Artificial Intelligence in Business: Insights from Both Sides of the Pond

The hype bubble around AI makes many executives think that if they’re not already in the loop, they’re too far behind and missing out on the power of AI transformation. The reality, however, is that businesses should think of AI just like of another helpful tool, simply one of strategic investments.

Not every process should be automated via AI just because it’s the latest trendy thing to do. The approach of “if you have a hammer, everything looks like a nail” is generally a bad idea.

A big part of the effective use of artificial intelligence in business settings is being smart about where to start. Mass corporate adoption of intelligent technologies is still mostly in its infancy, with possibilities and pitfalls growing by the day; it can often seem hard to even begin to work out what to focus on.

So how to figure out where and how to best use AI for your company, right now?

Judging by presentations from this year's O’Reilly NY AI conference, this question rings even truer with today’s growth and democratization of AI and ML. The overall vibe and message to businesses can be summed up as:

A good way to go about this is to look at the common starting points in recent use cases, the ‘whats and whys’ that successful AI adopters across the globe share. What possibilities and challenges do they take into account first? What are the prerequisites? What data and processes are good starting bets for AI integration? And all the while we should keep in mind whether the companies in question are large or medium-sized, from the same or adjacent industry, whether they operate in Europe, North America, and so on.

Intelligent business process automation largely deals with administrative, management and financial processes and is all about optimizing tasks and workflows through the combination of RPA and AI.

An example of that would be identifying and transferring data from various enterprise databases into a unified record system. Think moving data from emails and call center systems in a CRM or CDP business information systems—with sorting, structuring and assignment of data to corresponding customer profiles, all automated and happening on the go with the help of artificial intelligence and RPA. Managing credit data and transactions also comes to mind, as well as automated continuous checking for payment inconsistencies and format compliance in real time.

This is about analyzing data and identifying patterns for automated interpretations and predictive analytics, like fraud detection, actuarial analysis and assessment automation, as well as predicting what service, content or items a profiled customer might want to purchase.

These are good candidates for an AI makeover, so you may as well go for it if your enterprise can square its tasks and processes into any of these categories while being fully data-driven.

Paraphrasing Ben Lorica, the majority of successful artificial intelligence applications are not about reinventing the wheel and completely starting from scratch, but essentially building on the existing data and BI frameworks, clearing and structuring the former and layering machine learning on top of the latter.

Last year, O’Reilly named AI skill gap the largest barrier to widespread AI adoption. With corporate demand skyrocketing in the last few years, there’s just not enough experts. Another issue is that the educational programs often can’t keep up with the rapid pace of innovation and the surge of new discoveries within the field.

Sought-after AI software developers have up-to-date university-backed training, upskill constantly, and have considerable on-the-job experience. Companies that have just started adopting AI strategies into their operations need seasoned professionals like that to take leadership roles, shape AI and data strategies, and oversee the processes.

Aside from much-needed changes in official educational programs, there are, however, other ways to find answers to that problem. Retraining current employees in-house while employing skilled supervisors can be a better solution in the long run, particularly for bigger enterprises. Delegating AI implementation to external providers and consultants is also an option, especially well-suited for SMBs and organizations at the very first stage of the process.

Quality data is crucial when it comes to machine learning and analytics, yes, but the quantity of the data is another pain point that is often overlooked in discussions about commercial adoption of AI.

Take one of the more impressive and well-known examples of machine learning and AI: image (facial) recognition and subsequent face generation.

With the latest GAN 2.0 by NVIDIA, we can now produce hi-res realistic images of people who have never existed. The algorithm was trained on ImageNet dataset—which has 14 million hand-annotated images in it, and all of them were used for machine learning and image generating.

The problem is that while computers now are incredible at recognizing patterns and learning things, they still need all that big structured data to learn to do something at the business-grade level. Gathering millions of clean, annotated data points for machine learning and effective AI implementation might simply prove too expensive or lengthy a process for many companies, especially SMEs.

To better illustrate it, I’ll borrow an example from the hub.berlin presentation by Ruchir Puri (Intel) on the future of AI for enterprises. Contrary to what’s often suggested in the media, there are fundamental differences between the way machines and humans learn to recognize images and objects in front of them. A toddler only needs a couple of pictures or photos of a cat to be able to recognize whatever cat, living or depicted, she’ll encounter in her life from that moment. Meanwhile, an ML-based cat recognition process will take labeled images of about a thousand of various cats in different angles, lighting, shapes and forms to reasonably identify photos of real-life cats (it can fail still, and will not recognize toys, cat-shaped ceramic figurines or drawn cats).

Moving on to enterprise-level examples, imagine an ML-based software that needs to audit and perform fact checks, compliance checks and cross-reference checks of legal documents. To be able to successfully do this, hundreds of thousands of specific documents (if not more) will be needed; structured, labeled, annotated in a particular way by qualified lawyers—so the algorithm could learn on that data. The scope, time and resources that’ll be necessary to acquire a dataset like that are far beyond the means of most organizations.

Not every data is big data, and especially with the current tendency toward hyper segmentation and using micro data pools for analysis—for example, when working with personalization—the need for new ways of assessment and processing becomes obvious.

At the moment, the ability to provide that data or training algorithms to operate on small data to learn more from less remains one of the biggest obstacles to enabling effective and widespread enterprise AI adoption. Major players in the field are working on it, with Intel reporting recent successes (though currently still moderate).