At 6am, the AI hands the plant supervisor a new schedule for today’s production which cuts changeovers, and on paper it’s great. But the second job needs a certified operator who’s not on shift, the material for the third has just failed inspection, and the last job is for a customer who was let down last week – their order needs to be shipped first.
As a result, the Plant Supervisor can see that whilst the schedule is mathematically optimal, they can’t actually use it.
This gap between how an employee understands the context of an operation and where AI falls short is something I often hear when I talk to manufacturers about their use of AI. It’s backed up by Infor’s recent Enterprise AI Adoption Index, which finds that nearly three-quarters of industrial manufacturers (73%) believe off-the-shelf AI doesn’t adequately address the requirements of their industry.
You might expect this to put people off, and in my conversations, as many as half of the manufacturers I speak to haven’t started their AI initiatives yet. But I don’t see this as resistance. Our research shows that more than half of manufacturers (54%) plan to increase their AI investment in 2027. They know what AI can do for their speed of decision-making and execution. So what’s holding them back, and what are the leaders who are making progress doing differently?
In my conversations, four habits stand out. Firstly, the customers who I see making progress start by identifying a high-value operational problem and defining the desired outcome. Only then do they decide what kind of AI is appropriate. In practice, the answer is typically a combination of technologies, including machine learning, optimisation, and large language models (LLMs), rather than any single tool. Generic AI can provide support for activities where context is less important, whilst industry-specific AI solutions can be implemented to manage processes that require deep industry knowledge.
Success is measured against a business outcome – such as less downtime, faster quote turnaround, improved quality or reduced waste – and not against whether an organisation has deployed AI.
Secondly, they treat data as the route to context. Manufacturers have a unique strength compared to many other industries. They collect vast amounts of data from equipment, production systems and operations, and 70% of manufacturers believe their data is mature enough to support reliable AI outcomes.
Unfortunately, the data often stays locked in a reservoir. Think back to our Plant Supervisor: the schedule failed because the AI couldn’t see operator certifications, inspection results or customer history. Industry-specific AI can make better use of manufacturing data because it understands the relationships between equipment, production orders, inventory, quality and supply-chain processes. That’s what allows it to detect faults early and even predict defects.
The next thing successful customers do is to keep people in the loop. Manufacturers typically operate in environments where mistakes can affect productivity, quality, compliance, or safety. To make sure AI adoption sticks, they not only need to trust the decisions and recommendations being made, but also understand how they are made.
Successful organisations may appoint a named owner who is responsible for AI’s decisions and actions, and that person has the knowledge and authority to make responsible choices based on what the AI is telling them. An ideal process is one where AI recommends – human approves – AI executes – human audits. In effect, it’s the Plant Supervisor’s judgement, built into the process.
CFinally, they build governance in from day one. Manufacturing has one of the larger AI governance gaps, with 10.5% reporting no clear AI owner. Customers have shared with me that they value applying governance alongside the initiative, rather than afterwards.
As a result, I often talk about proportionate governance, where each use-case is risk assessed and the level of governance is matched to the risk level should an AI recommendation be wrong. If the impact of a bad AI recommendation is user frustration or a small budget overrun, then a lower level of governance is deployed. But if the risks extend to customers or data confidentiality, then more rigorous testing, auditability and human accountability are needed.
If you’re not there yet, you can begin to emulate these customers by choosing an initial use case focused on a genuine business challenge, and supported by data that is accessible, trusted and preferably drawn from limited sources that you can unpick. You might be surprised by how quickly you see results too, as many AI projects meaningfully deliver results in around 90 days.
If you start with the problem and give AI the context it needs, it will begin to understand your business as well as your people do.
Author: Andrew Kinder, SVP, Manufacturing Industry Strategy, Infor
For more articles like this, visit our Industrial Data & AI channel.
