Why Developing AI Products Needs Agile on Steroids

Artificial intelligence. It’s all the rave. I remember walking into the office at the start of the year as ChatGPT was being rolled out. My colleagues were exchanging experiences about using ChatGPT to summarise a 30-email thread to five key points or get some inspiration for an upcoming presentation. It is quite impressive.

But AI doesn’t stop there. It will continue to reach deeper into our everyday lives, challenging many norms and hopefully being a genuine force for good to help us tackle today’s challenges, arguably some of the greatest humanity has faced.

So, what is important to keep in mind when developing AI products? The short answer is that you need Agile on steroids. Consequently, and similarly to agile product development, the key factors are absolute clarity on where you stand and where you want to go, realistic and adaptable timelines, and user centricity, including feedback.

Defining the starting points

We have to deal with a relatively high degree of uncertainty during the planning process in software development. That is why in Scrum, for example, we try to keep iterations, or sprints, short. Add AI into the mix and it becomes even more challenging to plan work with certainty since causality in AI is not rule-based or linear.

While in more traditional software development, adding B to A will get you C, in AI this simple combination can produce hundreds or thousands of outcomes because there is no singular B. Instead, there are multiple Bs that have a different level of relevance depending on A, but also multiple possibilities of Cs. After that come Ds, Es, etc., all of which depend on the context of other letters or factors. In other words, the complexity seen in AI is far higher due to the learning nature from interrelationships between data.

To deal with this, as a leader (and as an entire team), you must always be clear on where you stand and where you start. Define your objectives unambiguously, and account for all considerations and assumptions you make. An agreement such as the definition of ready (DoR) takes on a compelling role in guiding you. It makes you aware of the assumptions (lack of actual knowledge) you are dealing with and, therefore, understanding the risks. Consequently, remaining ready to adopt new scenarios and solutions as new learnings become available is essential to stay on the right track.

Development timelines: Keep them short!

Now that we know that the planning process in AI product development requires many assumptions and deals with a high degree of uncertainty, there is little point in planning AI projects in great detail. There should, of course, be time horizons. But going for detailed planning is a waste of time because it will inevitably lead you to miss targets you set very soon.

Instead, ensure that you deal with complexity by shortening sprints, having a stable team composition, and encouraging the adoption of new skills and techniques as an investment into delivering faster. Don’t plan a fixed long-term scope. This time is better spent to get users to review your results.

Reviewing the results is more important than ever

AI is there to generate results. The overall objective is that the results can match or exceed those humans produce. For AI to know the right results, it needs access to data and (initial) human input.

In the process of developing a machine learning model, you and your team will be faced with numerous instances where you need to choose a way to go. While you should encourage experimentation, the results must be reviewed to determine whether there has been an increase in quality and relevance. This needs to happen regularly, even daily, where applicable. Relying on the logic of system demos and doing reviews every three couple of months, while problematic in many contexts, will not cut it in AI.

Remember: You keep making decisions, assumptions, and choices during the development of an AI model. If you review after only the 50th decision, you could have been going down the wrong path for quite a while. As a leader in AI product development, don’t try to avoid ‘negative’ feedback. Mostly, it is there to help you along the way to your overall objective.

Where do you go from here?

AI product development is very similar to traditional software development but with a higher degree of complexity and an even greater need to have clear objectives, plan and be ready to replan, review with users and customers, and learn as you go. While traditional software development may allow for some slack in living the agile mindset and practice, AI product development is much more likely to punish the inability to inspect and adapt.

 

Photo from BlackJack3D / istock

Geschrieben von

Tihomir Vollmann-Popovic Tihomir Vollmann-Popovic

TEILE DIESEN BEITRAG

Kommentare sind geschlossen.