Most manufacturers are familiar with terms such as machine learning and generative AI, and may well be deploying them within their organisations. However, an emerging AI technology has the potential to take artificial intelligence to a new level within the space. Tim Gaus, Smart Manufacturing Business Leader at Deloitte Consulting, puts forward the case for agentic AI.
To date, most manufacturers have leveraged AI as an order-taker or manual-analyst on factory floors – but agentic AI can be more. Agentic AI has the potential to be the full factory management solution of the future that companies need in an era where there isn’t enough human talent to get the job done.
By making strategic decisions, solving problems and keeping production running smoothly, it could be a major improvement from the basic AI programming we see today. And according to Deloitte’s latest State of Generative AI in the Enterprise report, half of leaders expressed interest in pursuing agentic AI (52%) and multi-agent systems (45%).
As manufacturing facilities become smarter and software-defined, AI agents will not only monitor inventory and identify machine issues to notify staff, but also automatically place orders, negotiate with suppliers and immediately reroute workflows to other machinery as needed to keep production on course.
AI use cases within manufacturing
As we all know, AI deployment within manufacturing is nothing new. Classic AI and machine learning has been used for many years. However, the pace of development has seen the technology progress to deep learning, which enhanced the potential applications of AI.
And, in the last two years or so we’ve seen the rise in generative AI which, as the name suggests, generates net new content based on large language models (LLMs) that have been trained to ingest data, summarise it and create ways of streamlining processes that didn’t exist before.
However, as Tim explained, there are limitations with generative AI as the technology is incapable of solving problems on its own; that’s where agentic AI comes in. “Agentic AI has the agency to actually go out and proactively solve and find problems that have not yet been identified and bring them to the forefront of industrial processes to find a solution.
“That doesn’t mean agentic AI implies humans are taken out of the loop, but it can play a different role than just content creation – it’s actually out there solving problems.”
Evolution of AI within manufacturing
We’ve all heard the term ‘rubbish in, rubbish out’ when it comes to data, and for that reason, the deployment of AI within the manufacturing space is something of a mixed bag. Many manufacturers have found specific pockets of value for AI, but because of the dependence on the quality and the management of incoming data, those manufacturers who have more of a disordered data legacy have struggled to scale AI effectively.
More modern techniques have emerged to manage that legacy complexity and are allowing a rebirth of truly impactful AI solutions. “This has seen generative AI receive something of an uptick because it’s a little easier to deploy,” Tim added.
“Generative AI has many use cases and allows manufacturers to upload maintenance manuals and SOPs for example, and get to outcomes pretty quickly. However, many businesses have found that generative AI, though interesting, hasn’t quite got to the value creation they expected in many scenarios.”
Tim went on to explain that, when plotted against a traditional hype curve, generative AI is currently at the trough of disillusionment stage. In turn this is contributing to the excitement around the concept of agentic AI, which essentially features the ability to use the structure of data to achieve better outcomes, but in a more autonomous and independent way.
He added: “As we all know, agentic AI is a very new term. It’s still in its infancy so we’re not seeing a broad spread currently, but I do expect to see the adoption curve increase fairly quickly.”
Key benefits of agentic AI
When splitting out the key advantages of agentic AI in comparison to generative AI, Tim explained that it boils down to a route creation of content, versus actually addressing problems. An agentic AI solution will go out and proactively look for anomalies and come back with resolution recommendations.
While a generative AI solution may merely offer a summary of how to resolve a maintenance issue, for example, agentic AI will send a notification if a parameter is out of condition, recommend fixes and adjustments to resolve the problem and offer an explanation of how to do it correctly. Agentic AI moves beyond just responding to providing actual practical insights and information.
This is presenting a different challenge, however, when it comes to multi-agent versus single-agent problem solving. In single-agent systems, only one agent is involved in solving the problem. The agent perceives its environment, makes decisions and takes actions to achieve its goal.
In a multi-agent system, however, various agents interact with each other and possibly with the environment in the process of solving a problem. These agents can either cooperate, compete, or be independent in their actions.
Most businesses will have started out with single-agent systems, particularly those that navigated to the use of generative AI very quickly. With single-agent systems, you essentially have one solution that is trying to solve a complex set of problems. As an example, a manufacturer may ask a single-agent system if a machine is out of condition. If so, the system will then be able to provide information around where it is likely to fail next, the likelihood causes based on order history, the last technician who resolved the same issue and the work they undertook.
“From this example, it’s easy to see how you can cascade through a variety of different questions with a single-agent”, added Tim. “That is great, but it struggles to scale. What we’ve seen is that the complexity of getting a single-agent to tackle the breadth of challenges within a modern manufacturing facility, means that single-agent AI can quickly become overwhelmed, and there is a real risk of the system having hallucinations.”
As a consequence the manufacturing space is seeing increased levels of discussion around multi-agent AI, which allows businesses to really define a singular task for a given agent. Whether it be an MES or a historian agent, multi-agent AI systems mean that they are completely fit for purpose – to only look at one source of information and provide an answer around that single source. Backing that up, organisations can then deploy an orchestration agent that triggers the individual agents to answer questions, and then brings it all together to solve the problem.
“That’s the fundamental difference between single-agent versus multi-agent systems, and why agentic AI is going to move to multi-agent needle very quickly”, Tim added. Not only that but if a manufacturer is already deploying other methods of AI it is possible to leverage from past investments. If data is already coming in, Tim explained that it is relatively easy to assign an agent against it. Moreover, there can also be an interaction where the agents themselves can use the machine learning models to answer their own questions.
He continued: “This means that we’ll be able to create new digital workers or new digital continuous improvement engineers, that will avail themselves of former work but also complement it with elements that would be harder to tap into previously. It’s complimentary and an acceleration of past work.”
Filling critical workforce gaps
Agentic AI is also a way to fill gaps in the workforce, and make them stronger and more effective. Continuous improvement engineers will no longer have to deal with spreadsheets. Rather, they will be able to start with a prompt of what they should try and tackle, to the point where the agents will autonomously begin to get into their own ‘five whys’.
Similarly, a production supervisor will be able to walk into a shift meeting and instead of having the agenda of trying to find out what’s happened over the course of last shift, they will actually come in knowing what happened and with a proactive recommendation around the topic of conversation in terms of correct course of action.
“From time spent in the office and how to be more proactive, to troubleshooting and problem solving, agentic AI multi-agents will help address the critical skills gaps that manufacturers are experiencing, particularly at the supervisory level”, added Tim.
That’s said, Tim predicts that humans will continue to be in the loop for the foreseeable future. So even if a business has a great multi-agent, orchestrated, problem solving solution, it is still going to need to rely on a human to look at the outcome and ensure it makes sense.
Maximising implementation of agentic AI
As a new technology, Tim highlighted that it’s going to operate very much within a manufacturer’s OT systems layer. So the individuals who have traditionally maintained those systems are going to have to either learn new skills and ways of interacting with technology, or they will have to be augmented with others who understand how to tune LLMs or SLMs, as well as other elements that are required to make agentic AI successful. Whether that sits within the four walls of the plant or is a support structure outside, in all cases, it needs to exist.
Outside of that, Tim also highlighted several other factors within the skills space he sees as important when it comes to successful agentic AI deployment. He added: “We’ve oftentimes in the past, struggled with people adopting AI solutions because they don’t trust them. And agentic AI will be no different. Therefore, manufacturers will need a workforce that’s willing to embrace technology, find new ways to use it, and really try and push themselves to that next level of activity.
“On the flip side, one of the fears I hear from many of my clients is if the agentic AI agent gives a recommendation – and if we have lost our ability to evaluate it – will that then lead to humans not critically thinking about the actions they’re taking , rendering the technology to be less helpful?
“It’s vital therefore, that the workforce deployed to use these agents don’t just accept whatever is given as being the right answer, but still apply critical evaluation of what’s coming back to make sure it’s not giving you a hallucination or an answer that’s based on bad data.”
He added that enterprises that manufacturer goods are currently very aware of agentic AI at an enterprise level. However, it has yet to create true excitement on the shop floor. No doubt this is not helped by the fact that many have become more than a little overwhelmed with the speed of technology change. Generative AI use cases are still being pursued in several instances, and now all of sudden there’s a new AI kid on the block. That inevitably has created a layer of skepticism and weariness around the volume of new tech being thrown at the sector right now, which is in turn slowing down true adoption.
Barriers and top tips
We’re nevertheless at a very interesting moment in the realm of agentic AI. The importance of data quality when it comes to the impact AI can have is well-known, so if manufacturers have made past investments in rebuilding that foundation to ensure their data is clean and of high quality, they are going to be able to move quickly and extract a lot more value.
However, if manufacturers have not spent the time and effort to get at least some of their legacy systems in a better condition and more accessible, there will be a large foundational lift required to achieve the sought after, high-value use cases. “That being said, I always encourage my clients to not only focus on the foundational elements, because it’s easy to get trapped in spending a lot of money and time on things that are seen as important and foundational, but frankly, don’t drive ROI. That’s when companies can get exhausted with the spend and progress stops”, Tim added.
It is important not to fall in love with tech for its own sake, and to establish a true value with a desired outcome. “If you can’t achieve true ROI on any technology investment, your people aren’t going to like it. So, don’t make this harder on your people, make it easier – make it a win for them. If you don’t have that human element you’ll have a huge challenge in getting people to actually adopt these solutions.”
In terms of quick wins and low hanging fruit Tim highlighted that there are areas which manufacturers can turn to quickly, and even if high-quality data isn’t in fact available, value can still be extracted in a different way. Businesses may just have to think a little bit more broadly than they have in the past. “I’m actually quite bullish in how we’re going to get that first tranche of interesting questions and prompts brought to businesses and built up over time. For me, agentic AI is the classic case of start small and scale fast. Don’t try and go too big too soon, but definitely get it moving.
“The idea of Lean manufacturing and Six Sigma processes have always been core to manufacturing and I see agentic AI playing very strongly in that space. I can see a world in which a problem exists and the agent actually does the five whys, comes to a resolution and automatically takes action on that resolution itself.
This will lead to a much more self-healing manufacturing systems, with humans first in the loop who will eventually be informed as actions are being taken, without even having the confirmation. So it gets you to a pretty exciting space of how this particular technology could be quite transformative.
“Slightly further out, there’s a very interesting intersection of agentic AI in relation to the future of robotics. If the ability exists for systems to be more problem solving on their own, that opens up a whole new tranche of how agentic AI could be applied to robotics to actually operate our processes more independently. This will move industry closer to that goal of having less reliance on humans to do the manufacturing, though still having them drive the future of our manufacturing base.”
For more articles like this, visit our Industrial Data & AI channel.
