The role of (artificial) intelligence in inventory optimization
What is artificial intelligence (AI) and how can it help with inventory optimization?
A working definition of AI might be anything that makes a computer appear intelligent. And relative to inventory optimization, we might define AI as the ability of machines to take inventory “decisions”, such as when and whether to place an order, and of how much.
Inventory optimization is a complex endeavour, as we have outlined elsewhere. As such, it is desirable to enlist the help of advanced technology to help with it. Several tools on the market promise inventory optimization, some of which use techniques that are described as AI.
One problem with a buzzword like AI is that it is used to mean different things. In this article we are not so much interested in a theoretical discussion of AI as in considering the potential of inventory optimization technology more generally, whether powered by something you consider to be AI or not. And how reliant such technology is on the old-fashioned kind of intelligence that you find displayed by your people.
AI is one of a number of concepts, like natural language processing (NLP), data lakes, low-code, machine learning, robotic process automation (RPA), and so on, that are making great strides each year. There are wonderful opportunities for businesses to take advantage of all this innovation, but in these times of flux there is also a risk of investing a lot of time, effort and resources into interfaces that are out of date within a year. We believe there is always value in experimentation, building a platform for the next generation, but not all organisations have the time and resources to invest in something without a clear and immediate business case.
So what does the supply chain professional need to know about AI and how can you take advantage of it in your inventory optimization efforts? This discussion is also relevant to other fields in supply chain and beyond.
A working definition of AI might be anything that makes a computer appear intelligent. And relative to inventory optimization, we might define AI as the ability of machines to take inventory “decisions”, such as when and whether to place an order, and of how much.
Inventory optimization is a complex endeavour, as we have outlined elsewhere. As such, it is desirable to enlist the help of advanced technology to help with it. Several tools on the market promise inventory optimization, some of which use techniques that are described as AI.
One problem with a buzzword like AI is that it is used to mean different things. In this article we are not so much interested in a theoretical discussion of AI as in considering the potential of inventory optimization technology more generally, whether powered by something you consider to be AI or not. And how reliant such technology is on the old-fashioned kind of intelligence that you find displayed by your people.
AI is one of a number of concepts, like natural language processing (NLP), data lakes, low-code, machine learning, robotic process automation (RPA), and so on, that are making great strides each year. There are wonderful opportunities for businesses to take advantage of all this innovation, but in these times of flux there is also a risk of investing a lot of time, effort and resources into interfaces that are out of date within a year. We believe there is always value in experimentation, building a platform for the next generation, but not all organisations have the time and resources to invest in something without a clear and immediate business case.
So what does the supply chain professional need to know about AI and how can you take advantage of it in your inventory optimization efforts? This discussion is also relevant to other fields in supply chain and beyond.
Finding the intelligence in artificial intelligence
Artificial intelligence is, above all, artificial. It does not work the same way as human intelligence. A computer can process data 24 hours a day, 7 days a week without tiredness or error. But that is also its weakness, since it doesn’t question what it is doing.
A computer does what it is told to do, no more no less, following an algorithm. But this also means that if the algorithm is not universally valid, even if it works flawlessly, then the answer may be incorrect. This will be familiar to anyone using planning technology that they know to be inaccurate for at least some of the items they make or stock.
Underlying any intelligent approach to inventory optimization, whether artificial or human, you find algorithms. They embody the intelligence itself.
A computer does what it is told to do, no more no less, following an algorithm. But this also means that if the algorithm is not universally valid, even if it works flawlessly, then the answer may be incorrect. This will be familiar to anyone using planning technology that they know to be inaccurate for at least some of the items they make or stock.
Underlying any intelligent approach to inventory optimization, whether artificial or human, you find algorithms. They embody the intelligence itself.
Inventory optimization algorithms
A simple algorithm might be used to automate replenishment: if stock on hand is equal to or less than X, place an order for quantity Y. This type of algorithm is widely used in various technology solutions. Its benefits are threefold:
- When automated, it is superior to a human because it can’t be ill, delayed, or simply forget
- Its automation frees a human up to work on more difficult topics
- It is easy for a human to trust, because the algorithm is simple to understand and transparent
This type of simple, rule-based algorithm lies at the heart of a lot of software including, where transfers of information between systems are required, RPA.
But how, in the example above, do you determine X (the re-order level) and Y (the order quantity)? This is not so simple. In this case you could either go for another rule-based algorithm, just with a lot more conditions and equations, or you could apply an optimization algorithm, that will work out the optimal solution itself.
This more complex type of algorithm could use deterministic and stochastic methods, which is essentially to say that it uses known mathematical formulas, factoring in variability. Or it could use techniques like genetic algorithms and data mining, which are useful when you are not entirely sure what you’re looking for or how to find it, and which are much more computer-intensive.
These more complex algorithms also have benefits:
But how, in the example above, do you determine X (the re-order level) and Y (the order quantity)? This is not so simple. In this case you could either go for another rule-based algorithm, just with a lot more conditions and equations, or you could apply an optimization algorithm, that will work out the optimal solution itself.
This more complex type of algorithm could use deterministic and stochastic methods, which is essentially to say that it uses known mathematical formulas, factoring in variability. Or it could use techniques like genetic algorithms and data mining, which are useful when you are not entirely sure what you’re looking for or how to find it, and which are much more computer-intensive.
These more complex algorithms also have benefits:
- They can handle calculations that few humans can do themselves, without error
- They allow business to take advantage of advanced scientific techniques
- They allow planners to get much closer to truly optimal levels of inventory
However, compared to the simple algorithm, these complex algorithms also have a number of weaknesses:
- Like any algorithm, they are only as good as the data fed into them. As the algorithms get more complex, relying on multiple input data sources, this is a problem, in particular where human planners don’t know what data are being used or how accurate they are
- Unlike a human planner, they are constrained in their access to information. Software can be connected to multiple sources of information, and great progress is being made in techniques like natural language processing (NLP) to enable better use of unstructured information, but ultimately it does not know everything your human planners know. AI’s models and constructs of the real world are in their infancy and still have immense difficulty outside their own comfort zone
- Despite their complexity, they inevitably still rely on a number of assumptions that will vary in accuracy by item and over time. But human planners usually don’t know what these assumptions are or how sensitive the algorithms are to them
- Because of their complexity, these algorithms are not readily understood by human planners. When you also factor in a lack of transparency as to how they work, it should not be surprising when human planners don’t trust them
Artificial intelligence and trust
Trust is the critical concept when it comes to getting value from AI. Let us consider three scenarios:
Scenario 1, a simple algorithm. The human planner understands how it works and has experience of it working well.
Scenario 2, a complex algorithm. The human planner does not understand how it works, but knows through repeated experience that it does.
Scenario 3, a complex algorithm. The human planner does not understand how it works, and knows that it works only sometimes, but can’t predict when it will work and when it won’t.
We can already see that in scenario 3 it will be hard to induce the human planner to use the technology. And even model 2 might have limited usefulness in periods of volatility and uncertainty. (Witness all of the advanced forecasting models quickly ditched when Covid struck in 2020.)
Within complex challenges like inventory optimization there is actually a tension between simplicity and accuracy. The more variables you factor in, the more complex the algorithms become, and so less comprehensible to humans. Whereas the simpler you keep the algorithms, the less accurate and reliable the results. A lot of technology in this space is in the unhappy situation that the algorithms are already too complex for most humans to follow (even if they were made transparent, which they are usually not), but the results delivered are frequently too approximate or conditional to be trusted and used blindly.
Scenario 1, a simple algorithm. The human planner understands how it works and has experience of it working well.
Scenario 2, a complex algorithm. The human planner does not understand how it works, but knows through repeated experience that it does.
Scenario 3, a complex algorithm. The human planner does not understand how it works, and knows that it works only sometimes, but can’t predict when it will work and when it won’t.
We can already see that in scenario 3 it will be hard to induce the human planner to use the technology. And even model 2 might have limited usefulness in periods of volatility and uncertainty. (Witness all of the advanced forecasting models quickly ditched when Covid struck in 2020.)
Within complex challenges like inventory optimization there is actually a tension between simplicity and accuracy. The more variables you factor in, the more complex the algorithms become, and so less comprehensible to humans. Whereas the simpler you keep the algorithms, the less accurate and reliable the results. A lot of technology in this space is in the unhappy situation that the algorithms are already too complex for most humans to follow (even if they were made transparent, which they are usually not), but the results delivered are frequently too approximate or conditional to be trusted and used blindly.
Taking advantage of technological innovation in inventory optimization
So what should you do? Exciting developments are ongoing and failure to engage with them leaves you at risk of falling behind. The fear of missing out is great.
On the other hand, the promise of fully autonomous supply chains, with all data and systems connected in smart networks that can manage themselves, is still a long way off. As different technologies compete with each other, there is also a risk of committing too fully to the wrong one.
The best approach comes down to finding the right balance between experiment and risk mitigation, while not neglecting the human factor:
On the other hand, the promise of fully autonomous supply chains, with all data and systems connected in smart networks that can manage themselves, is still a long way off. As different technologies compete with each other, there is also a risk of committing too fully to the wrong one.
The best approach comes down to finding the right balance between experiment and risk mitigation, while not neglecting the human factor:
- Invest as much in people as in technology. While it is unrealistic to expect every single human planner to understand your planning technology fully, it is important to develop and maintain a strong level of that understanding internally, and to have a proper strategy to ensure the whole organisation can take advantage of it. Don’t create silos between analytics and planning teams. Who are your internal inventory experts and how do they bridge the gap between analytics and planning?
- Maximise the value you are getting from your existing technology. While there are for sure clear differences between different planning tools, and you should not be afraid to upgrade, change or even just experiment, you may just be swapping one set of limitations for another. We encounter many companies with under- or unutilized planning technology who could get much more value from it if only they invested more effort in understanding how best to use it
- Try as many new approaches as you can afford, but with limited scope pilots. You have to devote enough time and resource to a new piece of technology to properly assess it, but on the other hand global top-down approaches that enforce standardisation are much longer and riskier than nimble pilot projects. This approach may seem anathema to those whose vision of the future is of a central team pulling all the levers in a single data ecosystem, but having several horses in the race increases your chances of backing a winner
- Be very clear on the business problems you are trying to overcome and the results you hope to achieve. While experimentation for its own sake can be good in the right circumstances, embarking on grand “digital transformation” projects without clearly stated objectives and measurable outcomes is a mistake. If you’re building data lakes and analytical models, implementing forecasting and planning tools with an expectation that service levels will improve and inventory levels reduce automatically as a consequence, prepare to be disappointed. Working out what is causing your inventory imbalances in the first place and then systematically addressing those root causes is a much more certain and efficient way both to deliver tangible business value and to identify the tools you will need for the job! Too much technology is bought based on the appeal of slick interfaces rather than a detailed bottom-up definition of functional requirements. Inventory challenges vary by industry and even by company and planning tools well adapted to, say, automotive manufacturers may well not work very well for apparel retailers
If you were hoping to read here that AI is able to take away the difficulty of achieving inventory optimization then, for now and for the foreseeable future, you will be disappointed. But it is not a question of all or nothing. Technology available now can help you achieve a lot.
In the end, what it comes down to is this: to take best advantage of all that AI has to offer today, you need the most powerfully intelligent machine known. The good news? Every one of your people has one between their ears.
In the end, what it comes down to is this: to take best advantage of all that AI has to offer today, you need the most powerfully intelligent machine known. The good news? Every one of your people has one between their ears.
Appendix - how nVentic uses (artificial) intelligence to support inventory optimization
The discussion above is not of purely theoretical interest to us at nVentic. We have built our own inventory analytics technology that we leverage to help companies deliver significant improvements to their inventory positions. So what do we do and how do we navigate this tension between complexity and trust? We can split this into what technology we use and how we then use that technology.
Technology
We have synthesized the best scientific work on inventory optimization into a series of algorithms that work on large business data sets to identify optimal inventory levels by item. We use a number of techniques commonly referred to as AI, in particular hill-climbing, but we avoid using the expression AI, which is applied to such a wide variety of techniques, from the simple to the very complex, that it risks being meaningless. “We use AI to do it” wouldn’t really tell you anything.
Our technology approach is very much based on the algorithms themselves, which are fully documented and reference the peer-reviewed scientific work underpinning them. We have coded these algorithms in C++, supporting multi-threaded calculations, so machine-near and fast but not hardware dependent. We have also automated a number of routines that produce the analytical outputs in a selection of formats for easy and flexible use by our clients.
Our technology is not a planning tool, but an inventory diagnostic. Because of the many issues with producing numbers that inventory planners can and should trust, we break the algorithms down into their main constituent elements, make these transparent, and then put a lot of effort into building understanding of the results among client teams. Because we work outside of the source data environment the operational data cannot be compromised.
Application
At least as important as our technology itself is what we do with it. Our strongly-held belief is that there is a significant skill gap preventing organisations from taking advantage of the best inventory optimization technology on the market, including our own, without significant investment in people. The input to our technology is data, the output is information, but to get real value from it you need insight and this is where the human brain needs to come back into the picture. Insight into understanding what the analytics means and insight into what you need to do to take advantage of it.
In other words, in order to take full advantage of the best in AI, you need plenty of I.
Technology
We have synthesized the best scientific work on inventory optimization into a series of algorithms that work on large business data sets to identify optimal inventory levels by item. We use a number of techniques commonly referred to as AI, in particular hill-climbing, but we avoid using the expression AI, which is applied to such a wide variety of techniques, from the simple to the very complex, that it risks being meaningless. “We use AI to do it” wouldn’t really tell you anything.
Our technology approach is very much based on the algorithms themselves, which are fully documented and reference the peer-reviewed scientific work underpinning them. We have coded these algorithms in C++, supporting multi-threaded calculations, so machine-near and fast but not hardware dependent. We have also automated a number of routines that produce the analytical outputs in a selection of formats for easy and flexible use by our clients.
Our technology is not a planning tool, but an inventory diagnostic. Because of the many issues with producing numbers that inventory planners can and should trust, we break the algorithms down into their main constituent elements, make these transparent, and then put a lot of effort into building understanding of the results among client teams. Because we work outside of the source data environment the operational data cannot be compromised.
Application
At least as important as our technology itself is what we do with it. Our strongly-held belief is that there is a significant skill gap preventing organisations from taking advantage of the best inventory optimization technology on the market, including our own, without significant investment in people. The input to our technology is data, the output is information, but to get real value from it you need insight and this is where the human brain needs to come back into the picture. Insight into understanding what the analytics means and insight into what you need to do to take advantage of it.
In other words, in order to take full advantage of the best in AI, you need plenty of I.
Would you like to receive more content like this, direct to your inbox? We publish white papers on a range of supply chain topics approximately once every one to two months. Subscribe below and we will notify you of new content. Unsubscribe at any time.
Would you like to talk to one of our experts? Contact us