonomy is profound. Automation has always been about efficiency - faster workflows, fewer manual steps. Autonomy goes further: it empowers systems to act independently, making decisions and executing tasks without waiting for human instruction. This brings enormous potential for productivity and innovation. However, it also introduces new questions: how do you govern systems that are non-deterministic, that act differently depending on context, and that surface risks that may not be immediately visible?

Berkeley’s research outlines five major barriers that slow adoption: technical infrastructure, organizational design, financial investment, human factors, and security concerns. These hurdles are real, but they are also governable. In fact, they are the very reasons governance needs to be rethought for the agentic AI era.

In the sections that follow, I’ll expand on each barrier through the lens of four priorities for enterprises: accelerating adoption, managing risk, governing with context, and assuring compliance and security.

Accelerate Adoption: From Pilots to Enterprise-Ready

Berkeley’s research highlights technical infrastructure as one of the most persistent barriers to adoption. This resonates with what we see across enterprises: sprawling technology landscapes with multiple cloud providers, siloed teams, and layers of legacy systems that require entire toolchains just to track who has access and when systems were last updated. Into this already complex environment, new AI pilots are often added as stand-alone experiments, fragile layers stacked on top of an unstable tower.

This approach creates fragility rather than resilience. It slows adoption because each new system adds complexity instead of reducing it. Berkeley’s insight is that infrastructure is a bottleneck, but the opportunity is to turn it into a platform for scale.

The reality is that adoption is already underway in leading industries. Banks are experimenting with agentic systems to automate compliance checks and detect fraud in real time. Manufacturers are piloting agents that orchestrate supply chain adjustments without waiting for human approval. Healthcare organizations are beginning to use agents to coordinate diagnostic workflows across departments and pharmaceuticals are knocking years off of drug discovery timelines with AI testing. These are no longer isolated experiments; they’re early signals of scale.

Moving from pilots to enterprise-ready adoption requires a shift in mindset. AI cannot be treated as a bolt-on tool. It has to be embedded into the flow of work, connected to core systems, and governed like any other enterprise-grade capability. When AI is integrated this way, it doesn’t add complexity - it helps tame it.

The organizations that accelerate adoption successfully are those that use governance as a confidence-building mechanism. AI adoption happens at the speed of trust. With resilience designed into integration, leaders can move from small-scale experiments to broad deployment without fearing collapse. This is adoption as a managed transformation, rather than a series of disconnected trials.

Manage Risk: From Uncertainty to Governable

Berkeley identifies financial investment and organizational design as central adoption hurdles. Leaders often find themselves caught between the need to invest in AI and the difficulty of proving ROI, while also struggling to align processes and cultures around new technologies.

This is where governance plays its most empowering role. Too many AI pilots are measured with “vibe metrics” - time saved, effort reduced, or other estimates created from subjective reporting. These metrics are useful for storytelling but rarely hold up under scrutiny. To move past this barrier, organizations need operational observability: systems that capture empirical data about AI’s impact, not just anecdotal evidence. This makes ROI demonstrable and builds the case for further investment.

On the organizational side, Berkeley’s research rightly calls out the difficulty of redesigning roles and processes around AI. The key is to flip the approach: instead of starting with the tool and asking, “Where can this fit?”, leaders should start with the problem. What gap are we trying to close? What does “good” look like when it’s solved? When framed this way, adoption becomes about adding capability, not just new tools. Using this approach, AI integrates naturally into workflows instead of forcing people to rebuild them from scratch.

Finally, the human factor. Berkeley frames this as a hurdle, but it is also the greatest accelerant. Many employees still see AI through the narrow lens of tools like ChatGPT or Perplexity, game-changing and useful, but hardly the full picture. By demystifying AI and showing it as a spectrum - from Excel formulas to anomaly detection to diagnostic discovery - leaders empower employees to think creatively about its applications. Adoption succeeds when people are inspired to innovate, rather than adapting reluctantly.

Risk, in other words, is not something to be avoided. With the right design, measurement, and training, risk becomes governable and governance becomes the foundation for moving faster.

Govern with Context: Seeing How Agents Behave and Why

Among the barriers Berkeley highlights, human factors of trust and transparency are perhaps the most underestimated. For traditional software, testing and certification provide assurance. However, agentic systems are non-deterministic, meaning they don’t always act the same way twice. Pre-deployment testing alone cannot guarantee performance or safety.

This is where contextual governance becomes essential. Instead of relying on static validations, organizations need continuous visibility into how agents behave, why they act, and under what conditions. Context matters: the same agent may make different choices in different scenarios. Without visibility into the “why,” trust remains fragile.

Berkeley describes cultural adoption as a hurdle with employees struggling to trust or understand AI. We see this differently: cultural adoption is the accelerant, provided the right governance is in place. When employees, customers, and regulators can see why agents act as they do, trust is no longer blind faith. It is observable, explainable, and accountable.

Transparency doesn’t slow adoption - it enables it. Contextual governance gives leaders confidence to scale by showing that AI systems are not black boxes but explainable systems operating within known parameters. This transforms governance from a compliance function into a growth enabler.

Assure Compliance & Security: From Claims to Demonstrable Proof

Berkeley’s research identifies security concerns as one of the most critical barriers. This is especially true for agentic systems, which operate more like autonomous contractors than static software. They take actions, make decisions, and sometimes surprise even their creators. That autonomy is their strength, but it also requires a new model of assurance.

Traditional guardrails are insufficient. Agents operate at machine speed, and static rules can be bypassed or outpaced. What’s needed is adaptive security and compliance frameworks that evolve alongside the systems themselves.

Here, operational observability becomes the cornerstone. Compliance can no longer be claimed; it must be demonstrated in real time. This allows leaders to show regulators, customers, and partners that systems are operating safely - not as a promise, but as proof.

Far from slowing innovation, these frameworks accelerate it. Just as brakes allow a car to go faster by providing control, compliance and security mechanisms allow enterprises to deploy agentic systems confidently. Safety and innovation are not opposites; they are complements.

By reframing security as an enabler, not a barrier, organizations transform one of the biggest hurdles Berkeley identifies into a source of competitive advantage.

The Urgency of Action

Berkeley’s research makes it clear: agentic AI is not a future possibility, it is a present imperative. The barriers are real, but they are not reasons to delay. These barriers are the very areas where governance, done well, creates advantage.

Adoption is already underway across industries, though unevenly. Some enterprises are moving decisively from pilots to scale, while others remain stuck in experimentation. The difference lies not in technical capability, but in governance: the ability to accelerate adoption, manage risk, govern with context, and assure compliance and security.

The leaders who act now will not only capture efficiencies but will also shape the standards for their industries. And they won’t do it alone. Partnerships with technology providers, regulators, and peers will be critical in ensuring adoption happens responsibly and consistently. Ecosystem collaboration is how resilience scales.

Conclusion: Empowered Adoption

Agentic AI is here. There’s no question it will transform industries. The opportunity is there for leaders to govern adoption so that it empowers their people, customers, and partners.

The pathway is clear:

• Accelerate adoption by moving decisively beyond pilots.
• Manage risk with operational observability that makes outcomes measurable.
• Govern with context to provide continuous visibility into agent behaviour.
• Assure compliance and security through demonstrable, adaptive frameworks.

Berkeley’s research rightly identifies the hurdles felt by businesses globally.

Our perspective is that these hurdles are also the opportunities. With the right governance, they don’t slow adoption - they enable it.

Adoption done this way doesn’t just protect the business. It empowers it. It creates the confidence to move faster, scale further, and deliver greater value. And not just for the enterprise, but for customers and partners who gain trust and clarity in every interaction. Enterprises that govern adoption now will build the trust customers expect, the resilience partners need, and the confidence regulators demand - creating lasting advantage in the process.

‍

‍

‍

‍

‍

Additional Reference:

1: Berkeley Study (2025): https://cmr.berkeley.edu/2025/08/adoption-of-ai-and-agentic-systems-value-challenges-and-pathways/

‍