The future of AI deployments reaching production is bright in 2021
Most artificial intelligence (AI) projects fail. About 80% never reach deployment, according to Gartner, and those that do are only profitable about 60% of the time. When we take a moment to consider the signs of successful AI all around us, these numbers may come as a surprise. We have voice assistants for our phones and homes, optimized online product searches, advanced fraud detection at our banks, and more. Yet as it stands now, we’ll never see the fruits of the majority of AI endeavors. In the final part of our five-part series on 2021 predictions, we look at the future of successful AI deployments.
These statistics might seem disheartening for companies that are turning to AI for positive impacts like greater revenue, lower costs, and more personalized, effective customer experiences, but we’re seeing signs of promise. In 2021, we predict that companies will start to overcome the 80% failure rate of deployment. Gartner has further predicted that by 2024, 75% of organizations will shift from piloting to operationalizing AI. This change in momentum will be driven by greater accessibility to data and the development of highly flexible models to adapt to specific business needs.
Why do AI projects fail?
Considering AI has been around since the 1950s in some form, one might wonder why we don’t have a perfect blueprint yet for deploying it successfully. In reality, there are tons of variables that go into building effective AI, which makes it difficult to prescribe set steps that will work well every time, for every company. Still, progress is being made in gathering best practices (mainly through learnings from success stories and failures), and with those, common patterns are emerging on what often leads to failure. Here are a few areas where companies can go wrong:
- They didn’t define a narrow business problem. Many organizations choose the wrong problem to solve. They may select something that’s too general, resulting in a model that’s useless for specific business use cases. They might choose a problem without enough data to support the solution. They also might choose a problem that would be better solved through something other than AI. Any time there’s a misalignment with business priorities, problems will occur. It’s also important to ensure all stakeholders, from the top down, are clear on the objectives of the project.
- They don’t have the right team. AI has a talent gap problem, which means companies often have to scramble to recruit team members with the right skillsets for building effective AI. Most organizations are currently ill-designed to support scalable AI ventures, requiring re-orgs, new hiring efforts, and leveraging of third-party resources.
- They don’t have enough high-quality data. To make accurate predictions, AI models need a massive amount of data. These models must be trained to handle any potential use case they’ll face in production, which means datasets must cover a wide range of use and edge cases. Many companies fail to collect the appropriate amounts of data for their models, and have poor data management techniques for accurately labeling that data. This results in poor decision-making by the model.
- They didn’t confront bias head-on. In our diverse, global business world, it’s vital to take a responsible approach to AI from the start of model build and beyond. Most companies don’t set out intending to create biased models, but do so accidentally by failing to include diverse perspectives and data in their processes.
Of the projects that are deployed successfully, many face challenges with model drift — or changing external conditions — that lower the model’s accuracy or even make it obsolete. Models must consistently be retrained with new, relevant data to overcome this hurdle.
With all of these factors in mind and the many additional aspects that weren’t highlighted, it may be clearer now how difficult it is to deploy (and maintain) AI successfully.
Overcoming the failure rate will be challenging for companies large and small, but the future’s looking brighter for a number of reasons. The first is that data is more accessible and abundant than ever. When we recall that effective machine learning is built upon a bounty of data, we should expect to see models that produce greater accuracy, models that cover more specific use cases where data may have been hard to come by in the past, and models that work better for more end users.
The second trend we’re seeing is that we’re learning. The wealth of information available has allowed companies to see what others are doing in the space. Companies are conquering the mistakes of the past by gaining knowledge around best practices and common pitfalls. More resources are available than ever before. For example, Alyssa Simpson Rochwerger and Wilson Pang’s upcoming book, Real World AI: A Practical Guide to Responsible Machine Learning, includes real-world success and failure stories, and clear action plans for lucrative deployments.
With these tools under their belt, we expect more companies pursuing AI to succeed. Despite the pandemic, there’s a great chance AI will accelerate because of it. Already, social distancing has resulted in the development of AI that’s more flexible to changing supply chains and customer demands. This flexibility, bolstered by ML techniques like reinforcement learning, may create AI solutions that are increasingly adaptable to change, and therefore more effective post-pandemic and longer-term.
Considering these factors collectively, the future for AI projects holds promise. The companies that learn, adapt, and mobilize quickly will be the frontrunners in the space; they’ll be more equipped to reach production, and ultimately, the holy grail of profitably. 2021 will serve as a critical turning point for AI ventures, as the tides shift away from failure and toward success.
At Appen, we have spent over 20 years annotating and collecting data using the best-of-breed technology platform and leveraging our diverse crowd to help ensure you can confidently deploy your AI models. For a practical guide to responsible machine learning and successful deployments, check out the upcoming book by Alyssa Simpson Rochwerger and Wilson Pang, Real World AI: A Practical Guide to Responsible Machine Learning.
Wilson Pang is CTO at Appen.