There’s a conversation we’ve had with operations leaders in manufacturing and distribution. It tends to go like this.
They tell us they know AI is significant. They’re seeing it everywhere: at conferences, in the trade press, in their competitors’ announcements. They’ve had conversations with their ERP vendor about it. They may have even run a pilot or two. Their question: how should we be using AI?
It’s a good question. It comes from exactly the right place: a smart operator who can see that something significant is happening and wants to be deliberate about how they approach it rather than reactive.
And yet: we believe it’s the wrong one.
The gap between what you expect and what actually matters
A recent CohnReznick study found that about three-quarters of manufacturers are already investing in and applying AI. When asked what they’re most worried about, the answers are predictable and sensible: cybersecurity and privacy (60%), cost (46%), employee concerns about AI.
These are reasonable concerns. If you’re about to bring a new capability into your operations, you should be thinking about security, cost, and how your people will respond.
What’s interesting is what happens when those same companies actually implement. The challenges that turn out to matter most aren’t the ones they anticipated. The Manufacturing Leadership Council found that the single biggest challenge manufacturers face in AI and digital adoption is “developing a cohesive strategy,” cited by 42% of respondents. Not the technology. Not the budget. The clarity about what they’re actually trying to do.
MIT Sloan’s research on AI adoption in manufacturing firms found something similar. On average, companies see a short-term decline in productivity after they begin using AI, before experiencing longer-term gains. The researchers link this pattern specifically to the time required for process redesign, organisational restructuring, and workforce learning, not the technology itself.
Companies go in worrying about security and cost. They come out having discovered that their real constraint was knowing clearly enough what they wanted to change, and being willing to change the processes around the technology.
That’s the gap between the question most companies are asking and the question that actually determines whether AI produces competitive advantage.
What the right question is
Here’s how we’d rephrase it.
Instead of “how should we be using AI?”, ask: “Where in our operation does the improvement constraint sit, and where can AI help us address it in a way that wasn’t possible before?”
The difference is bigger than it looks.
The first question starts with the technology and works backwards to a problem. You end up looking for places where AI can be applied and applying it there. The constraint your operation actually has may or may not be anywhere near those places. There’s no mechanism that routes AI toward the problems that most need solving.
The second question starts with the constraint and works forward to the tool. You identify where improvement creates the most competitive value. Then you ask: is there a way to address this constraint using AI that produces a materially better outcome than what was possible before?
The answer is sometimes yes and sometimes no. When it’s yes, the improvement is real and measurable: you designed the engagement around the problem rather than around the tool. When it’s no, you’ve avoided an expensive pilot that produces good-looking reports and no bottom-line movement.
What this looks like in practice
Most operations in manufacturing and distribution have a short list of where the real constraints sit. Unplanned downtime that disrupts production schedules. Inventory positions that are either too heavy or too lean at the wrong time. Quality escapes that get to the customer. Demand signals that don’t reach the production plan fast enough. These aren’t abstract problems; they show up in your margin and in your customer relationships.
Now ask, for each one: is there a way AI addresses this that wasn’t previously practical?
For unplanned downtime, the answer is often yes. For example: reading failure patterns in sensor data across dozens of machines simultaneously, at a granularity no maintenance team can match manually, and surfacing specific maintenance interventions before the failure: that’s something AI changes materially. The process redesign needed to capture the value isn’t complex: maintenance scheduling moves from a calendar to what the model recommends. But without that redesign, the model sits in a dashboard that nobody acts on.
For the quality escape problem, it depends entirely on where in the process the escape is occurring and what data exists around it. AI may be the right tool. It may not. The point is that you can’t know until you’ve defined the constraint precisely enough to ask whether AI helps.
This is what’s meant by “developing a cohesive strategy,” the challenge that 42% of manufacturers identify as their biggest barrier. It’s not a document. It’s a clear enough picture of where your operation most needs to improve that you can evaluate any technology or methodology against that specific problem.
The practical version
When a manufacturer asks us how they should be using AI, here’s what we ask them back.
Pick the three operational outcomes that would create the most competitive value if you could move them meaningfully. Not “we want to be more efficient.” Something specific. Cycle time on the line. Inventory days. Customer complaint rate. Equipment availability.
For each one: what’s currently stopping you from moving it? Is it a data problem, a process problem, a skills problem, a decision-making problem?
For each constraint: is there a role for AI that addresses it in a way that wasn’t previously available? Not marginally better than what you’re doing now. Materially better.
That exercise surfaces two or three AI opportunities worth pursuing seriously, rather than a list of everywhere AI could be applied. The ones it surfaces will be the ones where, when you redesign the process around what AI makes possible, the constraint actually moves.
The manufacturers who get this right don’t have better access to AI tools than the ones who don’t. They have a clearer picture of where the constraint is. Everything else follows from that.
If you’re working through where AI addresses a specific constraint in your operation, that’s exactly the kind of conversation we’re built for.