Industrial AI and the Orchestration Problem

IDC · September 24, 2026

A couple of months ago I was at Siemens Realize LIVE in Amsterdam, and over dinner at the media and analyst reception, a conversation with Tony Hemmelgarn, CEO of Siemens Digital Industries, took a turn I did not expect. From industrial AI to why a Delhi intersection, and a UK roundabout arrive at the same […] The post Industrial AI and the Orchestration Problem appeared first on IDC .

Somewhere in there the UK driving exam came up too (that thing is brutal!). Tony had lived in the UK for a while himself, so he knew exactly what that experience is like. Stay with me, it made more sense than it sounds.

I grew up in India and have lived in the UK for several years now, which means I have spent enough time driving in both places to appreciate just how different they really are. I also happened to be back home in Delhi this summer for the holidays.

If you have driven in Delhi, you know exactly what I mean. There is a kind of ‘beautiful chaos’ to it. Cars, bikes, buses, auto-rickshaws, pedestrians, delivery guys on scooters, all making their own decisions at the same time, somehow without a plan. There are traffic lights and lane markings, sure, but anyone who has actually spent time on those roads knows the rulebook is only half the story. You are reading the situation constantly. A gap opens, someone takes it. Someone else has already anticipated that. A scooter appears out of nowhere on your left. You adjust. And somehow, everyone just keeps moving.

Now put that next to driving in the UK. Almost comically different. Everything is explicit here. Lane discipline. Signalling. Right of way. Mirrors, always the mirrors. Anyone who has sat the UK driving test knows that knowing how to drive and proving you can drive by the book are two very different things. It took me more than one test (meh!) to pass the exam. Unlearning Delhi’s driving and then learning the UK’s is not a joke.

The UK runs on predictability and rules. Delhi has rules too, but what really keeps things moving is interpretation, anticipation, and a kind of collective improvisation. Being neck deep in all things ‘industrial AI’, I couldn’t stop seeing the parallel, because the real question underneath both is who’s in charge when the rulebook and the improvisation are happening at the same time. That’s basically the orchestration problem, just with more scooters!

My last blog was about why the context layer is becoming such a big battleground in industrial AI. The next part to this is what happens after a system actually has that context. Knowing what is going on is one thing. Deciding what to do about it, especially when nobody has seen the situation before, is a different problem entirely.

For decades, industrial automation has been really good at that second part, as long as we can write the rules down. If X happens, do Y. Temperature crosses a threshold, trigger the alert. Machine stops, stop the line. Inventory drops below a level, reorder. That is honestly why automation has worked so well for so long. We take messy physical processes and turn them into decisions that repeat the same way every time. Nothing wrong with that at all, when the world is predictable, rules are genuinely powerful.

A machine starts behaving differently because the material batch changed. A supplier is late. Energy prices spike overnight. A production order gets rewritten at the last minute. An operator does something slightly different because they know something the system was never told. Or, my personal favourite, three unrelated things go wrong at exactly the same moment, and now it is not one system’s problem to solve. It is a question of who is in charge, and who decides.

Usually, the problem is not that the system does not have enough data. It might be drowning in it. The problem is that nobody wrote this particular situation into the rulebook, because nobody thought to.

Scaling AI, automation, and digital transformation is now the single biggest business priority for manufacturers globally, ahead of cutting costs, ahead of innovation, ahead of supply chain resilience from IDC’s own Worldwide Manufacturing 2026 Survey. And yet only around one in five manufacturers have actually gotten AI to scaled production use across multiple workflows or sites. Most are still piloting, or stuck running it in a limited corner somewhere. That gap between what people say they want and what is actually running on the floor is, in my mind, that same problem playing out at scale.

We tend to talk about AI in manufacturing through use cases. Predictive maintenance. Quality inspection. Scheduling. Forecasting. All fine, all useful, but underneath all of that sits a much simpler question: can a system make a good call when the situation was never anticipated in the first place? That, to me, is really the line between automation and autonomy. Automation runs a known response. Autonomy has to read a situation, weigh a few options, and decide what happens next.

This is where the traffic thing earns its place, I think. Delhi has not beaten the UK at this. The UK has not beaten Delhi either. Both systems genuinely work, they just carry complexity differently. The UK gives industrial AI something it cannot do without: rules, constraints, predictability. You do not want an agent treating a safety requirement as optional, or skipping a maintenance step because it spotted a quicker path. Delhi gives you the other half of the lesson, which is that the real world does not always follow the plan, and when it does not, something needs to be able to interpret and adapt in the moment.

Traditional automation asks what should happen when a condition occurs. AI lets us ask something harder: given everything happening right now, what should happen? The first question can be programmed in an afternoon. The second needs judgement, including knowing when to ask a human. It also needs an understanding that the locally smart move can quietly create a bottleneck or a maintenance headache somewhere else entirely.

The decision stops being about one machine or one KPI. It becomes about the whole system, which is probably why I think the next real fight after the context layer is orchestration, agents working across ERP, MES, EAM, supply chain and energy together, not just sitting inside one application doing their own thing.

So maybe the future of industrial AI needs a bit of both: the UK’s rules and Delhi’s improvisation, with AI doing the very unglamorous job of figuring out which one a moment is asking for, across every system involved, not just the one closest to the problem.

Underneath all of this sits a broader challenge: how intelligence works across systems, processes, people, and increasingly other agents. That’s orchestration. And as industrial AI moves from individual use cases into day-to-day operations, it is becoming harder to ignore.

For manufacturers, I think this changes the discussion slightly. Most organisations already have plenty of AI ideas, pilots, and use cases. What feels less clear is what happens when those systems need to work together. How does an AI-driven decision in one part of the operation affect everything else that sits downstream? Who is coordinating across ERP, MES, EAM, supply chain, and energy systems? And when things do not go according to plan, how does the system decide whether to act on its own or pull a human into the loop?

Those feel like the more interesting questions because they are not really about AI models at all. They are about how intelligence, whether human or machine, is orchestrated across the operation. And that may be one reason why so many organisations still struggle to move from promising pilots to something that operates at scale.

Contact our experts to explore how industrial AI orchestration applies to your manufacturing strategy.

The post Industrial AI and the Orchestration Problem appeared first on IDC .

Источник: IDC