Robot hands ticking off a checklist on a clipboard

Companies need to digitalize their processes to stay competitive in today’s market. For that, many rely on AI start-ups, which offer everything from machine learning (ML) to decision trees and promise everything from 100% accuracy to immediate results.

But not every start-up is successful in what it offers or promises.

Early signs of a failing AI start-up go beyond technical issues. Much of the problem actually comes from unrealistic expectations—both on the client side and from within the start-up itself. Identifying success factors within a start-up comes down to a simple question: Did the start-up deliver what the client asked for? That can be harder to achieve than it sounds. Our research team spoke to several experts who lamented that there is still a kind of “magic potion” attitude toward AI. Many clients ask for a ML solution to boost sales before there even is something to sell.

This is similar to the dilemma Stuart Russel presents in his book Human Compatible: Artificial Intelligence and the Problem of Control: a customer keeps changing his mind on pizza toppings, driving a robot waiter crazy. How do you deliver what a client asked for when the client doesn’t know what to ask for?

Short answer: by ensuring you’re growing your AI start-up the right way. To that end, here are some crucial dos and don’ts.

Things You Should DO

DO focus on value.

Clients want to see the greater value of the solution they purchased, whether they’re working with an AI start-up or a traditional one. After all, companies want to boost sales and reduce costs—that hasn’t changed for decades.

Every AI start-up should start with research, even before meeting with a client. That way, they’ll be prepared to explain both the problem and how their solution will solve it. Next, they need to think about cost—not just implementation, but also maintenance. AI projects aren’t cheap, nor are they normally paid for upfront.

Research and practical cost consideration will help you position your start-up to provide value to clients.

DO understand the value of specialists.

When people think of AI, they imagine the machines. But there is a remarkable number of hands involved in the creation of an AI model—highly specialized hands. AI professionals are rarely interchangeable, so someone working in industry sector A may find it’s difficult to get hired by sector B. Sometimes, even moving from the purchasing department to the sales department of the same company can be an issue.

Accenture PLC research by Paul R. Daughterty and H. James Wilson predicts that the future will have three new job types: explainers, trainers, and sustainers. Specialists who communicate the best use of AI will fill the explainer role. Therefore, it’s to the benefit of AI start-ups today to understand and leverage that potential for tomorrow.

DO make plans for scaling up.

There’s one big, often overlooked thing that will make or break an AI start-up: scalability.

Having worked in a multinational organization, I’ve suffered first-hand the long purchasing cycles that often make it impossible to work with small start-ups. Big clients lack flexibility, and cloud operations can be more complex and costly than traditional approaches. VCs, too, want scalable products and cheap implementation.

Making an AI start-up scalable is, in itself, a huge project, but a good place to start is with company culture. Never underestimate the importance of keeping coders motivated. The partner of one consulting firm told me that he makes a point to alternate “stimulating” and “profitable” projects to keep coders on their toes. Without happy coders, there’s simply no way to even start thinking of scaling up.

Start-ups lacking scalability will find themselves stuck in a whirlpool of small consulting projects that never quite push them into the big leagues.

Things You Should NOT DO

DON’T forget the human element within AI.

Every company that uses AI still has humans working for it, and those humans will need to work in harmony with the machines. AI start-ups and their clients therefore need to identify what level of human-machine interaction their solution requires. Will the machine simply assist the humans, like a grammar corrector? Or will it augment the abilities of a skilled professional, like a skin cancer screening system that narrows down potential prognoses? Will there still be manual elements to the solution, or can we attempt full automation?

AI is about interpretability, but that interpretability will change significantly depending on the goal and level of human interaction. A 90% interpretability rate is outstanding, but reaching 99% is exponentially more complex and still  human driven.

Don’t forget to factor in the human element, or your whole operation will fall to pieces.

DON’T discount the importance of forecasting.

AI start-ups should plan for scenarios both general and extreme. Like playing chess, you don’t just need to move—you need to move in anticipation of your opponent’s next several moves. You’ll need contingencies for data drift, black swan events (like COVID-19), and attempted application of the AI outside its valid domain. Failing to do this could lead to major setbacks, exhaust your resources, and ultimately ruin the project.

A great forecaster performs frequent updates and introduces new, competing models. The same should apply to AI projects. If nothing else, be sure you can answer these questions: How often does the model get updated? How often is new data integrated to make the solution more accurate over time?

DON’T confuse desktop pilots with reality.

The “magic potion” of an AI solution that works everywhere, every time is a fairy tale. Here in reality, AI models can require as many iterations as any other product. Data collected during and after implementation will lead to scaling up, customizations, and a final product far removed from anything seen during testing. Not to mention, AI inevitably encounters scenarios in which it doesn’t perform as expected. Data evolves, so higher accuracy in a model actually makes it harder to find data sets to identify parameters.

ML master Andrew Ng warns that desktop app testing—a way of testing software functionality, security, etc. prior to deployment—solves only very specific problems. The solutions, therefore, are not transferable to other companies or projects. Many start-ups discover too late that successful pilot projects fail as a solid business model.

So make sure your expectations are aligned with reality, not fairy tales.

Despite all the bells and whistles, AI businesses are still businesses. That means they can fail just as much as any other business—and for many of the same reasons. Savvy AI start-up founders know to offer great ML solutions without forgetting to learn about business. Understand the skills you need and the market you serve, and you can define the best value propositions for your customers.