This is the 2nd post in the AI Value Catalyst series. The previous post can be found here.
Every day, every business in the world is likely wasting significant amounts of money due to suboptimal and inefficient decision making. In B2B, that includes the customer’s business.
In the first article, we looked at the big picture of how to generate value from AI by optimizing decision making. Rather than starting with data or algorithms we should start where the business decisions are being made and then work backwards through the steps of simulation, understanding, and observation. That allows us to create value from AI more reliably and with greater ROI.
In this article, we go into the details of that first step.
One can think of businesses as fundamentally “decision machines”. Our business model determines which decisions will be made, on a daily basis, to generate value and thus profit. Optimizing decision making therefore – and obviously – optimizes our business. From customer service through marketing campaigns and all the way to strategy, a myriad of decisions are made every day. Although AI is transforming this, humans still make most decisions in most companies. This takes time and effort, and mistakes and inefficiencies are inevitable.
AI can enable people to make better decisions faster. And keeping humans in the loop brings the power of human intuition to bear. In some cases people and AI working together yields the best results. In almost all cases, AI can improve decision making.
Better decisions through AI
Jeff Bezos once called the internet “a thin enabling layer across all industries.” Well, we can think of AI as a strategic capability that allows us to optimize decisions across our entire business, no matter which industry we are in. That is a very powerful concept!
By searching for the right decisions to optimize with AI we minimize the risk of investing in AI projects that end up producing no real business value.
You can even apply this logic to the decisions that your customers or users are making.
If, however, you have limited experience with AI projects, I recommend that you start with the internal decisions of your own business. This will make it easier to work with users – after all, they are your staff – and obtain the necessary data, both crucial steps further down the line.
Focusing on external users can be more impactful; that is how we can create new revenue streams with AI. But it is also harder to do. To keep it simple we will focus on internal decisions – but the principles would be the same when applied to our customer’s decisions.
So, we want to search for business decisions that are based on inaccurate predictions or that are difficult to make presently. How do we figure out which actual decisions to target?
An AI catalyst decision map
To answer this question we will create an AI catalyst decision map. (It’s really a glorified table but hey – “map” sounds better, right?)
We will examine our business through the decision machine lense, searching for decisions that are good choices for AI optimization.
First, we need some method of iteration or enumeration. There are many ways of going about this. A simple approach is to go function by function: operations, sales, customer service, supply chain, and so on.
If we already employ tools to manage standard operating procedures we just iterate through the processes already captured. That makes it easy to enumere the key decisions that form the backbone of each business process.
We need to involve staff who actually deal with making these decisions and together identify the strongest candidate decisions to optimize with AI.
Consider, for example, a wind power company operating several large off-shore wind turbine farms.
Say we start with operations. Here we enumerate three key operational decisions:
- When should a particular drivetrain bearing be replaced?
- How far apart should turbines be placed in new wind farms?
- Which turbines should we purchase for a new wind farm?
As you can see these range from the tactical to the strategic level. We should not restrict ourselves at this stage to a particular level.
Which one of these decision candidates should we start with?
How to discover the potential value of each decision
The potential value that can be obtained through optimizing a decision breaks down into two components:
- The decision cost
- Savings potential.
The decision cost is simply how much it presently costs to make the decision. That can include salaries for human decision makers, subscriptions to software tools, and so on.
The savings potential is calculated as the present operational cost minus the operational cost we could reach if every decision was made perfectly.
Let’s look at the first decision candidate in more detail.
A predictive maintenance example
Replacing the drivetrain bearings is expensive, requiring manual labor and downtime. To make things worse, it is very hard to predict when they are close to being worn out and need replacement. If the bearings are not replaced before they break down the turbine may suffer serious damage and the entire plant will be forced to be shut down for a long time.
The decision to replace a bearing at any given time is made by a human engineer. But he can’t see the true state of the bearing so he follows the manufacturer’s prescription – namely to replace the bearing after a certain number of operating hours.
This policy steers us clear of the worst-case breakdown scenario – but it also causes a lot of unneeded downtime and waste of time and cash as bearings are replaced that are not actually worn.
How much money could our business save if the decision to replace was always correct, i.e. only replacing worn bearings and detecting wear correctly? In other words, how much could we save by advancing from preventive maintenance to predictive maintenance?
The total staff cost for these procedures we estimate at roughly $1,000 per month. This is the decision cost. (The engineers use a piece of software to manage maintenance but it is also used for other things so we don’t count that cost.)
In this case, a number of bearings are discarded every month that did not have to be. Replacing them also costs manual labor and turbine downtime, i.e. opportunity cost. In total, this amounts to around $14,000 per month. There have been no catastrophic breakdowns as far as the logs can tell so there we’ll add no cost for that.
Since the breakdown of a bearing is not a smooth, linear process it isn’t possible to estimate exactly how much longer the replaced bearings would have lasted. After discussing this with our most experienced engineer we agree that it is reasonable to assume that bearings that clearly had no wear would have lasted at least one additional month.
Factoring in the corresponding reduction in labor and downtime we estimate that the minimal parts maintenance cost would be $8,000 per month.
Thus there are potential savings of $6,000 per month “hiding” in this decision. Added to the decision cost of $1,000, we end up with a potential value of $7,000 per month or $84,000 per year.
You do the same for the two other decisions, but they come up with smaller potential values – see the table.
It is sensible to start with quick and dirty estimates of potential value, as long as they are likely to be unbiased. The most important thing is ranking the decisions approximately correctly.
Note that we won’t waste time and effort on the bottom of the table. This is one way this approach is already increasing ROI and reducing the risk of our investment in AI.
Next step: simulation
Having identified the most promising decision target we next go about working out what kind of simulations we will need to actually optimize each of the top decision targets.
For our top candidate we will need to predict the likelihood of each bearing breaking down given its operational history. Our simulator must also be able to put some time frame on these predictions.
That is the subject of the next article in this series. Stay tuned!
In the mean time – if you’d like my help with leveraging AI strategically for your business – and your career – contact me.