AI Projects Fail for a Reason. How to Make Yours Succeed?
Markus Lippus | 19. June 2019
Don’t start from your data. It’s the problem that matters most.
“We have tons of data. How can we make use of it with AI?”
That’s the question an increasing number of companies are asking on a daily basis.
The craze for artificial intelligence is creating a sense of FOMO (fear of missing out) among organizations large and small. What if we all fail to make use of AI and are left behind the competition?
Certainly, there’s a fair share of truth in asking all these questions. However, there’s an important caveat.
You must approach your AI projects from the right angle.
How companies make use of AI
Working in a company specialized in helping others catch this wave, we’re often talking to executives and project managers that want to bring AI to their companies as well.
With some exceptions, these companies fall into two categories:
There are visionary companies that know what they want to do with AI and that have a fairly detailed action plan figured out.
However, often enough, we encounter companies with an action plan that looks something along these lines of this:
Apply machine learning
Now by no means is this 4-step action plan stupid or 100% wrong. You do need relevant data. Machine learning willprobably help to translate it into something useful. And of course, you’ll want to get some value out of the project.
There is only one thing wrong with this plan.
It’s missing some key elements.
After working on a fair share of data projects, we’ve spotted a couple of key things that make some projects succeed while others fizzle and die.
So what are the magical ingredients that will make your machine learning endeavours more likely to return a positive ROI?
Let’s start by understanding where companies, maybe including yours, stand today.
Most businesses, even small ones, generate vast amounts of data. This can be any kind of data. Think, for example, about any type of logs: usage logs, communication logs… Everyone has logs!
Now imagine you’re sitting on a pile of business data… The initial conclusion for most company leaders would be…
“We have tons of data — could we make something interesting out of it?”
Let’s call this the Data First approach.
The problem with the Data First approach
The plain reason people like to start AI projects from data is that this is what they have in their hands.
And because it’s really, really difficult for humans to make sense of large amounts of data using spreadsheets and SQL, we want to apply machine learning to make sense of all it all.
Wouldn’t it be nice to give all our data to the machines and watch them turn it into business ideas and profit?
But that’s usually not what happens.
What is more likely to happen is that a lot of smart people will form hypotheses and test them against their data. As a result, they may find a few patterns that have somerelevance to the business side of the company.
However, most of the Data First machine learning projects are far from optimal.
“AI projects fail because the questions companies ask are mostly wrong or irrelevant to the business.”
Add to that the costs of deploying the results and integrating them into your team’s workflow. At some point, someone will realize that the gains don’t cover the costs. Disillusionment in machine learning ensues and all further plans of riding the AI hype train screech to a halt.
Put simply, if you start your AI projects by asking questions that you think the AI can answer, you’re likely to…
ask questions that don’t need AI to solve them
spend a lot of time on solving low-impact problems
So what are you to do? Not apply machine learning to your business data?
That’s far from the optimal solution.
Enter the Problem First approach
There is an alternative way to apply machine learning to your business data. And we’ve seen far more companies succeed with this second solution.
This other option is called the Problem First approach.
Let’s imagine that you’re running a mature business that already has loads of data. And you want to find a good purpose for it.
Only this time, instead of starting to blindly search for patterns in this data, you’re going to talk to the people who are in charge of various processes in your company. Those people know the different sides and operations of your company the best and are able to bring out the issues with the largest impact.
“The people most fit to ask the right questions from AI are your team members.”
Ask these people to define, as specifically as possible, a number of problems they need solving. This could also be a process that needs to be improved.
For example, the problems solvable with AI could be:
What triggers the churn of our customers?
How to make our onboarding process more effective?
How to cut the costs in our production line?
How to avoid faulty product being shipped to customers?
“Only after you know the problems you need solving can you ask if the answer could be found in data.”
How to ask the right questions from AI?
Your employees have no previous experience in AI.
That’s why we should keep in mind that a large part of defining this set of problems rests on people not necessarily well-acquainted with the capabilities and limitations of machine learning.
It makes sense to give your team members involved in the ideation process at least some shorthand to decide whether a business problem can be automated using machine learning.
As a general rule, I find the following two questions to be useful for this purpose.
1. Could a person do it in less than a second? — credits: Andrew Ng
“Humans are great at pattern recognition. We see patterns everywhere. Even if there are none!”
If a pattern detection task is simple enough that a person can do it in less than a second, there’s a good chance you can teach a machine to do the same task with similar accuracy, but a lot faster.
These tasks include simple perception tasks anyone can do, like deciding if there’s a cat or a dog in an image. But they also include far more complex tasks like monitoring sensor data to decide if a component is soon in need of maintenance.
Some good and relatable examples of machine learning in this field are facial recognition and voice commands used on phones.
One common issue when automating human tasks with AI is that the more complicated the tasks get, the more training data you’re going to need. For example, building a machine that can recognize the shape of a human face is fairly simple.
However, building a machine that can accurately distinguish between a billion different faces is an endeavour so difficult that it’s unlikely to have a positive ROI.
That’s why you should always think how complex of a solution you really need.
There is also another question you can pose…
2. Is there a pattern here that I could see, if only I could fit it all on a spreadsheet?
“Do you think there’s sound logic to believe that the answer to your question can be found in the data you have?”
“The data concerns the usage patterns of my product, so it makes sense that the reason for churn could be deduced from this data.”
Making the right connections between your data and a problem can be difficult. After all, you need to make a judgment call, and your logic may turn out to be wrong.
So what are the right situations for using this approach?
Look for high-impact projects that are worth trying to solve — something that could potentially generate a lot of value to your company and is worth a small risk.
Prioritize your AI projects
After you’ve gathered a bunch of questions or problems to be solved, start prioritizing them into a list.
It is highly important to have both machine learning experts and your key employees involved in this process.
Prioritize your problems based on:
The potential value solving the problem would generate
The difficulty of the machine learning project
The urgency of solving each issue
And finally, there is one more important question to be asking:
“Is the machine learning project the optimal solution at all in each situation?”
It may well be that a computer vision project would save you X amount of money every day, but as it has a high risk and takes more than a year to complete, it makes little business sense.
After thoughtful consideration, you will have a list of actual business problems you can solve by using your data.
Data First vs. Problem First approach
So what is the main difference between Data First and Problem First approach?
For one, the Problem First approach requires some extra effort of bringing together a bunch of busy people and asking them to spend their time on elaborating on your organization’s bottlenecks.
On the other hand, by putting in this extra work, you’ll arrive together at a set of very specific and well-defined problems that are both solvable and hold a long-term benefit for the company.
“With the Problem First approach, you’ll arrive together at a set of very specific and well-defined problems that are both solvable and hold a long-term benefit for the company.”
Another advantage of the Problem First approach is that although the initial estimates will be far from specific, everyone involved will at least have a general idea of the difficulty of the project and results to expect.
How to ensure you have the right data?
This process of asking the right profit-oriented questions sounds great, but what if you don’t have the right type of business data yet?
On the other hand, you do have a product that you believe could benefit from some additional intelligence.
What’s the most efficient way to start collecting the right data to solve your problems or improve your product?
What works best here is to think about how your product currently works. Ask yourself what are the main issues slowing down or hampering its work or speed at different points.
Usually, products or services can best be improved:
Where some human involvement is required — ask yourself if you can help those people be more efficient or remove the task altogether.
When there’s a large drop-out of customers — can you make your product stickier or more engaging?
When the task is being performed at a suboptimal level — are there any actionable steps that could be automated or even left out?
After you’ve made a list of the steps you’d like to improve, arrange them by the level of value they can potentially generate. Next, find a machine learning expert who helps to evaluate the level of difficulty of each item on your list.
It may well turn out that some problems are unsolvable with machine learning or that it would be more efficient to have a person working on the task on a daily basis.
“It may well turn out that some of your problems are unsolvable with machine learning or that it would be more efficient to have a person working on the task.”
After you’ve figured out some of the fields that could be improved with AI, you should make a plan of how to collect the data required for solving the problem. Also, make an assessment of how much time and resources the whole process from data collection to execution will take.
As a result, you will have a clear roadmap for improving your product/business processes with AI. You will also know where’s the most potential value.
You read through the article, but are a little unsure how to apply all of it to your company’s benefit?
Here are some key points to take away:
Use the problem first approach
Engage your team in the decision process
Make sure you actually have to use ML/AI to solve the problem
Prioritize your AI projects based on their impact
Make sure your data collection process is optimal
Thanks for reading! If you liked it, please support by sharing the post.
Markus’ career started as a bioinformatician and he was soon drawn towards data science and developing machine learning projects. While acquiring his Master’s degree in Computer Science, he spent two years at STACC (Software Technology and Applications Competence Center) where he developed data analytics driven tools for personalized medicine applications. Markus also has experience with deep learning algorithms, natural language processing, and image recognition.