Solving the AI scalability problem

It is undeniable that AI technology is going to revolutionise business practices and radically transform the business environment. The problem, however, is not knowing that businesses need to integrate AI solutions in order to remain competitive, but rather how. Far too many businesses fail to move beyond the proof of concept stage into wide scale deployment.

Firms often become all too focused on incorrect details or building a model that, instead of solving a problem, merely proves a point. According to research conducted by Harvard Business Review, three-quarters of C-suite executives believe that if they don't adopt AI solutions to scale, they risk collapsing. The significance of the problem, coupled with current shortcomings thus requires a radical solution: ignore proof of concept and scale up immediately.

The Harvard Business Review's study surveyed 1,500 C-suite executives from 16 different industries in 12 countries. It found that although 84% know that scaled-up AI solutions are critical in achieving their strategic growth objectives, only 16% have moved beyond the AI experimental stage. The companies that successfully managed to implement scaled-up AI solutions had one common denominator: they all skipped the proof of concept stage.

The firms that adopted this method doubled their scaling attempts, succeeded at scaling AI initiatives twice as often and - as the projects were structured correctly and involved a steep learning curve - the completion of projects was not only quicker, but less money was also spend on pilots and fully scaled deployments. The end result of this tactic? An ROI 3x that of their lower performing equivalents.

In purely financial terms, as the mean spend on AI by a company totaled $215 million in the past 3-years, the 54% difference equals a $115-million gap in lossed returns from AI development and deployment. As an aside from money, the firms that scaled successfully found a marked improvement in satisfaction to workforce productivity and customer service.

What is wrong with proof of concept?

To show why proof of concept doesn't work, let's look at a practical example. A company allows for the six month development of a customer experience optimisation platform as a proof of concept in order to better customer service. It gets up and running, it is confirmed it does the job - as many others have proved before - and they move to scale. It sounds perfectly logical however there is a crucial mistake.

The firm has only proved it technically works but they have not thought about the requirements necessary to put it into production. These requirements can include data bias, ethical considerations, model risks and data privacy. As a result, the firm is in technical debt as the concept was never built for scale in the first instance.

A firm that the Harvard Business Review worked with, Nordea, demonstrated how skipping POC works. Nordea, the largest banking group in the Nordics needed to relieve their customer service team from menial enquiries so they could focus on more complex customer issues. To overcome this Nordea developed a chatbot. Nordea had already established the structure for testing and development which included ethical frameworks, organisational design, the right data inputs and talent. Thus, they missed out the POC stage and scaled up immediately.

With the right data, they built a minimum viable product, gave it an avatar and waited to see how it interacted with customers and, in turn, how customers interacted with it. Nordea piloted their chatbot on a limited basic so a few hundred thousand customers. Customers immediately started understanding it resulting in email and telephone traffic reducing by 20% immediately and the use of the chatbot, along with related webpages, increased by 30%. This was all without proving to C-suite executives regarding how a chatbot operates.

If no POC, then what?

The research by Harvard showed that only one in five AI applications reaches the production phase. These are low percentages, but if POC is ineffective and now anachronistic, what replaces it? The research shows that effective businesses followed these practices.

Pilot to prove - while a POC is an exercise that intends to test a specific design idea or assumption, a pilot is the testing of a fully functional product that is offered to a portion of your target users. A pilot is launched directly into the real world. This has major benefits over a POC as POCs are launched on a much smaller scale meaning the value of their tests in often much less. Furthermore, a pilot gives much more feedback into how an AI development is used practically as how you think a new tool will be used often differs to how it is actually used. This allows you to tweak the product accordingly before a full-scale roll out.

Focus - companies often initiate multiple POCs resulting in the spreading out of limited resources and fatigue across industries. Rather than having many ideas, focus on the ideas that will add most value to your operations. Less is more. Consider only a few valuable developments and conduct proper research with the goal of production.

Cross department collaboration - a common flaw with any innovation is that it is often developed in a solitary team or department - often IT. This often results in the siloing of an idea. However, when an idea has cross-department support and collaboration that is multi-dimensional it connects the innovation to business outcomes increasing the likelihood of success.

To the future

To take AI ideas into lift off, think big and start small. This means one must consider the implications of scale at the very beginning of the design process. Furthermore, it also requires an enquiry into what value is and in the future. It would be a mistake for firms to concentrate on their current relevance while overlooking how they can deliver value in the future. Firms must accept that AI is changing industries all over the world. Once this is accepted, firms must see this as an opportunity and exploit it.

Exploitation this opportunity does not manifest from wasting time proving a concept that already exists as consensus. Instead, firms should have scalability at the forefront of their innovations and look to pilot - so long as all the hard graft has been implemented to ensure it can be put into production.