A paradigm shift in production with machine learning

Jan-Hendrik Heuing and I have spent the last few weeks discussing new possibilities for production control. This is relevant for us, as we are considering how we should organize production at Accounto in the future in order to achieve “extreme efficiency”. The result is what we call, not without a twinkle in our eyes, a hyper-frequency production planning engine. About the future use of artificial intelligence (AI) in production.

(Reading time: 5 minutes)

Paradigm shift 1: Switch from individual production to assembly line

When Henry Ford introduced assembly line production around 100 years ago, it was an extreme productivity gain. The concept is still used today and has been greatly accelerated by robots.

In principle, production is always about allocation problems: Is the material in the right place at the right time, is the employee with the right skills in the right place at the right time. Can the planned work be carried out in the planned time? What are procedures if these processes do not run as planned?

Sequential assembly line production made it possible to solve these allocation problems better than before, thereby increasing material throughput and production output.

Paradigm shift 2: The machine that builds the machine

A second paradigm shift, I believe, has taken place in the last 5-10 years. More and more people in production are viewing production itself as their product. This is a process that should not be underestimated in terms of importance.

This is because it directs engineering’s attention away from the product to be delivered and concentrates on the machine that manufactures the products.

I think this is a quiet paradigm shift. A rethink by many people involved in production. And it’s a logical step: the more work steps robots take over in production, the more people are involved in maintaining and optimizing the production machines. This rethink is a basic prerequisite for the third paradigm shift that I expect to see in the next few years: production organized/supported by AI.

Problems of sequential production

The problem with today’s sequential assembly line production is that the allocation of resources is not dynamic. There is always idle time and it is not possible to speed up the cycle, i.e. the time available to perform an operation, indefinitely. On the one hand, people working in sequential production cannot (and should not!) always work faster. On the other hand, material maturing processes, for example, require downtimes that must be approached with sufficient leeway because external influences also play a role.

In addition, the production of heterogeneous products in sequential series production systems is a complex matter. Sequential cycle-based production was/is a huge step forward compared to one-off production. But it is by no means perfect in terms of today’s possibilities.

Paradigm shift 3: AI-supported production

With AI-supported production, you can break up the sequential process and switch to a decentralized model, making much better use of resources.

Instead of carrying out work processes in a clear, predefined order, AI-supported production should perform the non-interdependent work steps in the best possible (i.e. logically correct) order depending on the situation.

At the heart of this production is this hyper-frequency production planning engine, which reads all production data in real time and controls the entire production process. You have to imagine it a bit like a game of chess: the engine continuously calculates x-possible scenarios and constantly decides on appropriately optimized work instructions for the production machine.

For example, in the automotive industry, door A of model Y may receive the trim first and door B of model Y may receive the locking mechanism first. As long as the work steps are not physically interdependent (e.g. the door must be painted first so that the trim can be applied), the engine decides what the most efficient route is based on capacity utilization and material supply, basically using all available data. And thus controls production completely autonomously.

Reinforcement

For such an engine to be possible, as much data from the production machine as possible must be available in real time. That is already quite a challenge. And I maintain that in the area of human labor, we don’t even have the basics in traditional manufacturing at the moment. And we can’t do it without human labor at the moment.

What will make such a production engine really hyper-efficient, however, is that it learns independently from the effects of its decisions.

And what does this have to do with automated accounting?

You may wonder why this is bothering us. Do we want to produce cars or other physical things now? No, of course not.

But we consider the fully automated creation of auditable books as a production process. And just like in assembly line production, today’s processes are sequential. A document comes in, is classified, undergoes various checks, is broken down into its accounting-relevant components, which are then classified according to a regulatory framework.

Here, too, it is the same as with physical production. Not all work steps necessarily have to be sequential. Depending on the input volume and availability of resources, different sequences make sense. And the more human labor is involved, the more these different sequences can make sense.

This fluid, ongoing planning of the production machine therefore has particular advantages when human labor is involved. However, it also has advantages when this human labor is completely replaced by machine labor. This is because machine processes also have different and changing resources available. At the end of such an automation process, in our case an AI-supported production planning engine would act as a kind of load balancer for computing power, because we no longer have any physical processes involved.

Workbenches

Now, of course, it would be extremely exciting if we were already there. But we are not yet. The Hyper-Frequency-Production-Planning-Engine is a much-discussed vision at the moment. One that we will be tackling in terms of code over the next few weeks.

The basis for this is the organization of production away from traditional (digital) accounting processes towards what we internally call workbenches. A workbench always comprises various related tasks, which can be performed either by bots or by employees. In our concept, the employees help the bots to learn to take on more and more work. Until all that remains for the humans is (fanatical :-)) customer service remains.

This reorganization is currently underway and is no easy matter. Thank goodness we are still so small. Above all, it is not easy because working in this way is not exactly intuitive. However, the fact that we are positioning ourselves accordingly is a basic prerequisite for ensuring that more and more tasks can be controlled by the future engine. It is the first step on the way to a new production paradigm. I am extremely excited to see how traditional production companies will make use of these new possibilities. First and foremost, and for good reason, the automotive industry, of course.

I think the fact that we have a completely digital production that is detached from physical processes will greatly simplify implementation. That doesn’t mean it will be easy. But if we had to connect millions of sensors, as in physical production, the tasks would be scary.

However, I am confident that we will have the first real-time, automatically planned work tasks for employees and machines ready in a few months’ time. And that we can then build on this.