Click to learn more about author Vinay Seth Mohta.

Thanks to Artificial Intelligence, organizations are tapping into troves of data to improve decision-making processes in a way that data managers of yesteryear could only dream about. As such, many in the business want to capitalize on AI as quickly as possible. Lean AI delivers such value to these businesses.

AI implementation in general requires the significant, coordinated effort of a diverse group of teams that are siloed in many organizations. When we started working with clients, we quickly found that it takes a herculean effort to round up business stakeholders, domain experts, developers, and data scientists and get them working seamlessly together. Perhaps you’re familiar with this problem from your own internal projects. We knew we needed a specialized process to keep team members aligned and working toward the same goal. The result was Lean AI.

Lean AI is a new, innovative practice and its principles should be widely recognizable. A number of existing systems inspired us in the development of Lean AI, including human-centered design at IDEO, the Lean Startup methodology, Agile Software Development principles, and the CRISP-DM approach pioneered by the data-mining community. Data executives — knowingly or unknowingly — apply many of these principles within their teams already. We just put it all together in a process and focused it specifically on AI.

Over time, we’ve refined Lean AI into six steps that minimize risk and focus on delivering clear results as efficiently as possible. While no step is inherently revolutionary, the specific order is thoughtfully designed to manage risk while quickly producing tangible results. You can try out the process on your AI projects and see the results for yourself (or, better yet, prove the results to the other execs at your company).

1. Understand the Problem — and then Understand the Data.

The first step in any AI project is understanding the business problem inside and out. Then and only then should you turn your attention to the data at hand — otherwise, you won’t know what’s high-quality data and what’s meaningless. High-quality data is the fuel for an effective AI solution; this step is about making sure that fuel is at the appropriate octane level before using it.

This first step typically involves workshopping to pinpoint a problem that’s solvable with AI in a reasonable timeframe. Then, you can figure out exactly what kind of AI is called for, what data is required to create viable solutions, and how long it will take to put the solutions in place.

2. Engineer the Framework.

In this second step, you’re preparing for the actual build that will solve the problem you identified in step one. A key practice for us is using Docker to capitalize on containerized Data Science, which delivers a tidier and more collaborative developer flow.

Your delivery team should include dedicated data engineers and dedicated machine learning engineers — you’ll need both skill sets, but it’s extremely difficult for employees to master both fields at once. Creating data pipelines is vastly different from creating machine learning models, and by bringing them together early on, you’ll enable team members to think about the final deployment with each early decision they make.

3. Create Baseline Models.

Once you have a solid foundation for your data in place, you can begin to construct models. But instead of getting worked up over the limitless possibilities, focus on the minimum viable product. Ask yourself how you can solve the problem you’ve identified while doing as little as possible. If you can deliver the core features that solve the business problem, you can worry about adding additional features later.

We truly believe that the most efficient problem-solving happens at a baseline level, and Lean AI revolves around this approach. Producing shippable models is the goal, so each step of the Lean AI process should be aimed directly at deployment.

4. Workshop for User Feedback.

AI is a powerful tool, but it still works best with a human being at the helm. Whether your end users are maintenance technicians or marketing professionals, try to get your solutions in front of them as quickly as possible.

The user feedback step is instrumental in the development of the user interface. Because raw predictions aren’t easy to work with, we typically need to develop a type of recommendation engine, in addition to any other tools necessary for post-processing, to aid interpretation. This fourth step also allows the AI model to begin earning user trust as soon as possible, as many people — especially those who aren’t well-versed in data — are skeptical of AI.

5. Deploy the Solution in Production.

Deployment is the ultimate goal, but one of Lean AI’s strengths is that you’re thinking about deployment from the very first day. This forward-thinking focus means working with the software and operations teams to determine the most important integration points, as well as the data requirements for your solution.

Ultimately, the upfront work you do in step two is the most important aspect of a seamless deployment. It’s easy to overlook, but developing clean interfaces for integrations both upstream and downstream will make the difference between a successful deployment and a nightmare.

6. Validate the Final Product.

In this stage, observe how users are interacting with the final product without actively soliciting feedback. If you’ve done your due diligence in step four, there should be few surprises. As you monitor model performance, one specific thing to watch for is whether biases from the training data are resulting in the model performing differently in production. If necessary, it’s possible to go back and make changes.

As you work with your team, offer as much support as possible in this step. It’s necessary ensure the clients’ internal teams have all the resources they need, we believe it’s important to pay particular attention to training your employees. A tool is only as capable as the person who wields it.

AI can be incredibly powerful, but it can also be complicated and difficult to develop within a large company with many stakeholders. Lean AI is about harnessing the strengths of different people on your data team to provide value in a quick and consistent way. If you’re struggling to get an AI project off the ground, switching to a leaner process might be the jumpstart you need to succeed.