Agentic AI is Here, But Most Enterprises Are Still Stuck. Why? 

Ensono’s Chief AI Officer and Chief Technology Officer weigh in on the gap between agentic ambition and production reality — and what it actually takes to close it. 

A Forrester report, The State of Agentic AI, 2026, finds that three-quarters of enterprise leaders at enterprise organizations report they are adopting agentic AI, but only a small minority have reached meaningful production. 

The contradiction revealed by this research is striking. The technology has arrived. The investment is flowing. So what’s getting in the way? 

We put that question — and several others drawn directly from the report’s findings — to Jim Piazza, Ensono’s Chief AI Officer, and Tim Beerman, Ensono’s Chief Technology Officer. 

Forrester reports that three-quarters of enterprises are adopting agentic AI, but only a minority have hit meaningful production. Does that match what you’re seeing with clients? 

Jim: It largely depends on the client and where they are in their lifecycle. I just got out of a conference last week where only 11% of companies had even really taken a first step toward doing something where they can produce meaningful, production-related stats — and that’s not just in the agentic space, that’s just in general. So it varies quite a bit. 

That said, there are clients who have hit some pretty interesting milestones. At Ensono, we’ve always wanted to be able to back up any claim we make with real data — and we can do that. When we look at the work we’ve done over the last eighteen months, we have meaningful metrics where we can say we improved in five areas that are all directly related to the experience our clients expect. And we made meaningful progress in a positive way across all of them. 

Tim: The maturity of what people consider to be truly agentic matters a lot here. The whole agentic framework has matured. It used to be — and even kind of now is — “built a chatbot agent.” Well, that’s an agent, but it’s not really what modern AI views as an agentic workflow. True agentic AI is actually delivering a work output autonomously, or mostly autonomously. And that’s where a lot of companies are still maturing. The industry is in this pivot right now, and the tools to really make this a reality are just now coming to market. 

When a client comes to you stuck in pilot mode — it technically worked, but they can’t get it to scale — what’s usually the first thing you diagnose? 

Tim: One of the first things is to understand what business problem they’re actually trying to achieve. And then if you start backing up from there — what state is the data in that you’re using to make this decision? A lot of the accuracy and functionality of AI comes down to how good the data is that you’re using to inform decisions. And that’s where a lot of people just really aren’t that mature yet. You’re taking things like a document repository or a knowledge set, and most people’s data is just not in a great position.  
 
At Ensono, we’ve spent a lot of time as we’ve gone down this journey just to get our data in order. So if people look at going from POC to production, the question is: do you have that scale, that knowledge base? Are you really solving that business problem and getting that measurable outcome? 

The Forrester report calls out four main culprits for failed production deployments: weak orchestration, poor governance, immature platform choices, and ROI uncertainty. If you had to rank those, what you rank number one? 

Jim: For me, the number one would be ROI uncertainty. If you don’t start understanding what the business problem is you’re trying to solve, what process you have in place today, and at least six months of historical metrics — you’re just throwing darts at a dartboard hoping for the best. You have to start in those places before you can do anything. 

Platform choices, I’ll be honest, that’s probably last. The platforms are evolving so quickly right now, you just have to pick a horse and go. If you wait and wait and wait, you’re going to fall further and further behind. The key is to make sure your platform is nimble so you can swap components out over time as different pieces mature at different rates. 

Between weak orchestration and poor governance, I’d edge governance above orchestration. Orchestration can be fixed — that’s a pretty easy one. But if you get governance wrong out of the gate, you’re going to be paying the sins of the past for a long time.  

Tim: I’d add one thing that cuts across all of this. A lot of people go down this path without reimagining how to solve the business case with AI in mind — taking an already-inefficient process and trying to throw AI on top of it. It really ties back to that ROI uncertainty. They’re not asking “how do I do this differently with AI?”, and that’s where things fail. 

Most companies believe they have AI governance — a policy document, maybe a framework. But you’ve drawn a distinction between governance as a statement and governance as a verifiable record. What does real, measurable governance actually look like? 

Jim: A policy document is a really good place to start, and it’s an absolute requirement. But as part of real governance, you also need a verifiable record. First, you have to be able to discover the agents that are out there — especially if you allow people to create their own agents within your company. Second, they have to have an identity. Whether you choose a human-based identity or a system-based identity is up to you and your business process, but having agents have their own identity is important — because that comes along with security privileges. And then an audit log of everything that agent has done needs to be verified and independently auditable. 

You put all three of those pieces together, and you can hand basically a digital envelope to somebody that says: this agent is known, has these permissions, and here is the audit trail record of what it did. That becomes immensely more important when you start allowing agents — or AI itself — to self-improve, because that audit log is going to tell you not only what it did, but the modifications it may or may not have made along the way. Without that sealed digital envelope, you’re going to be confused about why there was an outcome that was unanticipated.

The report talks about “agentic sprawl” — unregistered agents, duplicated logic, and privilege drift across environments. Is this something you’re actually seeing, or is it still more theoretical? 

Tim: It’s absolutely something we’re seeing. You’re always balancing the desire to enable your workforce to experiment and bring value to their department against what actually happens, which is a lot of people doing similar things, building something and not going through and cleaning it up.

This is where the governance Jim talked about becomes very practical. Can I see which agents are accessing the same data sets? Do I have a requirements process that minimizes duplication upfront? Agentic sprawl is a real risk and a real phenomenon, and it materializes itself in different ways depending on the level of governance a company has and the tools they’re allowing to be used.

The report flags nonhuman identity as a specific gap. Agents without proper credentials, ownership, or lifecycle management. How does that play out in practice, and why does it matter for production readiness? 

Jim: Every agent, whether it has a system-based identity or a human-based identity, needs that identity clearly defined. A human-based identity means the agent is allowed to act like a specific person — and that means you’re on the hook for what that agent did and what access levels it has. A system-based identity is a system account with a defined set of privileges attached to it. 

The difference matters because if you allow an agent to impersonate a human, that person may have elevated privileges because they’re trusted. And where you don’t want elevated privileges is when you’re working with production systems that may have unintended consequences. 

I’m a big believer in coupling probabilistic analysis with deterministic execution. If I’ve limited not only the role an agent can take and the privileges associated with it, but also limited its execution capability by making it deterministic — that really minimizes the risk of adverse reactions. You put those two things together and you get a more guaranteed outcome while still getting the benefits of AI reasoning. 

Tim: The only thing I’d add is the kill switch capability. Having the ability to quickly stop something if it goes wrong requires having the right level of monitoring and event handling in there. And you can’t stop what you can’t see — which gets right back to the identity and registration question. 

The report argues that companies layering agents onto legacy workflows “rarely achieve step-change value” — that you have to redesign the workflow itself. But full redesigns carry real risk. How do you help clients do that without it becoming a massive transformation program that never ships? 

Jim: I think you can augment and AI-fy an existing workflow and get some incremental value out of it, and you’re going to learn a lot through that work. That learning is going to give you great insights into how you need to reimagine that workflow into something that is more agentic. That’s where you get your step-function improvement. 

But think about what it would take to go back and redesign a quote-to-cash system from zero, for example. The amount of work to go from that to a step-function improvement is going to take months. Why not get the quick wins in first? You’re going to learn a lot, you’ll have the data, you’ll have the business know-how, and you’ll have the fundamental metrics you need to gauge whether you’re successful. 

If a client came to you today and said they want something in production in 90 days, where would you point them? 

Tim: Some of the things that have the fastest time to value — and this isn’t necessarily fully agentic — are around knowledge work. Working with large bodies of data, being able to search and provide value, whether it’s contracts, preparing for an RFP, drafting a SOW. As long as you get your data in a good spot, there’s really quick potential value there. That’s also where we’re seeing even our own internal teams get that value. It’s also an area that’s ripe for duplication, so you have to be careful. But that’s probably the one that comes to mind first. 

Jim: I’d reinforce that. Chatbots get dismissed sometimes because it’s hard to measure the productivity increase. But I think it’s a great entry point — and that work to collate and organize data is going to pay dividends down the road regardless. It’s not hard, it’s a quick win, and it will pay value back to the business longer term. 

Last question: What’s your single piece of advice for the enterprise leader who has budget, executive support, and technical capability — but is still stuck? 

Jim: Identify where your big buckets of work are. Determine how much time, energy, and resources go into those buckets. Then start breaking that down into smaller chunks you can execute on rapidly. Maybe every sprint cycle you get another one or five percent of that process understood and worked through — made agentic, whatever the goal is. You don’t have to tackle the whole thing at once. The Big Bang approach for existing companies with tech debt and legacy data is just such a daunting problem. Allow people to think in a microcosm until they get their head wrapped around it. And once they get a couple of wins under their belt, that will continue to breed more success. 

Tim: I agree with that. The thing I’d add is prioritization. If you have too many things going in different directions across different technology platforms, you’re not really organized, and that’s where a lot of the frustration comes from. “I’m spending all this money on all these different tools, I’m not sure who’s on point, and I’m not sure where the value is.” Prioritization should guard and focus limited resources. It shouldn’t be something that stymies you. It should be something that helps you focus and actually build something you can grow. 

Frequently Asked Questions

What is agentic AI?

Agentic AI refers to AI systems that can autonomously, or mostly autonomously, complete work toward a defined goal. Unlike a basic chatbot that primarily responds to prompts, agentic AI can reason through tasks, make decisions and take actions across a workflow. 

Why do agentic AI pilots struggle to reach production?

Common barriers include unclear business value, poor data quality, weak governance, inadequate orchestration and difficulty scaling AI across existing workflows. Establishing measurable business outcomes before deploying the technology can help organizations determine whether an initiative is actually delivering value.  

What does effective AI governance look like for AI agents?

Effective governance goes beyond a policy document. Organizations need visibility into the agents operating across their environments, clearly defined identities and permissions, and auditable records of agent activity. These controls become increasingly important as agents gain greater autonomy.  

What is agentic sprawl?

Agentic sprawl occurs when AI agents proliferate across an organization without sufficient coordination or oversight. This can lead to duplicated functionality, overlapping access to data, unmanaged permissions and difficulty determining which agents are operating within an environment.

How can enterprises move agentic AI from pilot to production?

Start with a clearly defined business problem and measurable baseline, then break the opportunity into smaller use cases that can be implemented and evaluated quickly. Early wins can provide the data, operational experience and governance foundation needed to expand agentic AI over time.