Scaling AI Agents Is Less About Building More Agents and More About Designing the System Around Them

Breaking the Bottleneck is a weekly newsletter and interview series covering manufacturing technology and physical AI. Want to chat? Reach out at aditya@machinafactory.org or connect with me on LinkedIn.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

"If it lives only in a dashboard, it won't move the needle. It has to show up inside maintenance, scheduling, or quality workflows where decisions get made every day."

You went from programming systems on factory floors to P&L accountability at Honeywell, then led global business units at Rockwell. Now you’re at a hyperscaler selling into the same customers you used to serve. How does sitting on the Microsoft side change the way you think about what manufacturers actually need versus what they’ll buy?

What changed for me isn’t the definition of “need,” but the scale of what’s possible when you can connect data, AI, and workflows across the entire value chain. On factory floors, “need” is still the same: safer operations, better quality, more throughput, and lower cost. If a solution can’t move one of those needles, it’s noise. Now sitting on the Microsoft side, I’m even more passionate about practicality.

Manufacturers don’t buy AI to check a box; they invest in outcomes they can measure in months, not years. And it has to show up in the way people actually work: the maintenance workflow, the scheduler’s daily decisions, the quality engineer’s root-cause loop—not another dashboard. What excites me now is being able to bring that together on a single platform, connecting data, AI, and workflows across the entire value chain in a way that wasn’t possible in fragmented systems.

Microsoft has deployed over 25 AI agents inside its own logistics operations, with a target of more than 100 by year-end. What has that internal experiment taught you about scaling agents that you wouldn’t have learned from customer deployments alone?

The biggest lesson is that scaling agents is less about building more agents and more about designing the system around them. In real operations, speed matters, but the ability to coordinate actions across workflows matters more.

Internally, you see quickly where agents deliver value today: monitoring inventory signals, tracking supplier commitments, flagging issues early, and triggering actions, not just insights. That shift from recommendation to execution is critical. Especially in supply chains where conditions are constantly changing, supplier delays, demand shifts, and logistics constraints, agents help teams respond in real time, not just react after the fact.

You also learn the harder part: how agents work together, how they integrate into existing processes, and where to lean on human intervention. Scaling isn’t just about autonomy; it’s about orchestration. If people don’t have clarity on “what should happen next and why,” you won’t build trust or adoption at scale. That is where a unified platform becomes critical, so agents are not isolated, but operating on shared data, aligned to the same workflows, and able to drive coordinated action across planning, supply, and production.

You’ve set a six-month goal for delivering measurable AI impact. What is the most common reason you see industrial AI projects fail to hit that mark, and at what point should a leader kill a stalled pilot rather than pivot it?

The most common failure indicator is trying to do too much at once with no clear measurement tied to throughput, quality, safety, or working capital. AI doesn’t fail because the model isn’t clever; it fails because it’s not embedded into the daily operating rhythm. What we are really talking about is a shift in operating model, from periodic, human-driven decisions to continuous, AI-assisted decision loops embedded directly into those workflows.

If it lives only in a dashboard, it won’t move the needle. It has to show up inside maintenance, scheduling, or quality workflows where decisions get made every day. As for “kill vs. pivot”: I’d give any pilot a clear checkpoint at 6–8 weeks. If you can’t get reliable data, real user engagement, and a leading indicator that the outcome is achievable, you don’t have a pilot, you have an experiment. Pivot if the outcome is right, but the workflow, data source, or scope is wrong. Move on if you can’t tie it to a real operational constraint and an owner who will run it after launch.

You frame AI as a “workforce multiplier,” not a replacement. But sentiment around AI is extremely negative right now, especially here in the United States, and frontline workers hear “autonomous operations” and immediately think layoffs. How do you address that cultural resistance in practice and get them to buy in, particularly when they’re the ones providing the critical context that agents on the factory floor need to function?

You earn buy-in by starting with respect for the people doing the work. I’ve stood next to technicians and operators who don’t want another screen, they want clarity. So we position AI as reducing friction: fewer fire drills, fewer surprises, faster root cause, and more time spent on high-skill work that only humans can do. Practically, that means involving frontline experts early, designing the system to explain “what to do next and why,” and putting clear guardrails in place so people understand where autonomy ends and human judgment begins. Often, the hardest part isn’t the technology, it’s change management. If leaders want trust, they need to pair AI investment with upskilling and a consistent vision, and they need to measure success in operational terms that matter to the team. That only works if the tools show up where people already work, in their daily workflows, not as separate systems they have to learn.

You talk about connecting every stage from design and simulation to production feedback, so manufacturers can test decisions before disruptions cascade. The “digital thread” is a vision the industry has been chasing for a decade. What is actually different now? Is it the AI layer, the data infrastructure, or something else that makes this achievable today when it wasn’t three years ago?

What we are describing is industrial intelligence, the ability to continuously connect data, AI, and workflows across the value chain and act on it in real time. What’s different now is the convergence of three things. First, the data foundation is finally catching up; more plants can bring together production, quality, maintenance, and supply signals in a usable way. Second, AI has advanced from “reporting what happened” to understanding patterns and cause-and-effect across silos, so you can see what is happening, not just what should be happening. Third, planning is moving closer to execution. Instead of static forecasts, teams can run frequent what-if scenarios using live production data, supplier updates, and demand shifts, so decisions get tested before disruptions cascade. So yes, the AI layer matters, but only because it turns the digital thread from a vision into something you can operationalize day to day. It is the combination of data platform, AI, and the applications people use every day that makes this practical, not just theoretical.

This newsletter is brought to you with support from our partners: Upkeep, Jiga, Industry 4.0 Club, Ironloop, and T51.