Stop Building AI Agents. Start Building the System to Run Them.

Agentic AI works in demos. It stalls in production. Ensono’s Chief AI Officer explains the five barriers keeping enterprises stuck in pilot mode — and why the path forward starts with orchestration, not more agents.

The Industrial Revolution replaced human muscle with machines. AI is replacing human brainpower with systems. That’s a fundamental shift, and enterprises aren’t ready for it.

According to Forrester’s The State of Agentic AI, 2026 report, “three-quarters of leaders at enterprise organizations report they are adopting agentic AI … but only a small minority have reached meaningful production applications.” The rest are stuck in what Forrester calls “proof-of-concept purgatory” — impressive experiments that never become operational systems.

The technology works. The demos are compelling. So why can’t companies scale?

Forrester identifies several barriers to scale. Having helped enterprises navigate this shift firsthand, I’d add that they all connect to a deeper problem: Most companies are trying to fit a fundamentally new capability into fundamentally old structures.

Automating Bad Processes

The biggest mistake I see is one we should have learned from the automation era: Just because you can automate something doesn’t mean you should.

Too many companies take a 20-year-old business process, stick an agent in the middle, save six minutes and call it transformation. That’s not a transformation; that’s a faster old process.

Agentic AI gives us a rare opportunity for zero-based process design where we can start from the desired outcome and work backward. The biggest value won’t come from making today’s company faster. It will come from designing the company you would have built if AI workers had always existed.

ROI Uncertainty

Proving value is the top barrier to scale. But the problem starts earlier than measurement — it’s a strategy problem first.

Companies need to ask the right questions first. Instead of starting with, “Where can I use AI?” try, “Where do I have an expensive business problem that AI could materially change?”

Here’s my rule: If you can’t tie your AI investment back to core business metrics you’ve tracked for years — such as customer satisfaction, ticket resolution, outage reduction or others — you’re doing something wrong. If you don’t see meaningful change in those metrics, then AI isn’t providing you with any benefit. In that case, you should reconsider what you want strategically out of AI, and what difference you want it to make in the business.

Governance that Can’t Keep Pace

Most enterprise governance was designed around humans asking systems to do things. Agentic AI changes the equation. Now software can decide what to do next.

You can’t govern that with a PDF. Governance must run at the same speed the agent does. Real-time. Automated. Embedded in execution. Most companies haven’t built systems for that, which is why even well-intentioned policies don’t prevent sprawl, drift and risk.

Agentic Sprawl

Shadow AI is the new shadow IT.

One team builds a procurement agent. Another builds a finance agent. Someone else builds an operations agent. Soon you have hundreds of digital workers talking to the same systems with overlapping responsibilities — and no one has a clean inventory of what exists.

That’s not an AI problem anymore. That’s a distributed systems problem. The question isn’t how many agents you have. It’s whether you know who they are, what they’re allowed to do and who’s accountable when something goes wrong.

Forrester found that “41% of genAI decision-makers cite coordination challenges or failures with multiagent environments” and “49% of security decision-makers cite agentic AI as a concern.” Both are justified.

An agent isn’t just another application; it’s an independent identity with judgment. It can access data, invoke tools, generate code and initiate actions. And unlike traditional automation, it’s probabilistic. You can’t predict with 100% certainty what it will do. You wouldn’t give an intern unrestricted admin credentials on their first day. That same logic applies to AI.

Platform Confusion

The vendor landscape is shifting fast. The model that was leading six months ago might be second-tier today. New players keep emerging at a lower cost. Companies stall because they don’t know which bet to make. The key here is that you don’t have to make a permanent one. The companies that scale are building flexibility into their architecture, not locking into a single vendor’s vision.

The Path Forward: Orchestration First

All five barriers point to the same underlying gap: Companies are adding agents without building the system to manage them.

Orchestration is foundational, not optional. Before you add more agents, you need a control plane: a registry, identity management, routing, permissions, observability. You need to know what agents exist, what they’re allowed to do and who owns them.

At Ensono, we’re building exactly that. Our vision is that agents will appear in our systems alongside humans. They’ll get performance reviews. Are they worth the cost? Are we getting the outcomes we expected? What’s the error rate?

That’s how you manage a digital workforce.

We’ve designed the platform to be malleable. We can swap agents, swap systems, mix and match — whatever best supports our clients. Because if there’s one thing I’m certain of, it’s that the landscape will keep changing, and we have to be ready to pivot.

What Separates the Winners

A year from now, we’ll stop talking about AI projects and start building AI operating models. Where do agents live? Who owns them? How are they governed, measured and held accountable?

Five years from now, companies will have two workforces — human and digital — working side by side. The competitive advantage won’t be access to smarter models. Everyone will have that. It will be how well you’ve designed your organization to operate with AI workers as a given, and at scale.

Your AI architecture may look more like an org chart than an application diagram. And the companies still stuck in proof-of-concept purgatory will be the ones who kept chasing the technology without building the foundation to run it.

Frequently Asked Questions

What is “proof-of-concept purgatory” in agentic AI?

Pilot purgatory is when companies have impressive AI experiments that never become operational systems. According to Forrester, 75% of enterprises are adopting agentic AI, but only a small minority have reached meaningful production.

Why can’t companies prove ROI on agentic AI?

It’s a strategy problem before it’s a measurement problem. Too many companies start with “Where can I use AI?” instead of “Where do I have an expensive business problem that AI could materially change?” If you can’t tie AI investment back to core business metrics like customer satisfaction or ticket resolution, the pilots won’t get funded to scale.

What is agentic sprawl?

Agentic sprawl is when teams across an organization build agents independently — procurement, finance, operations — without a clean inventory of what exists. Soon you have hundreds of digital workers with overlapping responsibilities and no clear accountability. It’s the new shadow IT.

What does “orchestration first” mean?

Orchestration first means building the system to manage your agents before adding more of them. That includes a registry, identity management, routing, permissions, and observability. You need to know what agents exist, what they’re allowed to do, and who owns them — before you scale.

Forrester, The State Of Agentic AI, 2026, 29 May 2026

Information in Forrester publications is based on Forrester’s efforts to compile and analyze the best resources reasonably available to Forrester at any given time. Opinions reflect judgment at the time and are subject to change. This report is part of a broader collection of Forrester resources, including interactive models, frameworks, tools, data, and access to analyst guidance.