Latest news

Product Management for AI Startups

No comments

How To Introduce AI Into Your Web and Mobile Products

When most people hear about Artificial Intelligence, they think about Vicki (Voice Input Output System) from Small Wonder (my favorite show).

Vicki has a set of capabilities = AI = { NLP (Natural Language/Speech)+ Capability to Learn from data and Predict (Machine Learning) + Super Powers (Deep Learning = vision, image processing etc) }

When we are talking about building AI-based applications, we are usually talking about using one of the above technologies (although all 3 technologies can cross into each other’s domain) to make our applications more useful for our users. Sometimes, usefulness can mean faster, smarter, contextual, easier and other such things.

Product Management for AI is not that different from building other software, there are simply more set of things to think about. Here, I will try to cover some basic things we have to think about when introducing AI to our software (Web, Mobile or IOT applications).

So first things first — you need AI, but for what? What are your problems? How can AI solve your problems? These kind of answers usually are part of your Business Strategy.

For example, if your business strategy demands innovation, then you might want to consider AI as a strategy. Thing is you might be dead as a business if you don’t adapt. So you might need to consider AI just to survive. For example, if you are a dating site that makes users spend hours of time without finding them the right date, then you face threat of a new app putting you out of business. So, you might want to use Machine Learning to to Analyze your users photos, profiles, descriptions, etc to generate a model that can best predict the the right date for your users.

If you are new, you can start thinking whether your application should be Bot-first or AI-first or Mobile-first etc (Product Strategy). for e.g. you can simplify food ordering by building a bot which needs just a couple of taps to order food (since it sits in the messenger), but whether you want to build a bot for it or not is a decision you have to take based on your goals (For e.g. adoption, retention and revenue goals).

If you are an existing enterprise, then you can use one of the various AI technologies to improve your User experience, to solve deeper problems of your users, solve new problems for your users or solve problems that your users will have in the future.

Continuing with the same example, if you are a Product Manager in existing Online Dating Business, you might want to detect fraud profiles and scammers on your dating website , then you look at the historical data and find out the likelihood of the profile being a fraud (scam/spam). You start by looking at few features and decide “ok, this is how we will decide if the profile is a fraud” which is nothing but an algorithm that is testing for the likelihood of spam based on certain parameters (factors). This will be a classification algorithm.

NLP vs ML vs DL — Or A Combination of Them All?

Regardless of your project or size, here are some basic questions you can ask yourself before picking an AI technology:

  1. Do I want to add Natural Language capabilities for your Application (, Amazon Lex, Pullstring, Lex etc). This includes Speech to Text and Text to Speech. (for e.g. for building assistants like Alexa, Google ALLO etc)
  2. Am I adding Machine learning intelligence to my app? For e.g. training data sets to predict outcomes for your users ? ML can be applied to almost any industry to do things that were not possible before. For e.g . finding the right job for the right person, predicting Patient’s diseases, predict house prices etc.
  3. Or am I trying to apply Deep Learning AI for a domain? For e.g. applying Vision Capabilities to Cars to make Self Driving Cars or to slowly introduce Smart features into my Cars.

Once you have decided to use AI as a part of your strategy, depending on whether you are a startup or enterprise, you will…

Pick an API to Achieve Your Technical Goals

If you want to add NLP to your application, then you will have to pick an NLP platform to design your conversations (few api’s mentioned above).

Things to consider when picking NLP platform

  1. Try it out. See how it deals with composite entities, whether it even allows it. As a Product Manager, you have to know the basics of NLP, otherwise you will struggle to build applications with medium to high complexity.
  2. Write your own NLP algorithms to cover for the cases your platform can’t cover. You can start by reading some basics online.
  3. Flexibility (to incorporate other platforms)
  4. The roadmap and background. What if you build a business around it and it costs too much or the project is shutdown by the parent company?
  5. Plan for various contexts your users will use Natural Language while using your applications
  6. Plan for failure and always have quick reply or route ready to deal with such cases.

If you are adding ML to your application, then you need to find your Initial Data Sources.

There is no AI when you start your application, but you want to use as much of an AI possible to start your application (begin with as much data as possible). So you need your first set of data and possibly grow the data based on user inputs, interactions.

Build a Candidate Model:

  1. Pick the Initial Dataset (for e.g. for this dating site, the millions of users profiles from various countries, their support data, conversational data etc)
  2. Preprocessing data (Cleaning, normalizing, removing empty sets/incomplete sets). This is important to reduce errors as well as improve accuracy of your results. This step also gives you an idea about the richness of your data, what kind of results you can expect, feedback about your dataset (so that you can pick better data sources).
  3. Select a feature set or factors based on which you will train your Model. For e.g in Dating Application, you can choose time to reply, time spent in the app, no of messages sent by the users (in each segment) as starting points to reach your ideal results.
  4. Apply learning algorithm to the data — Find Candidate model and iterate. for e.g. if you have site such or, you can apply regression algorithm to extrapolate and estimate the price of a car or rent of a home.

5. Deploy the chosen model to your application

6. Measure Results and go back to optimize the feature set or Algorithm, or may be even pick new feature set and algorithm altogether (if the results are not as expected)

We see the results, if we like it, then we mark the model as good and fine tune the algorithm to make it better. If we don’t like the results then we pick a different kind of algorithm and alter our feature set and see the results again.

Things to Consider as Product Manager’s Building AI Products

— Choose AI platform and API’s (based on costs, your type of application, future of your application, future of the platforms and other such things) (use tools such as cooperative Intelligence to compare the costs of API’s for your AI app.

— Good thing is you can start from scratch, Add AI, Remove AI, Update AI, but you need to manage your Product lifecycle. You can’t suddenly introduce AI one day and change it another day, it can lead to bad user experiences.

That’s why I believe collecting feedback data when you launch your AI startup is crucial. We want to be able to retract certain features that aren’t working properly before everyone uses your apps. That’s why big companies will use AI slowly. That is why google isn’t pushing its voice apps, because it has a big strategy to slowly introduce AI into our lives.

— Similarly, you have to think about how much of your App will be AI-driven and how much will be manual or usual software(for e.g Getmagic APP was completely manual in the beginning, but they now have tremendous amounts of data to help them automate the incoming user queries.

So your New Product Roadmap will have a mixture of these going along with your Features RoadMap. You might call it the AI ROADMAP that runs in parallel to your Usual RoadMap

— We should think if we are increasing the number of steps or decreasing the number of steps for a user (for e.g. voice support decreases the steps drastically when using Google voice search)

— Sometimes the value of what we are providing is very high, so even if we are increasing the number of steps, it’s ok. Got an example? Write in the comments

— We have to think about the barriers that exist for users to use it (privacy, not personal, personalization for a certain demographic, cultures, etc)

— Think about hiring a person from humanities background. Most bots are still not used by people because they are built by engineers. It’s important to build for people in the end.

— You are always launching something new in AI (because the field is new), the incremental launches need to be cohesive. This takes a ton of planning and working really closely with the engineers to know what is possible and what is not.

— I say keep a “Roll back procedure” ready and a letter from the CEO also, in case there are potential disruptions because of the new experiments.

— Everybody will tell you that you need a context for your AI bot (and even applications) and it’s the hardest things for your software application to learn. Spend a little bit of time writing down contexts and also write plans and flows to deal with each of them.

— What if your Machine learning engineer leaves in the middle of your startup? Can you afford to hire another one? Most universities in India have now started incorporating ML and AI in their Computer Science courses. Be sure that the demand of Engineers will be met!

How should you price your AI applications? You can pick various kinds of models. May be you will not charge for the software at all and only charge for the AI part of it or may be a combination of both or may be operate a completely different business model (such as the value of data as business model).

— — How are you going to communicate your new AI based feature to your users? You don’t necessarily have to say it. Google and amazon have used this for years. In fact that might be a better way to go about it. Amazon slowly introduced Alexa as a harmless addon in the beginning. It didn’t make a big deal about it, although Alexa might be Amazon’s biggest strategy in the coming years.

If you are B2B App, then your users might like that you are using AI and hence you might be able to retain your users for a longer time. You might be able to reduce the time it takes for businesses to do certain tasks, or the amount of people it takes for to do certain tasks. Businesses will simply love this.

— What are the key metrics based on which you will evaluate the success of your AI ? Goals achieved? Improvement in key metrics? Overall Market reaction?

Coming up with MVP of your CANDIDATE MODEL is Product Management for ML Applications.

How we are using machine learning in Faq Dog?

I have described here that we were building bots at faq dog, seems like noone cares about using bots (yet), but we are close to launching a SMB bot (with more a button based approach) and then add NLP to it by using PullString (previously we were using

As a part of the same efforts, we are very close to launching a QnA based conversational approach to get feedback from customers, users in App and Website. Here we try to predict what the user is wondering about when he/she visits your site, try to answer their questions (by starting with answers) and also help them use applications better. It’s a proactive way of getting feedback and exploring deeper engagement opportunities.

Here we use ML to find which question should be asked next (which answer to present first). We want to anticipate the question on the customer’s mind and start with the answers. So, we pick a dataset (set of features/factors ) that will be used to decide the next question (for e.g. age of customer, recent sign, cancelled customer, refund status etc etc) . Then we will create an algorithm or pick an algorithm, based on those parameters and dataset, and analyze the results that we get.

The algorithm basically is a model (a framework) = to decide about our goal (to find the next question).

Ok, this post is already too long, I am going to shut up 🙂

We will keep you posted about how it’s going and post details about our Model Building Process in future articles. We also plan to cover Product Management for Bots in depth in future. If you found any errors in the article, please point it out in the comments. I am learning ML since last 1 year, and I can use your inputs about how you are using ML in your applications.

If you want to learn and write about Product Management, please do so here..

Product Management Life

Couple of Resources to explore and Learn about AI and ML:

Product Management for AI Startups was originally published in Chatbots Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Chatbots Magazine

magnoliaProduct Management for AI Startups

Related Posts

Leave a Reply