AI through a product manager’s lens
If we observe Gartner’s Hype Cycle for Emerging Technologies 2020 closely, in 10 years horizon there are many AI/ML technologies edging up the hype curve. Some of them like GANs have already shown results. Huge pre-trained models(GPT-3) with more than 100 billion parameters, amazed us through interesting demos like this one very recently.
It is high time that product teams across industries take note of this trend.
Here is why:
Let’s start with basics.
A product or a service helps a user to get a job done and to achieve certain outcomes under constraints. There are many frameworks to define success metrics for a product. At its core, a good product should score high in metrics that measure users’ desired outcome. If the product has a good user experience, it will also score high in metrics that measure success of individual tasks that lead to the outcomes. AI could be the technology that is used to make the core product that gets the job done. Or AI could enable improving user experience and help in task success metrics.
Technology as an enabler and as a carrier of consumer insights
There are many examples where technology components informed product teams of insights that came from similar products or services in adjacent markets. Some times there are design components that carry this learning from one product in one market to another.
A typical example is search. There are many elements of product search experience in Amazon that you will find in other e-commerce sites. ( search suggestions, filters) In fact you will find the same in unrelated sites also.
Typically this feature (search) is supposed to help users find what they are looking for as quickly as possible. It is quite possible that the e-commerce site might tradeoff user’s task success for revenue or any other internal metric.
However the point here is that the same features became a standard for the search experience across multiple products. In fact some of the open source search engines (Solr, Elasicsearch) provide these features as part of their package.
Machine Learning is maturing
Let’s discuss machine learning — the popular application of AI based around the idea that we should really just be able to give machines access to data and let them learn for themselves. Recently there has been good progress made in all three paradigms of ML — supervised, unsupervised and reinforcement learning, and very good progress in the first two.
This is visible from the feature launches of big tech companies who have the resources to hire specialists and invest in computing resources to churn out better models.
AI/ML is accessible. Almost democratised
What’s interesting is that even if you are a startup business with limited resources to build ML models on your own, you can access most of these high end models through APIs for a reasonable cost.
Microsoft, Google and Amazon provide API-s for a broad spectrum of applications that product teams can use to build their initial versions.
If you are a startup building a customer support tool, you could use one of these API-s to do autocomplete and make your users (CS executives) efficient. You could also build a classifier to categorise incoming queries for prioritisation. You could even build a bot that can handle low priority information seeking queries and gracefully pass the conversation to a human agent if it goes beyond its capacity to solve.
Let’s say your product reached a scale where you have a lot of transactions. You think some of these features are very valuable and your customers like them. But you want to save on the API cost. As earlier, your resources are still limited. Not to worry, many AI research organisations have made basic building blocks of AI free and open source for the community to build upon.
They have used a concept called transfer learning to generate pre-trained models that can be used to build applications for your domain. For example word2vec, a numeric vector representation of english words, has been available for sometime now. Pre-trained models like GPT-2 and GPT-3 provide a lot of generalisability. Some of these generalisable models and their applications are being released periodically. For example, checkout huggingface
On top of these, platforms/libraries like Tensorflow, Keras abstract most of the math involved in backpropagation and optimisation and give you enough flexibility to customise model architectures for your application. Platforms like Colab give you GPU-s to compute for free for experimentation. In a way, building AI features is within the grasp of any team that has good software engineers with a flair for data.
A few years back, with the proliferation of SaaS products, business software started becoming as user friendly as consumer software. A good UX became a prerequisite for a CRM SaaS product. Similarly soon the UX benefits of AI powered features will make them ubiquitous for software products.
New product categories will also be established by AI. For example Grammarly at its core uses ML. We will be seeing more progress in smart homes, smart office assistants and smart <lot of things>.
Don’t add AI for investors
I have seen examples where product teams add chat bots, where a good search experience would serve the purpose much better. An AI bot might improve your chances of getting investment, but it might frustrate users who just want to search for some information.
Going back to basics, a product is good if it is helping its users get a job done and achieve their desired outcome. AI/ML enables it in ways we have not seen before.
