Thank you for getting in touch, we'll get back to you soon!

Do you have some insights to share with the product community?

GET IN TOUCH.

Building products with Intelligence

Written by
Anant Kapoor Anant Kapoor
Head of Product @ Raft
Published
Share
Software has eaten the world. It is now AI’s turn at the dinner table.

Before WeWork became WeCrashed, Adam Neumann was painting an ambitious future for the company in an interview with Forbes. When he threw out the phrase Artificial Intelligence, I couldn't help but chuckle. Just another buzzword. What did AI have to do with the future of coworking?

His next few sentences wiped the smug smirk off my face. There was a lot. From data driven selection of property to optimising coworking space and empowering architects such that they designed offices in five days what would have otherwise taken thirty. Whether or not it ever materialised, I was convinced by the possibilities.

Throughout my career, I’ve gone through this arc a few times: hearing the buzzword, feeling the cynicism and then becoming a convert as we build something magical but far more limited in scope.

At Masternaut, we tinkered with vast amounts of IoT data, continuously collected from devices in countless cars, to categorise driving patterns and identify exceptions such as crashes or stolen vehicles. We identified the most dangerous roads in the country and could detect whether a driver stopped to drop off a delivery or was just idling at a traffic light. At Improbable, AI helps create richer simulations for decision makers. And at Vector.ai, it is not only an integral part of the domain name but a pillar in our ability to deliver next generation automation to the freight forwarding industry.

Despite the limitless potential, I have found AI to be both an unhelpful term and an unwieldy tool. It comes with endless connotations and unrealistic expectations. In practice, it is just a series of messy algorithms hacked together. The data requires endless finicking and never seems to quite achieve perfection.

When done right, however, you can get some astonishing results. Over the last few years, I’ve learnt a few principles that help me think about how to build useful products that leverage AI.

Start with the user.

Few companies have the luxury to not start here. If you are building the next openAI or deepmind, you can disregard my advice and start with the technology. They solved challenges (like chess or go) that displayed the feasibility of their technology to secure funding to commercialise user-facing solutions in the future.

For everyone else, this is a hurdle that will hit you hard in the shin if not taken seriously. Often teams get carried away by the depth of insight they can produce. Just because you can do something with the data does not mean you should. Users care about experiences and outcomes not technology.

Resist the allure to build fancy dashboards that deliver game changing but one-off insights. Decisions like these are better served with projects not products. The software world is littered with decision support products that promised to help CEOs which country to launch next in.

When it comes to automation, the things I check are:

  • Is this an important decision?
  • Is this decision made on a regular basis?
  • Is this decision made by many users?

For example, everyone agonises over what to watch every day. A great decision to automate. Imagine the human potential it could unlock! On the other hand, the c-suite tweaks the company budget once a year – an important decision, but much harder to automate with a product.

Choose many AIs over one

Rather than intelligent, most algorithms are optimised for one thing. The head of product at Deepmind once told me that the best solutions are usually many strung together in a seamless way, such that the user has no idea what’s going on under the hood. The algorithms used to search for a website are very different to another performing the same function for videos or photos. Google just forces the user to pick what they are looking for.

This framework always stuck with me. Instead of focusing on a general model that solves everything, I like to constrain the problem as tightly as possible. Optimise for one metric that only touches on one user requirement. Then find the next one. Generalise only when necessary.

Tightly couple user input with your AI datasets

Switching back to the world of Netflix, their product will constantly ask for feedback. Hungry for ratings to fuel their algorithms. The quality of their learning is only as good as the quality of the user input.

However, it is easy to avoid pressing a thumbs up when it pops up on the screen. It works better when the data collection is an integral part of the user experience. It is less easy to not click on the most useful google search result. Giving you better results that you can click next time, it fuels the virtuous learning cycle.

Experiment like a scientist

Karl Popper said that a hypothesis can never be proven but only falsified – the role of a scientist is to proliferate the number of plausible theories and then systematically scratch off as many as possible using only empirical evidence.

I am not quite that radical when it comes to product management. I am also a little less radical than Marty Cagan who says that testing your ideas with real users is probably the single most important activity in your job as product manager”.

However, I do not believe that this can’t be done for complex technologies or AI products. Testing the value of data can be done with as little as a shared google spreadsheet. If the user checks the spreadsheet regularly, it is likely valuable. If not, you’ve saved yourself a lot of development time. Just like that, there are many ways to hack solutions that test your ideas before having to build elegant algorithms that do not get used.

End note

Hopefully some of the above frameworks help guide the way you go about building your next product or feature. They are by no means exhaustive and biassed by my experiences. If you think about developing these products in other ways, I would love to get in touch.

Share

ABOUT THE AUTHOR

Anant Kapoor

Anant Kapoor
Head of Product @ Raft

Related articles

Insights and thoughts from leading product people.

Looking to build your product team or find the perfect role? Let’s chat