Product Management of AI Products

Managing AI Products presents unique challenges.

The traditional product management works well for Web, Mobile and SaaS products. The agile process of wireframing, prototyping, writing user stories, feature prioritization, backlogs, development, sprint cycles, and continuous deployment do not translate well to AI products.

AI product Management

data - metrics - model - result - deployment

• Metrics

• Understand what is outcome that you trying to achieve (predicting a single no, predicting probability, grouping data into binary or multi-class, grouping into clusters or decomposing data)

• predictors

• Data

• Features - if traditional machine learning models, you would need to define features

• About 90% of the work will be spent on data

• Model

• Generally multiple models(algorithms) would be used (Reinforcement learning, NLP, Image/sound/emotion, GANs, Deep learning)

• Explainability of your machine learning prediction to humans to build trust [explainable AI - Deep explanation, model induction, machine teaching and recomposability]

• Result

• Confusion matrix

• ROC

• Deployment

• QA

• Scaling

• Post-deployment feedback

Prioritization

• ML Impact and user impact

• outcome mapping

Sprints

Generally you would not have sprint cycles.

AI Product Roadmap

• Research [Models, Optimization Algorithms]

• Engineering [Data Integration, cleaning, labeling, Trust and security]

• Product [Workflow integration, Continous dataflow, Quality checks]

Programs ==> Models

Debugging ==> Training

Patching ==> Retraining