Artificial Intelligence Applications — Where to start?

How do I create value from Artificial Intelligence for my company?

That’s what probably comes to mind of most business executives when confronted with the technology. Well, how do I find the right application or at least domain of artificial intelligence that will boost my business?

Not so easy! If you look at the many, many different attempts on segmenting the “AI market”, you’ll probably get very confused. Why? Because there are already more than 3,500 ‘real’ AI-startups out there worldwide, according to Roland Berger and Asgard. On top comes an uncountable number of fake AI-companies mentioning it as part of their tech stack while working with traditional technology. And, of course the big boys such as Alphabet, IBM, SAP, Facebook etc.

Why is AI different to other software solutions?

The traditional approach from a consultant or investor’s perspective would follow the major industry segmentation, such as healthcare, manufacturing, logistics or finance. And this is usually a good starting point because you normally know in which industry you are in.

But there are two problems here: First, AI will sooner or later impact almost every industry, therefore your list of segments will become very long if you don’t put it all together in one big fuzzy category that you label as ‘other’.

Second, AI tackles the fundamental functions that are underlying in so many different domains. This makes it is really hard to organize AI by industries. Is your Natural Language Processing algorithm more relevant for customer service in the finance segment or in transportation? Oftentimes, it is a question of input data that determines the actual use case, not the underlying technology stack.

A solely technological approach is also difficult. For advanced applications many fundamental AI functions such as Machine Learning, Perception, or Natural Language Processing have to be closely interconnected. Most segmentations, that you can find on the web, will therefore mix up industry-based and technology-based approaches.

Start with the problem

Let’s try to get a bit of clarification into this space: If you don’t care about the underlying technology and neither about the industry that you are in, focus on the problem, you currently have.

The precondition is always to have a certain amount of data to work with on an AI-related use case. Then imagine explaining your goal (preferred outcome that solves your initial problem) very detailed in simple instructions as if you would explain it to a child:

The three application layers

Okay, this might be a little bit abstract, but it gives you a clue that AI can in some way improve a lot of your processes. It is an attempt to create a foundational layer of core AI-infrastructure functions. Example applications in this set include:

• Text-to-speech / speech-to-text
• Image classification (e.g. person identification)
• Movement prediction
• Sentiment analysis

Infrastructure functions are rather generic and you can use them like a tool box and train them with a specific data set or combine multiple functions to solve a problem or improve a process. From a business perspective, they are not really plug’n’play but many of them will be easy to access via APIs on open source platforms. Simply add a bunch of good developers and start building your own applications!

A company is practically a collective of humans and (with an increasing share also) machines working together to achieve some self-declared goals. Independently of the industry a company is active in, there are several enterprise functions like HR, marketing, sales, finance, customer support that every company has to deal with. This is the second application layer — the enterprise functions. Example applications can be specifically trained and combined infrastructure functions to solve a dedicated problem:

• Social media analysis (e.g. classification, mapping, learning and interpretation)
• Customer support chatbots
• Budget planning
• Demand prediction
• Dynamic pricing

Enterprise functions are being already developed to solve commoditized problems. They show the highest potential for productization. Most of them will be offered in “as-a-service” business models designed to be used by non-tech staff. The perfect solution, if you want to apply AI into your company processes without reinventing the wheel. But take good care of the data you are working with, and that decisions are actually derived from the application. AI is based on logic, not magic.

On the third layer there are industry-specific functions, where finally the good old industry segmentation makes a lot of sense (healthcare, retail, mobility etc.). Applications on this layer are highly specialized to solve an industry-related problem:

• Navigation & routing (Transportation)
• Financial risk scoring (Finance)
• Autonomous driving (Automotive)
• Picking robots (Logistics)
• Gene analytics (Healthcare)

On the industry layer, we enter the solution business as most them require unique input data from the applying company and highly tailored interfaces to the corresponding requirements and already existing data processing systems. That is where the big corporates spend a lot of money. If you are able to offer a modular designed solution platform as an application developer, that might be one of the next big things.

Follow the data, follow the money

After trying to structure the application landscape, one question remains: Where to start applying AI? Considering the maturity of the technology in different areas, available computing power or focus areas of academic research can of course make a lot of sense. But to keep it simple: just start where you have the best data.

It’s important to say best — not the most. Every algorithm can just work as good as the input data is that fuels the system. Therefore you can look at industries where already very detailed, well-structured data is digitally available and accessible in order to improve efficiency through AI or enable value-added services that generate revenue from inherently data-driven business models.

The best data you can have as a company is ideally:

• Unique (brings additional value to the market)
• Continuously captured (to grasp even small changes)
• Well-structured (filtered, cleaned and organized in a data warehouse)
• Easy accessible (efficiently available for API requests, analysis and data enrichments)

It is also worth mentioning, to rather not start with the most mission-critical data of the company. AI ideally improves itself iteratively over time. And not only technology can fail, it is also humans that are confronted with new processes and ways of decision-making.

Therefore it makes sense to initially trial AI applications in areas, where failure does not cause fatal consequences. In a second step you can use up and running systems also for critical applications. Alternatively, start with a redundant test setup, measure the success and reliability and go forward, if you are confident.

Leading industries of the AI development

With social profile data, shopping behavior and personal interested data, we have already seen a gigantic value capturing in the e-commerce and digital advertising space led by Google, Facebook, Amazon, Alibaba and others. With the day-to-day increasing number of connected Internet-of-Things devices, data wealth is exponentially increasing in other industries such as manufacturing, logistics, retail, agriculture and transportation.

By digitizing more and more financial and medical records, the same development is observable in finance, insurance and healthcare. Though, regulatory and compliance issues have a much higher impact in these industries.

Conclusion

Coming back to the initial question on how to derive value from AI applications. Instead of browsing through hundreds or thousands of different use cases and vendors, start breaking the problem down by asking three simple questions:

Which role is AI supposed to play for the business model?

• Enhancement of existing processes
• Inherent value of the business (e.g. social media analytics provider, product recommendation engine)

Where does the business capture the best data?

• Common enterprise functions
• Industry-specific sensor, customer or communication data

Does the business have its own resources and capabilities to develop AI solutions?

• Yes
• No, solution and implementation partners are required

With this concept of different layer applications, it should be easier to narrow the focus now. Picking the actual application and right vendor for implementation depends highly on the existing structures and preferences. We will dig deeper into this topic in one of our next posts.