Artificial Intelligence (AI) presents an opportunity for businesses in nearly every industry to evolve and improve business operations. The numbers speak for themselves: Data from Fortune Business Insights shows that the AI market size was valued at $27 billion in 2019 – a figure projected to reach $267 billion by 2027. It’s no wonder enterprises are grappling to get involved in what is likely the most prolific technology of our time.
But implementing an AI strategy is challenging, especially for legacy organizations and those who simply don’t know where to begin.
How to plan an AI pilot project: 4 tips
By definition, pilot experiments are carried out before large-scale quantitative research to avoid using time and money on an inadequately designed project. Pilot projects are really no different: They offer a “try-before-you-buy” that helps businesses understand the challenges and potential value a new solution or process can bring to their company. While most follow the same basic principles – define the problem, desired outcome, cost, and resources needed and measure results – the ever-changing field of AI has some particular nuances that you should understand. Consider these four criteria before selecting an AI pilot project.
[ Check out our primer on 10 key artificial intelligence terms for IT and business leaders: Cheat sheet: AI glossary. ]
1. Clearly define the business outcome your AI project should deliver
AI has gained a lot of steam over the past few years, and it’s still a hot topic of conversation. While some aspects feel positive – AI helping accelerate COVID-19 vaccination rates, improving your shopping experience – and some negative – such as possible job losses or ethical dilemmas – AI is here to stay. Smart businesses are finding ways to strategically use AI for everything from marketing to hiring.
But despite its broad appeal, AI isn’t right for every business process, and it’s certainly not one-size-fits-all. Before embarking on an AI pilot project, make sure your problem and the outcomes you hope it will solve are clearly defined. Are there business measures in place to track your progress?
Once you define the project and sell it to the leaders within your organization, it’s time to tackle the next challenge.
2. Choose the right approach
Open source, cloud services, products with “built-in” AI features, hiring a data scientist to build your own solution – there are many options for tackling AI initiatives, with pros and cons to each of them. Off-the-shelf cloud solutions, for example, are widely available and usable but are trained on data that’s not yours and can be very expensive. Hiring a data scientist to build your own solution can provide a level of customization you may not find elsewhere, but getting the right level of talent and domain expertise is hard.
To achieve the results you want on the path of least resistance, you need to consider the challenges that come with the solution you ultimately choose – none are perfect. Weigh what fits best within your existing IT ecosystem, the resources needed to not only get your pilot project up and running but to deploy and continually improve upon it, and the stakeholders involved to keep the initiative on track.
[ Want best practices for AI workloads? Get the eBook: Top considerations for building a production-ready AI/ML environment. ]
3. Anticipate a learning curve
AI models need to be trained. They get smarter over time (only if properly engineered), but the learning curve isn’t just for the humans interacting with it; it’s for the models themselves. Even using an out-of-the-box solution from one of the major cloud providers takes tweaking and fine-tuning to fit your business needs, inherent biases, and target metrics. Understand that this will take time and effort, so consider how much of both are necessary for a successful project.
[ Public data sets can help with training. Read also: 6 misconceptions about AIOps, explained. ]
On the other side of the spectrum, you can have best-in-class technology in place, but if proper training isn’t provided for the people using it, what good is it? For example, an algorithm can understand the results of an X-ray as accurately as a human, but if an understanding of the technology and processes is not in place in the healthcare organization putting it to use, it’s a lost cause.
4. Understand testing vs. production-ready
Business needs and data are constantly changing. To get the best possible results from your AI initiative, ongoing testing and retraining of models are necessary to deliver accurate results to your customers.
And this doesn’t go away after implementation. One-time acceptance testing works for traditional (static) software, but not for AI systems that must change on their own as the world around them changes. Plan for a different kind of monitoring, online measurement, and retraining pipelines to correct for data and concept drift, potential biases, and other bumps in the road that may not be apparent in early testing.
By taking these four criteria into consideration, you can choose the most impactful AI pilot project for your business and provide the fastest and least-risky learning path for your organization on how to best apply AI.
[ Get exercises and approaches that make disparate teams stronger. Read the digital transformation ebook: Transformation Takes Practice. ]