Dec 30, 2021
Implications on product roadmaps for AI-powered startups
Development on a web app contributes linearly to your value proposition, while development on an AI contributes exponentially to your value proposition. What does this imply for your product roadmap?
Circumstances for promising AI-powered startups
If you want to stand a chance against FAANG but also build an AI-powered product for a large market, you need to pick an area where the data is hidden for incumbents. It turns out, that those rare opportunities tend to have very fragmented data sources.
The implication of this situation is that you cannot go to kaggle and download a suitable dataset or adapt some existing model. So you have to put in a lot of work to build a working AI which adds value to your customer.
Known challenges for product roadmaps without AI
Rodrigo Martinez lays out a very thoughtful Hidden Roadmap for founder-CTOs. And I agree absolutely in the context of companies, where only the User Experience is most important.
If you build a product which doesn’t rely on AI you should follow exactly his guidelines:
- Complete focus on building a working prototype to validate your assumptions
- Start designing your product and organization for scale
- Scale
There’s one source of uncertainty: the customer. You think you know what she wants. You go back and forth a couple of times until you figured most of it out and the you build it. It takes time, discipline and skill to build it, but many people before have built web applications.
New challenges for product roadmaps involving AI
If part of your value proposition involves AI, you get a bonus source of uncertainty: can you actually build it?
I was responsible for putting User Experience (Frontend), Backend and AI together at Coral Innovation. Before I was working on the web application of Papershift.
Based on these two experiences I noticed a couple of differences in the development process.
Development on a web app contributes linearly to your value proposition
When you spend time working on a web application you usually go from feature to feature. In doing so, you’re trying to cover one requirement at a time.
In general, it follows that the price your customer is willing to pay or the number of customers you can acquire increases linearly with regards to your product features.
At Papershift, there was often the situation that a possible customer signaled that she was willing to buy the product if one feature, which was really important for her, was added.
In contrast, spending time working on our technology discovery AI we build at Coral Innovation showed a different course of progress.
Development on an AI contributes exponentially to your value proposition
Exponential growth means that for a long time there is almost no growth at all. At that is pretty much what we experienced. We spend almost 6 months not making ANY progress.
Remember the circumstance:
- Fragmented data sources
- No preceding models
The dataset we used when we started building has almost nothing to do with the dataset we’re using currently for the deployed models. And the same is true for algorithms, features, preprocessing, etc.
There was no best practice, guide or example dataset that was ready for use. And in the early stages our AI was not testable with a customer. The output was so bad, that she would have always said“it sucks”, “it sucks”, “it sucks”, “it sucks” on every iteration.
Minimum viable AI
Maybe you know this picture illustrating the MVP concept:
If a major part of your value proposition comes from an AI, is it possible to actually build an MVP?
You can test the expectations of your customer using real intelligence simulating artificial intelligence. And this gives you a good idea of where you need to go. But it doesn’t give you any directions how to get there.
With all the new, abstract solutions for building AI, you might think that it should get dramatically easier. The tools for building AI are becoming a commodity, but the actual creation process is not.
Implications
Add some slack: It’s very hard to predict when your AI will provide value for your customer. Plan ahead without considering time. If your technical readiness gets stuck at some point, your team members can still continue working on the business investment (social) readiness.
Management of expectations: I was desperate at one point regarding our progress with our technology discovery AI. So you have to keep on looking. For a suitable data set, for useful features, for working algorithms, for relevant quality metrics, etc.
As soon as your customer sees some value: stop. Minimum viable AI achieved, focus on other things before improving AI. We consider AI as a (very powerful) tool, but not an art. We don’t want it to be perfect, but to add value. We don’t care about the last 0.2% accuracy, but about the first 80%.
I am very curious what experiences others have made and whether they coincide with mine. Please let me know!
