Lab to Market: Commercializing AI
Where can Product Managers add differential value while bringing AI to market
Bridging the gap between groundbreaking AI research and real-world applications is a challenging yet rewarding endeavor requiring a unique blend of market insight, product vision, technical knowledge, and leadership.
As a product manager with a background in data, AI/ML, and strategy roles in startups, and tech giants, I've had the opportunity to bring frontier AI to market. In this blog, I'll share learnings from commercializing frontier AI at Apple Intelligence, Uniphore (a $2.5B conversational AI unicorn) and Samsung Research US.
What differentiates commercialization PMs?
Counterintuitive to a traditional product manager role building tech to solve existing user problems, a commercialization PM's goal is often to create a new market/product from breakthrough tech/research.
To do so, PMs identify "user wants" that can become monetizable user needs. This transforms user behavior at a foundational level - hence the GenAI hype triggered by ChatGPT, is compared to internet or iPhone debut.
AI Product Lifecycle
Based on my experiences, I divide an AI product lifecycle into 3 buckets:
AI stack - focused on developing and improving AI algorithms and models
Production stack - focused on infrastructure and systems to deploy and serve the AI model
Real world applications - the deployed AI model applied on real-world applications and services
A successful AIML PM adds value to all 3 buckets but can add differentiated value in identifying real-world applications.
Key superpowers of Commercialization PMs
1. Identifying commercially viable AI Research <> GTM use-case fit
A superpower of commercialization PMs is the ability to cut through ambiguity to identify high-impact use cases that customers want i.e research -> market -> product fit.
Tactically, this looks like a lot of data and market analysis to identify use-cases and then prioritization of a high-impact use-case considering ROI, business goals etc to ensure product-market fit.
For eg - As a visual intelligence & search PM at Apple, I identified use-cases that would make augmented reality (AR) experiences mainstream from camera and search, for everyday users. Here’s how:
Create a GTM hypothesis: First, I defined what qualifies as “mainstream” - eg entertainment/utility or high-frequency scenarios best solved through AR
Market research & validation: Then, prioritized 2 verticals after qualitative & quantitative analysis
brainstormed 40+ painpoints across 8 verticals (eg education, retail, manufacturing etc) to prioritize 2 verticals
backed the prioritization with quantitative proxy data eg “what categories of search queries would benefit from AR as an output?”
Prioritizing use-cases: Within prioritized verticals, identified specific use-cases that qualified the “mainstream” definition. Then, scoped the MVP for those
Prototype and iterate: Lastly, hacked with a small team of researchers and engineers weekly to build AR prototypes and evaluate feasibility.
This, in turn, led to some fascinating breakthroughs in AR interaction feeding into research improvement as a feedback loop.
Rapid prototyping becomes crucial in evaluating market potential and feasibility of research before investing significant resources. Close collaboration with researchers to understand strengths, limitations, and real-world applicability was a game-changer.
Given the trillion $ market-cap of Apple, the goal was to simply “surprise and delight”. However, monetization is often an end goal from commercialization.
In hindsight, the skills I cultivated during VC and MBA helped me ask the right questions to drill down to what matters most when assessing different markets and what makes particular innovation relevant to re-define a use-case/vertical.
2. Overcoming Technical Challenges
After validating the MVP, the focus shifts to scaling and commercialization. While AI research may yield promising results in controlled environments, real-world deployment can be challenging and expensive. As tweeted:
Technical challenges are across data availability, quality, model performance, and infrastructure.
Data availability & Quality challenges -
AI models require large, high-quality datasets for training. Most AI startups today crawl heaps of open-source data from internet for model training. But, companies struggle to acquire enough quality training data slowing commercialization. For eg -
At Apple, I realized the industry level challenge hindering AR progress was not a lack of a GTM use-case but lack of 3D assets to render AR experiences. To build a collection of 3D assets at scale, I proposed a data generation strategy across hiring, partnerships and generative 3D tech investments.
At Samsung Research US, an entire computer vision model trained only on synthetic data showed high accuracy in controlled testing but struggled in real-world testing. Hence, we ran targeted QA tests and user-studies to gather real-world data to fine-tune the model for realistic performance assessment.
Model performance & Infrastructure challenges -
Ensuring models perform well on real-world hardware can be a challenge.
As chief of staff to CTO at Uniphore, I knew one of his top goals was to reduce cloud and compute costs. This led to a significant number of build vs buy decisions and evaluating multiple partners for deployment, fine-tuning (eg, Anyscale, Galileo) to simplify ML operations for an enterprise AI platform.
Similarly, at Samsung, the team made significant advances in a computer vision solution for detection and labeling. However, it was critical to assess model performance on mobile devices with limited compute power. To accommodate that, we switched to Tensorflow Lite.
3. Building the Right Team & Culture
Commercializing AI research demands organizational changes and cultural shifts. Research orgs and commercialization orgs operate at very different speeds and both require an independent culture to be successful. A good PM can be invaluable in being a bridge to ensure rapid development.
For eg - When Uniphore acquired two startups (one being a research startup), it became crucial to rethink company culture and proactively address organizational resistance. Hence, an in-person offsite with all leaders across India, Israel, US was instrumental in addressing current challenges, communicating the value proposition and building camaraderie among stakeholders.
Conclusion
Commercialization PMs can unlock the full potential of AI by identifying promising AI research, building MVPs, scaling/commercializing effectively, shaping organization cultures and staying ahead of emerging trends. The ambiguity and creativity involved in bringing next-gen tech to market is what makes it so exciting.
Having tackled the challenges of contextual computing, it was rewarding to see user delight on Vision Pro. It showcased the potential of new markets created, industries transformed and power of re-defining user behavior, as tweeted :
While there are infinitely more problems to solve and more to unpack (e.g. commercialization strategies, legal/ethical constraints), I hope this blog sparked inspiration for product managers and AI enthusiasts.
Stay tuned for more practical tips, case studies, and thought-provoking discussions. Twitter DMs here.