Machine Learning Won’t Save Your Startup
A common situation: Things are going fine, but not great at your SaaS business. Your sales team’s win rates aren’t quite high enough. Your marketing pipeline isn’t quite as full as you’d like. Your customers are happy but some are considering other solutions. Overall, your product doesn’t quite feel differentiated enough.
A natural temptation that I’ve seen in this situation is to look to machine learning as the solution to your differentiation woes. If we just sprinkle a dash of the ol’ ML on this bad boy, the thinking goes, we’ll have a product that stands out in the market and that everyone will love. In general, this strategy just doesn’t work.
Machine Learning Is Not Unique
First – at this point in time, everybody vaguely knows what machine learning is and has a rough sense for its capabilities. You’re not getting a jump on the market by declaring that you’re going to make your product “powered by AI.” Every Gartner report has some checkbox about building intelligent or predictive features, and it’s no longer a secret that there appears to be some magical pixie dust out there that you can drizzle on your product to make it special.
Using machine learning to differentiate your product is like driving a fancy sports car to stand out when picking someone up for a date. It can be cool, it might even fit your persona, and some people will be impressed. But ultimately it isn’t revolutionary or inherently game changing.
ML is a buzzword of the moment – and investing in buzzwords is not the route to enduring differentiation. There’s always some technology that’s gotten enough mindshare that everyone is sharing blog posts, frantic manifestos are being written, and investors are hot and bothered. 2 decades ago it was cloud, it’s currently (roughly speaking) machine learning and crypto, and who knows what it will be next.
This doesn’t mean that ML features are useless – if it’s useful, people will pay. But there is essentially no novelty left in the ML play. Even if the underlying trend is meaningful (for example, cloud transformation was and continues to be a big deal!), the biggest trends rapidly become table stakes.
A Lot Needs to Go Right for ML to Work
The barrier to building a profitable, differentiating ML-driven product is high – not only in a technical sense, but also in terms of the role that it solves for your business. It can’t just be slapped on top of your product like salad dressing:
You need to have an ML application that fits your business model – for example, if ConvertKit (which Stay SaaSy uses for our newsletter) adds a crazy send-time-optimization product, that might simply not matter for sites like us that use them to email out blog content
You need a problem that you can solve but others can’t. In reality, you are probably not orders of magnitude smarter than your competition
You need to solve a problem that has a tangible business impact – only the most frivolous buyers will purchase something just because it’s cool technology
You need to actually prove that whatever ML magic you’ve built actually solves a key problem better than anyone else can, or for that matter solves a real problem at all
A situation that checks all of the boxes above is the holy grail, but you need to be honest about whether that’s the case. It’s very possible that you’d be better off trying to differentiate your SaaS product by creating a suite of functionality or investing heavily in UX – both strategies that many companies have used to construct defensible product moats.
The concept of provable value is one of the most unknown or unappreciated elements of building a sellable ML product. The more that you can prove that you’re adding revenue or reducing costs, the stickier your revenue will be. When the rubber hits the road and something needs to get cut, the products that add value in a provable and (ideally) deterministic way are the ones who survive.
Many customers want to verify the results of anything that they’re paying for, and black boxes understandably scare them. Hype has created a litany of startups peddling ML snake oil and buyers are rightfully skeptical. This increases the barrier to entry and means that it’s much harder to build an enduring ML product. Not only does your product need to work, but you often need to build significant reporting functionality that allows customers to dissect your product and verify the advantage that you claim to provide. And keep in mind that many customers will approach the problem of analyzing whether your algorithms add value adversarially – teams that wanted to build ML products in-house rather than buying yours will be especially aggressive critics and naysayers.
The Promise of Machine Learning as a Silver Bullet is Distracting
Perhaps the worst part of hoping that you can dust magical machine learning on top of your product is that it trains you to look for silver bullets. ML feels like a new, magical solution that will solve all of your problems. In reality magic solutions are near-mythical and it’s dangerous to believe in them.
The earlier your startup, the worse the optics of saying that you’re going to dominate your market by sprinkling machine learning on your product. If you have terabytes of unique data and claim that it will unlock the portal to Machine Learning Narnia, or if you have a unique approach and track record that indicates true expertise, reasonable listeners will hear you out and not immediately assume that you’re full of it. But if your non-specialized team is trying to raise some kind of seed fundraising on the back of a machine learning story (“We’re just going to do X, but with ML”), then it should be obvious that you don’t have anything vaguely proprietary. At best, you look overly optimistic; at worst, you look like you have no idea what ML actually entails. Unfortunately, some people will pump you up and make you believe that this strategy is viable – the rest will roll their eyes at your “.ai” domain once you leave the room.
Why does this happen? I think it’s because the talk track “we’re going to do X but with AI” does work on a certain kind of unsophisticated observer. If boasting about ML products is like driving a McLaren, there are indeed investors / buyers who are the equivalent of someone who is really, really captivated by a fancy car. You can’t count on this to be an enduring advantage.
Where Machine Learning Matters
Where I’ve seen machine learning make the biggest difference is in extending a lead that you’ve already earned. An example of where ML can really set a product apart is Photoshop’s ability to do automatic skin smoothing and sky replacement. This is some hardcore stuff, and Adobe has checked the box on the narrow domain where ML products can make a huge difference:
They have the data and team to solve this problem well
They have a platform (Photoshop itself!) where automatic image manipulation software can be plugged in to add significant value
You can actually observe that their technology works.
Photoshop already exists as the most sophisticated graphics editing software, so adding more sophisticated features on top of it extends their advantage on the market
Due to the technical moats that it can create machine learning functionality can be a great accelerant when it works well. But you simply can’t rely on it to be the singular differentiator for your business.
