With the ease of availability and access of AI tools and technology, people are putting AI into a wide range of products and services, and even in applications where AI is a dubious fit, at best. Many times, organizations are feeling the motivation, “fear of missing out”, and perhaps customer or shareholder pressure to add AI capability to their offerings. It should come as no surprise that many of these AI projects are often half thought-out, at best, and often fail to deliver the desired results, if the results have even been considered ahead of time.
Sometimes, AI projects have a high-level, big vision, where the AI efforts are focused. Other times, AI applications are being focused on smaller tasks, or shoehorned into existing applications. The challenge is that for AI projects to be successful, there needs to be a combination of a larger vision for where AI could add value while at the same time, smaller, focused projects that allow organizations to ensure value in the real-world before diving deeper into AI capabilities and investment.
“Thinking Big, Starting Small, Iterating Often” is the Best Practice. But What does think big but start small mean?
As discussed in an AI Today podcast on this topic, the best practice for any high risk emerging technology project that has ill defined goals is to think big and start small and iterate often. But what does it really mean to think big? You can do a wide range of things with AI across the Seven Patterns of AI. You can build autonomous vehicles on the one hand, or sentiment analysis and NLP applications on the other. You can do predictive analytics and forecasting, or image recognition tasks. The applications, while all different applications of AI, are really very different, with different technology, data, iteration, investment, and ROI time horizons.
This is one of the risks that we have with AI, because it is such a broadly adoptable technology. In many ways, it's not like a website or a mobile app, where there's only so much you can do with these. With AI, we can do so many different things, and this is a problem. So we have to start with this idea of thinking big.
When we say think big, what this really means is we're trying to solve a big business or organizational problem that is worth solving with AI, because there are a lot of big problems that we can't really solve with AI or are better suited to solve with other approaches. In addition, the big idea that drives the AI investment needs to be long-term enough so that you don’t have to think big every two weeks.
So that means that the thing that you're thinking big about has to be a big enough problem that it can exist across multiple iterations, but it also can't be so big that you can't even figure out where to start or that you can never solve it. It's a kind of Goldilocks problem, in which you don't want to be too small in your thinking big vision, because that is too trivial, but way too big either to be impossible to accomplish.
This means that the think big vision has to be a substantial organizational or business driver with an identified outcome, such as saving costs, increasing revenues, reducing time or resource investment, improving compliance or security, reducing risk, or otherwise providing a significant competitive or other advantage. It goes without saying, then, you should not be doing AI just for AI’s sake. This isn’t any sort of thinking big.
Setting Goals for AI Projects
A big problem worth solving needs to be divided into smaller, addressable parts. So, the “think big” vision needs to be done in such a way that you don't boil the ocean. You're not going to get really far, and you're just going to get super frustrated. And, of course, your AI project will fail. Not every part of the big vision has to have the same priority. Break it down one level further into the different patterns of AI. If a big plan has multiple patterns, each pattern should have its own smaller project. And within each pattern, it could be broken down into even smaller projects, prioritized in a way that will get you to success.
If what you're trying to do maybe incorporates three different patterns of AI, by starting small, we can pick just one of them, and we don't need to be doing all of them simultaneously within that same iteration. This is the core of the first phase of the six-phase best practices CPMAI methodology that is being used to achieve success with AI projects.
The first phase of CPMAI, the business understanding phase, requires you to find the smallest subset of this “think big” problem that you can implement with the greatest return in the shortest amount of time. The reason that you're going to do that is you want to make sure that you are having early wins and early successes. You’re going to have to justify your ongoing investment in AI, and the best way to do that is to show many incremental, quick successes and positive returns. Think about that first small problem that has the highest return in the shortest amount of time to get that early win.
When choosing that first small iteration you might discover a roadblock or a challenge that would stop your AI efforts, and therefore better to realize that quickly and early before you are too far along. For example, you might discover that you don't have access to the data you need, or maybe can't implement this model the way you want to, even if it is small. Maybe you need more data engineering to address data quality issues, or you need some data labeling, or maybe you’re doing some prompt engineering and need time to iterate prompts. Or, maybe there is a trustworthy issue that could kill the entire project. So that means we may have to choose another small project, because again, we want to be successful. These roadblocks are part of the “AI Go / No Go” in the CPMAI methodology.
How to Iterate AI Projects
The next thing to consider is how to take those early successes, or lessons learned from attempting those small projects, and iterate for further additional small projects that lead to the think big strategy. This is the “iterate often” part of the best practice approach. AI projects need to be short and outcome focused. Iterations should not be many months long, or even many weeks long. They should be quick enough to deliver a real outcome in a short amount of time, which means a week or two at most, and sometimes just days. This allows you to discover roadblocks and challenges early and address them before they derail your whole AI effort. You want to determine if the small iteration has met the business understanding objectives that you set up in your first phase of iteration and provide an outcome that supports the “think big” overall goals.
So at each iteration ask the question: has this met the objectives? And of course there's only two answers for that. If the answer is yes, that you've done it for that project or iteration and then can move to the next project that supports the overall goals, or iterate the current project in a way that delivers those goals even better. If you haven’t achieved that outcome or run into a problem, then you can go back to reconsidering the project, re-scoping your goals, changing to a different pattern of AI, or breaking down the small project into even smaller projects that help you get to your desired outcome better.
Ensuring that AI Projects Meet Real Business Objectives and Goals
The overarching goal is for AI to deliver real value. Otherwise, what are we doing here? While AI offers all sorts of promise to accomplish many seemingly magical tasks, they will all seem like just games if they aren’t accomplishing something tangible that provides solid positive benefit. This might seem obvious, but for whatever reason, people are not actually doing some of these fairly obvious things with their AI projects and realizing a ridiculously high rate of AI project failure.
Much of what we are talking about here is really bread-and-butter project management in many ways. Taking a big project and dividing it into smaller projects and iterating to success has long been a project management best practice. This means the key to AI success isn’t yet another AI model or tool, but rather better methods and processes.
The idea of project scope is what we’re talking about here with regards to breaking down larger AI projects into smaller ones. When it comes to thinking big and starting small, you can divide up your project in any of a number of ways. The way that you organize your project and the order in which you tackle these projects is a very elemental key to determining success.
Likewise it’s important to determine how AI specifically provides a solution in a way that a non AI solution wouldn't do better or more optimally. Because then you’re not going to do AI for AI's sake. Developing a detailed AI specific project management scope that tackles a business problem worth solving, and iterating your way to success with successive small, narrowest solutions so that we can reduce the amount of variables, the quantity of data that's needed, the resources and additional things to get to a potential solution the fastest with the greatest chance of success. In this way, AI is providing the value we are all hoping and expecting it will.
(Disclosure: I’m a managing partner of Cognilytica and co-host of the AI Today podcast)
