Q1. Could you start by giving us a brief overview of your professional background, particularly focusing on your expertise in the industry?

I currently lead the AI strategy for sales and operations at a global manufacturing company. Before this, I spent about twenty years working in operations, supply chain, and digital transformation—mainly focused on process improvement. That’s how I earned my Lean Six Sigma Black Belt, and it’s also where many of my instincts around AI were formed.

These days, my main focus is applying agentic AI in industrial environments. In practice, that means figuring out which parts of the value chain can be handed off to autonomous systems, what needs to change around those systems for the transition to really pay off, and where people still need to stay involved. Most of the failures I see aren’t technical—usually, it’s because a smart agent gets dropped into a process that no one bothered to rethink. The work keeps flowing the old way, so any gain from the technology never really spreads beyond one person’s desk.

I’ve had the chance to take multi-agent systems all the way from design into live production, and I also led a division-wide AI task force that brought together twelve different functions to agree on a shared roadmap. So I’ve worked on both sides of the challenge: understanding what the technology can do, and what an organization is actually willing and able to adopt.

Q2. Is resistance to enterprise AI investment actually growing because companies are struggling to prove ROI, and what are you seeing on the ground that tells you whether this is a temporary hurdle or a deeper problem with the economics?

Yes, and it is a problem with how the money was budgeted, not a rough patch.

Four things have to be paid for before AI pays back: 

• Getting the data into a usable state, 
• Redesigning the process around the tool
• Training people and deciding what their freed hours 
• Infrastructure to run and monitor it

The only thing that shows up as an invoice is the license. The other three—internal effort from people who already have full-time roles—never get formally budgeted or assigned. As a result, the program only gets a fraction of the funding it actually requires. The missing resources weren’t rejected; no one ever formally requested them.

Then the project goes live. The data isn’t in good shape, so everything gets delayed. No one has changed the underlying process, so the tool just ends up layered on top of the old workflow, and any improvements only benefit whoever happens to use it. Fast forward a year, and the board is reviewing the costs, but there’s no clear line item in the accounts to explain the spending. From their perspective, it’s a mystery—though everyone working on the project knows exactly why it happened.

This matches what we’re seeing in broader research. For instance, McKinsey’s State of AI 2025 survey—covering about 2,000 respondents—shows that only 39% report any EBIT impact, most of it under 5%. Just 21% of organizations had actually redesigned a workflow at the core, and the top performers are in that group. While it’s self-reported and only shows correlation, not causation, the trend lines up with what you’d expect if budgeting for change is the real problem.

This is exactly what I see in heavy industry. The real obstacles aren’t usually about the AI models themselves—it’s more about messy data, the challenge of scaling beyond a single site or department, and the fact that governance only shows up after deployment, instead of before. Demand isn’t the limiting factor either; people will use these tools whether they have official approval or not.

So I think the skepticism is justified—and honestly, I wouldn’t want it to disappear. It’s forcing boards to ask two questions that often got skipped in the first wave: What kind of return would make this effort worthwhile? And what were we measuring before we started? If you don’t set a baseline up front, you’ve already made it impossible to prove the program’s impact.

Q3. Based on your experience deploying agentic AI, which business processes are proving most suitable for autonomous or multi-agent workflows, and why?

Three things decide it, and none of them are about the model.

Can you check the answer cheaply? 

If it costs just as much to check the work as it does to do it yourself, then autonomy isn’t really saving you anything. This is the kind of problem that can slip by unnoticed—because a wrong answer that sounds convincing looks exactly like a right one. Make sure to check up front, or you might not spot the issue for months.

How expensive is a mistake to undo? 

That’s really what determines how much autonomy you can give a process, and it’s more important than whether the work is analysis or execution. Many execution tasks—like order entry, document handling, or routine transactions—are great candidates, as long as any mistakes are easy and inexpensive to undo. The most straightforward way to keep things safe is to limit what the agent is allowed to touch, instead of just piling on extra checks.

Where does it sit? 

If you place an agent somewhere outside the main bottleneck, it might work flawlessly but not actually increase overall throughput. It could just make one person’s job easier while shifting the backlog further down the line. Yet, most programs still choose where to deploy based on excitement and volume, rather than real impact.

When it comes to system architecture, I’d recommend keeping your AI models easily swappable. This part of the stack is evolving faster than anything else, so what seems like the best option today might be outdated in half a year. But there’s a catch: swapping models also shifts your quality baseline. That’s why it’s important to make sure your quality checks sit outside the model itself. Otherwise, you won’t really know if a change made things better or worse.

Q4. Looking ahead 3 to 5 years, which enterprise AI applications or capabilities do you believe are most likely to scale significantly, and what evidence are you seeing today?

Let’s start with the less glamorous side of agentic AI: ticket triage, claims handling, dispute resolution, procurement, and first-line support.

These are the areas that scale up first, mostly because they’re already set up for success. The decision rules are clearly documented, the data is all in one place and ready to use, and the tasks repeat often enough that you can see results in just a few weeks. Plus, it’s easy to spot good output from bad without having to redo the whole job. The heavy lifting—getting the data and processes in shape—was taken care of a long time ago.

But the real number to watch isn’t how many agents have been deployed—it’s how much work the organization is willing to accept without a human double-checking it. That’s where you see real trust in the system, and it tends to stay low in places where mistakes are costly. Counting deployments just tells you how much risk you’re taking on; tracking accepted output shows the actual value being delivered.

Beneath the surface, there’s another layer that I expect will quietly scale up: the “plumbing” of agent management—things like agent identity, permissions, evaluation, and monitoring. For most organizations, the challenge isn’t about whether agents can do the work anymore; it’s that no one really knows how many agents are out there or who’s responsible for them. That’s a problem that will get fixed simply because it has to be.

Looking further ahead, the bigger opportunity is with physical AI in manufacturing. Heavy industry usually ranks low on adoption charts, which might seem like a weakness—until you dig into the reasons. The real hurdles are integrating operational technology (OT) with IT systems and improving data maturity, not the AI models themselves. Right now, that foundational infrastructure is being built. Once it’s in place, we’ll see agent layers emerge around physical operations: things like operator assistants, grade changes, maintenance, safety, and compliance. The reality is that we’ll see autonomy on specific units, with people moving more into oversight roles, rather than fully autonomous sites.

Q5. If you had to give industry leaders and investors one lesson from what you have actually seen deploying AI in enterprises, what would it be, and what have you seen that makes you believe it?

Figure out how you’ll use the time you free up—before you roll out any new tool. If no one is clearly responsible for this, you won’t see the benefits show up in the numbers.

The time savings are real—people do get work done faster, and it’s not just their imagination. But those savings come in small doses, scattered across lots of people, and they never really make it into the official numbers. Instead, the extra minutes just get absorbed: workloads grow to fill the gap, expectations go up, and a year later, the balance between work and resources looks just like it did before.

To actually put those freed-up minutes to use, you need something concrete to channel them into—like tackling a backlog, reaching a service level you couldn’t hit before, improving coverage, or reducing a cycle time that matters to customers. If you name a specific goal, you turn scattered minutes into a meaningful metric someone can own. Saying “we’ll find uses for it” isn’t a real plan—that’s how any gains quietly disappear.

What really strikes me is how often this step gets skipped. It’s not a technical problem, so it falls outside the delivery team’s responsibility—and they can’t answer it for the whole organization anyway. Meanwhile, business cases are usually built by multiplying time savings per person by headcount, which ends up producing a number that never actually shows up in the accounts.

If you’re evaluating a company on this front, ask them what the freed-up capacity is supposed to achieve—and who’s in charge of making it happen.