Harnessing real business value with AI

Artificial intelligence is a booming field of research getting a lot of attention in the media due to its impressive applications such as image recognition or self-driving cars. The topic is however usually depicted from two fallacious angles, i.e. either largely excessive expectations in terms of what it is so far capable of, or from an irrational fear that AI may eventually try to replace, or even kill us. Beyond these journalistic fantasies, let’s look at what is in the box and how you can leverage AI to create actual business value.

BENJAMIN BIERING | FORMER HEAD OF DATA SCIENCE, 2021.AI | JUNE | 2019

How does AI deliver value for your business?

AI is commonly leveraged to take on tasks and processes which would either have to be done manually or which are too complex to be performed by a human. AI is particularly interesting when the tasks at hand can be automatically learned by “just” looking at data. For instance, you would like to classify a set of flower pictures into two groups: roses or camellias. A human could easily perform this task but if you have one million images to manually classify it may take you a really long time.

Another way to go about it, is to get some measurements of the flowers (e.g. size and number of petals) and use this data to train an algorithm which will automatically determine the cut-off values in these measurements in order to make its decision – but first you would have to “engineer/collect” these measurements – not so appealing either. Last but not least, you can feed the images as raw input to an algorithm and let it look at the images’ pixels, learn what a rose typically looks like and let it make the decision itself.

Now you may think, why should I care about flowers? How does that relate to my business challenges? Well, the flower classification problem is primarily a classification problem. AI algorithms tend to be data agnostic, they don’t care whether the numbers they crunch are measurements of petal length or customer lifetime values.

Therefore, you can use the same type of algorithms to classify flowers, customers that are likely to churn, or bad vs good credit, etc. You can have a look at a sample of typical problems that can be tackled by AI across industries.

In expert terminology, this algorithm classifying flowers is called a “data product”. It’s a classical software product which contains an algorithm trained on data. In terms of product management, it doesn’t change much if compared with the development process, it only takes the product manager to be aware of the typical pitfalls of developing algorithms on top of regular code.

However, in order for this data product to create real business value, it has to be alive and maintained, connected to your systems and your business processes.

Let’s consider the customer churn example. A good solution for this problem would be to collect as much data as you can on your customers and train an algorithm to classify whether a customer belongs to the category “likely to churn” or “unlikely to churn”. Then, you will have to put this algorithm in production – this means that the algorithm is placed on a server, just like a website and can be called to make a prediction about a particular customer – just like a website would be requested to serve a certain web page.

Putting algorithms in production

Putting algorithms in “production” means that you can talk to them from anywhere and more importantly that you can connect them to your internal systems such as your databases, CRM or ERP systems. Thus, when one of your customer success managers meets up with a customer and looks at the CRM system, he can get an estimation of how likely that customer is going to churn, perhaps also for which reasons and which marketing actions he could undertake to prevent that from happening.

Of course, since people change (i.e. personal situation, tastes, financial means, etc.), it entails that the algorithm may have to follow up on these changes to always provide the most accurate prediction. In turn, you will have to put in place mechanisms to continuously assess how well the algorithms are performing and whether you should “retrain” them, i.e. updating the models with the newest data. Particularly, if the algorithms are used as decision support tools to make critical decisions about your customers. Industries with such requirements will need to be able to explain why a decision was made. Fortunately, there are ways to explain the outcome of an algorithm using mathematics, but this will be the topic of another blog post – stay tuned!

What does it take to get started with AI?

At the end of the day, AI is just like classical software development: you will need to think about systems, databases, teams, skill diversity, and code maintenance. Here are a couple of things you should consider when starting an AI project.

AI Project

1. Get your business goals straight

As Simon Sinek would say, start with the “Why”. There are many reasons why you should try to leverage AI in your business: are you trying to relieve a pain point, are you searching to automate a manual process or even create a new data product?

A notorious example is the company Stitch Fix which understood from the very beginning that data would be at the core of their services. Therefore, they integrated algorithms in every aspect of the company from warehouse assignment, to product recommender systems, demand modelling and new style development.

AI is here to stay and refusing to think on how to leverage your data to strategic ends will put you on the sidelines of the competition.

You should not consider AI as a fancy buzzword that will probably go away like big data. In e-commerce, data- or AI-based services such as product recommendation, personalized content or dynamic pricing were cutting edge features a couple of years ago, now they are standard must-haves.

You do not want to be competing on the same features as others, you want to be ahead of the curve. This requires a vision and some level of investment because like any other software project, it takes time to implement, gain experience and reach an operational mode.

2. Get hold of your data

Data is one of the most paramount aspects of an AI initiative. Are your data sources accessible, complete and valid? Do you even understand your own data? The answer to these questions will most likely be negative. I’m still waiting for the day where I will encounter a company having full control over its data. So you will have to put efforts into shaping up your data sources to your projects requirements.

There is a saying going around that data scientists spend 80% of their time grooming data and only 20% working on modelling – sadly, in practice it’s true. I have observed that data usually sits on (legacy) BI systems only to be used in a particular context, and the architecture of those systems wasn’t thought out to support that kind of AI/data initiatives. Thus, having a healthy data infrastructure is of the utmost importance to any AI initiative you plan to undertake.

An advice is to not go down the path of believing the tales of consulting BI experts preaching that if you buy this “big data” infrastructure costing a small fortune only then will you reap the benefits of data. Start small but start smart. Plan for gradual and incremental scaling of your data infrastructure. Focus on servicing the data and stakeholder that matter to your initiatives first. Designing, building and maintaining a data lake is expensive and time-consuming, so make your effort worthwhile and leverage every bit of data you store.

3. Algorithm development

Developing a data product answers to the same development rules as any other product. Do not be afraid to start with a simple MVP or even a dummy model in order to test the data flow and infrastructure around it. Focus on proving the feasibility and getting the infrastructure around the product right.

Once the framework is in place, it will be much easier to iterate, refine and complexify your product at will. Google provides a number of rules to develop machine learning models and most of these are more related to engineering problems than scientific ones. Adopt a lean methodology (Build-Measure-Learn), where you implement something simple at first, measure the results and decide on how to proceed based on the learnings you have just made.

One can always try to buy out-of-the-box solutions for your problems to minimize the cost, but usually, real life is slightly more complicated. If you want a real AI solution, you will have to rely on the expertise of data scientists. Those with experience may be hard to find, but since machine learning (ML) is taught in many different fields (i.e. physics, engineering, bioscience, finance, etc.), you may go for candidates with quantitative background, hands-on experience with ML, and most importantly solid coding skills – preferably Python (as of the year 2019).

4. Pushing algorithms to production

As pointed out by Google in a famous article on machine learning technical debt, the machine learning code (here referred to as algorithm) only constitutes a tiny part of the full production architecture. Putting algorithms in production is hard because of many reasons. There is no one-size-fits-all solution and solutions provided by major cloud providers force you to fully enter their ecosystems and impose strict limitations on how to deploy. However, they take a lot of complexity off your shoulders. You will have to decide for yourself how you want to approach this problem.

5. Algorithms maintenance

Your algorithms have finally made it to production and deliver value to your company on a daily basis. Now, you definitely want to ensure that they will keep on doing a good job. In order to do so, you will need to monitor their performance with regard to a baseline. On the one hand, if the algorithm was delivered with an average performance of 70% accuracy, you should not expect a production accuracy of 90%. On the other hand, it should not fall much below 70%, otherwise, this could be an indication that something is wrong either with the algorithm or with the data and this signals the need for a diagnostic.

Algorithm monitoring will require to consistently record the data the model is fed with along with the predictions it outputs in order to compare what the model predicts to what actually happens in real life. This boils down to an engineering problem that must be thought out from the start of the project.

Finally, when unrolling a new version of your algorithm, think about doing A/B testing to ensure that the new version will effectively produce better results than the previous one.

Summary

There is a huge opportunity for companies across industries to reap the benefits of AI without having to make massive investments upfront. Start by aligning your business goals and divide a plan on the different ways AI can help you achieve these – don’t hesitate to hire trusted external consultants to uncover these opportunities.

Commitment to a good execution is essential. Going through the five above mentioned steps will require a good dose of discipline and rely on skilled data scientists, engineers and/or partners. The author’s opinion is that AI-powered features in software products will become increasingly pervasive and it will require an ever-greater amount of expertise and tooling to stay competitive.

You might also like…