How Product Managers Can Adapt Agile for AI

Main idea: Businesses are moving beyond the exploration phase of AI/ML and increasingly integrating AI/ML features such as predictions and recommendations into their production software. This entails collaboration between software development teams that typically use agile processes and frameworks with data science teams that are usually unfamiliar with agile. Product Managers can play a key role in adapting agile processes for AI powered products to drive successful product development outcomes.

Agile software development: Agile refers to the practices adopted by software teams that focus on satisfying customer needs through early and continuous delivery of working software. While practices vary across companies large and small, teams typically work in 2 to 3 week long increments or sprints and aim to deliver functional software at the end of each sprint.

Data Science workflows for product development: A high level data science workflow involves problem definition, data collection and cleansing, data exploration and modeling (or use of MLaaS) followed by iterative inferencing and retraining of models. When the intended outcomes from such workflows are for business reporting or are intended as Proof of Concept AI projects(PoCs), data science teams can function in collaboration with the product manager largely independent of the extended software teams. On the other hand, after the AI/ML POCs are successful, integrating these AI/ML capabilities into production software applications at scale becomes a key imperative for product teams. This makes it critical for data science teams to adapt their workflows to the typically mature agile software development processes and a product manager can play a key role in this given the benefits that this accrues.

Actions product managers can take to positively influence outcomes:

• Educate data scientists about the basics of agile: Data scientists lack familiarity with software project tracking and release tools such as Jira and are unfamiliar with work estimation practices such as story pointing. Product managers can invest upfront in light weight efforts on common tooling and practices. Allow for a "run in"period in the planning for data science teams to gain traction and be flexible.

• Define "Done" for AI/ML appropriately: Defining "Done", a key measure in agile software development for measuring progress at the end of each sprint, is 'relatively' straightforward (e.g. "Create a functional Download button"). For data science projects, product managers should define "Done" in terms of what constitutes as completion of a data science experiment. This will instill confidence in the data science team to leverage the benefits of agile processes while accommodating the uncertainty associated with the outcomes from data science projects.

• Mitigate risks and actively communicate dependencies across data science and software teams: Encourage data science teams to adopt practices such as Point of Views (POVs) and Proof of concepts(POCs) as part of sprints to gain better visibility into sizing problems. Communicate dependencies with software development teams and plan for mitigating them.

• Set expectations with key business and technical stakeholders: Evangelize the potential benefits from integrating AI/ML features into production systems while also highlighting the potential uncertainties that these entail. Machine learning operations( MLOps ) is an emerging field that aims to unify the release cycle for machine learning and software application and will likely mitigate many of the early challenges over time.

Home