Digitate’s Ugo Orsi offers this commentary which asks the question of who is holding the reins on agentic AI. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.
According to Deloitte’s 2026 State of AI in the Enterprise report, nearly three in four companies plan to deploy agentic AI within the next two years. That statistic comes as no surprise, as leaders are understandably excited about self-directing AI agents. No-code AI technology lets organizations build systems that can manage complex workflows, make autonomous decisions in real-time, and perform tasks that previously required 24/7 human monitoring.
But here’s the uncomfortable truth hiding behind that statistic: only 21% of enterprises have mature governance around those agents.
As organizations race to capture the competitive advantage of agentic AI, they’re creating a new problem – agentic sprawl. Fragmented deployments, siloed workflows, disconnected automation efforts, and agents operating without oversight are becoming a defining feature of early enterprise AI programs. The result is not transformation but, often, complexity.
For manufacturing leaders rethinking digital architectures in a cloud-first world, the question isn’t whether to adopt agentic AI, but rather who’s really holding the reins.
Fragmented Islands in a Data Lake
Most IT workflows were designed around human-led systems performing linear processes. Not anymore. Modern ops environments require systems that can respond to multiple data streams simultaneously – and many of those decisions will need to be automated.
Agentic AI thrives under that kind of pressure. Software agents let organizations break free of linear workflows by decentralizing complex, autonomous tasks and making decisions at speeds no human team could match. It’s powerful stuff. But as adoption trends show, there’s a trap.
In their hurry to roll out the next agent, too many leaders deploy them in complete isolation. Siloed agents are designed to optimize one task. Few ask how those agents fit into the overall operating ecosystem. Next thing you know, your autonomous components don’t communicate, they don’t share data, and they’ start working against each other. Agents become fragmented islands that increase complexity instead of decreasing it.
Slowing deployments to think about integration seems counterintuitive. It’s not. Think about agents as you would new employees. Connect them. Every new agent should know where it fits into the larger ecosystem. Clarify how new agents interface with existing systems and with other agents. Formalize data access so every agent can scale its decisions across your operation. Instead of increasing fragmentation, each agent should help unify the ecosystem.
Governance Multiplies Agentic Value
One of the most persistent misconceptions in enterprise AI is that governance limits agentic potential. The fear is understandable: organizations worry that imposing structure on autonomous agents will constrain their ability to operate and adapt, producing systems that can’t do much more than their initial programming.
That fear is misplaced. If you want agents to realize their full potential, give them freedom within guardrails.
Governance isn’t a bottleneck – it’s what allows agents to scale thoughtfully. Organizations that establish governance models early don’t just mitigate risk; they create the guardrails agents need to improve their decision-making over time. The most effective governance frameworks balance control with autonomy, giving agents the flexibility to adjust based on incoming data while maintaining clear accountability for how decisions are made.
There’s an important distinction to draw here: an agent that never violates your rules is not necessarily a good agent. Rigid rule-following isn’t the goal. The goal is to create an agent that can learn, adapt, and improve within a defined trust framework – one that gives human teams full visibility into what decisions are being made and why.
For manufacturing environments where operational continuity is non-negotiable, this kind of governance-by-design isn’t optional. It’s the foundation everything else is built on.
Disconnected Agents Create Liability
When asked about their agentic AI deployments, most enterprises will point to a handful of isolated pilots. An agent managing a procurement workflow here. Another monitoring production line anomalies there. Isolated use cases. Point solutions. Agents are not connected to anything else.
Disconnected agents introduce unnecessary risk. An agent can only operate as well as the data it has access to. Siloed agents will always try to fill in blanks with assumptions. At best, your agents become bottlenecks. At worst, they cause problems you’ll have to manage elsewhere.
Connecting your agents should be non-negotiable. One rule to live by with agents: if an agent is only responsible for a single workflow, you’re doing it wrong. Integrate agents with each other and shared data sources. Not just to reduce risk, but to improve decisions across your operation. Connected agents have a network effect. Alone they’ exponentially more valuable. Together they can drive your business.
Stop Piloting and Start Scaling
Is there anything wrong with pilots? Of course not. There is nothing wrong with experimentation and pilots have a place in any responsible AI program. However, many organizations never get past the pilot phase. They’ll launch an agent into production. Call it a pilot. Test drive that agent for months, wondering why they’re not seeing value, then move on to the next shiny object. Pilots without end users are symptoms of a bigger problem. The simple fact is that AI agents in production are either creating value or they aren’t. Organizations stuck in piloting purgatory spend more time growing complexity than capability.
My advice? If you want to scale AI in your organization, try scaling away from pilots. Start by defining your deployment plans – and what success looks like along the way. When does an agent graduate from pilot status? 90 days? Six months? How will you know when it’s ready to operate beyond its initial use case? Plans should be flexible enough to adjust as you learn more about what you’ve built. But every organization should have a plan. Because without one, your AI initiatives won’t grow. They’ll just drift. And that drift is costly.
Conclusion
Agentic AI isn’t magic. But for organizations who take the right approach to architecture, governance, and scale, AI agents can be transformative. That said, unlocking that value isn’t easy.
Think of deploying an AI agent the way you would onboard a new hire. They don’t arrive knowing your systems, your culture, or your workflows. But with the right structure around them, they learn. They improve. And eventually, they stop working alongside your organization and become part of it.
The enterprises that treat governance as an afterthought will spend years untangling the fragmentation they created in their rush to deploy. The organizations that do it right won’t be the ones that simply deployed the most agents the fastest. The ones setting the pace will be those that take ownership now, ask who’s holding the reins and build the right infrastructure first, putting orchestration frameworks, governance models, and integration strategies in place that allow agents to collaborate, learn, and scale responsibly.
