Lessons from an AI startup: Seldn.ai
‘In a lot of ways, his failures were more wonderful than his successes.’
— G. Hardy on mathematical genius Srinivasa Ramanujan.
Between 2014–2016, I ran an ambitious predictive artificial intelligence (AI) startup Seldn (yes, named after Asimov’s hero). Here, I highlight five highly distilled key learnings from that adventure. These lessons are also a big part of why we decided to fold. The patterns I observed in my own venture are consistent across the whole AI (software) landscape. Now, when I advise an AI startup or help a VC on AI diligence, these fatal flaws stand out. My goal here is simple: save founder and investor resources by avoiding these repeated traps.
Note: these lessons apply to any AI enterprise startup in software. Some of these will indeed apply to any startup. There is a lot packed into each of these bullets, and some are further unpacked in specific separate blog posts.
1. Be Goldilocks
One of my key learnings from being a founding VC at Shell’s venture capital arm, is that a startup’s product needs to be better than a nice to have. A ‘nice to have’ doesn’t give you enough enterprise pull and cannot survive the lengthy enterprise sales cycles. Surprisingly, an offering that ties in directly to massive enterprise value will not quite take off either. Reason: most big corporations will not trust a new startup product with massive high- $$ value decisions. A successful AI offering needs to have tiered layers of value where the product offerings can be quickly tested in live, low-risk situations. Otherwise, the road to gaining trust in actually using the AI insights will be long and likely impassable. This counterintuitive Goldilocks problem deserves its own post — you can read more on this here.
2. Enterprise Catch 22
Almost every machine learning or data science startup I’ve run or talked to has experienced the famed enterprise ‘catch 22’. Enterprise sales has two phases. Phase one: first, you need to convince your Fortune 500 clients of the value of your brilliant machine learning/data science product. Proof that your predictions are true and that the insights are actually relevant and actionable. This takes longer and more $$$ than expected, even accounting for enterprise sales cycles, because it is hard to overcome human biases, especially that big glaring one called ‘the gut’. Business decisions have been made by gut for all of human history. And your clients, especially leading executives, have not only survived but have thrived because of their past decisions — all made by gut. Your AI software, no matter how much it is vouched by the mathematics or the real world, is not their gut. Empirically, the difficulty of this problem is independent of how good the guts in question actually are at the particular decisions you are helping with. This is the first challenge you have to beat. Given enough of an impressive technology, time and grit, startups will beat this. But then comes the real catch.
Phase 2: assuming the startup survives through the repeated meetings, multiple demos and pilots, you will have your early customers. But just as sales start inching up, your clients will start exploring building an internal team attempting to do the same! Internal teams, the reasoning goes are faster, cheaper, safer and more controllable. Plus, you have just convinced them of the value of data science and repeatedly pointed out that the magic is in the algorithm, not people. From the outside, all data scientists are created equal.
So, how does one crack the enterprise catch 22? The key here is to survive long enough, for all the incredible difficulties you’ve already encountered to educate the once and future customer. Eventually, companies realize that not all data scientists are created equal. Plus, the product features, design, smoothness, continued innovation and growth all clearly stack the deck in favor of external AI products. If this is not the case, it means you are building a feature, not a product. Otherwise, external democratized AI builds are always going to beat out internal solutions. If you see this happening with your customers, you can make your advantage more obvious by add features, such as industry benchmarks, that internal teams can’t possibly provide. Whatever you do, you must survive for the next ~2 years until this illusion fades.
3. Eat your own AI dog food
Startups offering predictive AI offerings alone cannot become massive successes. Let me say that a different way. Predictive software startups will not become standalone multi-billion companies. This is the reason why most predictive AI or insight companies today are stuck in some form of “consult-landia”. Only a company that imbues its own predictions & uses its own predictions to run a core business, such as, infrastructure, insurance or trading is unbounded. On the other hand, when AI insights feed into a hardware or an infrastructure, i.e., a core business, then it is poised to be HUGE. Simply put, you have to put money where your AI mouth is :). If you can and do, the benefits compound and you very quickly stand out. The successful AI companies take the form of service infrastructure or an insurance company (nonideal for a startup). But for an AI startup (of any size), it is extremely difficult to add on the hardware aspect: that would be two different companies! This is part of the reason, we wound down Seldn.
4. Incentives, Incentives, Incentives
As an early stage startup, you need to quickly understand and map out the incentives of your entire stakeholder landscape. Are your users really incentivized to use the product, and the buyers to buy? What exactly is the geometry of the value chain between your customers, buyers and users — a simple triangle or something more complicated? And while we are at it, do you know who your real customers are? Accountability, responsibility, and authority frequently lie with different groups in a big enterprise. The quicker you learn the concept of moral hazard, the faster you will figure out product-market-founder fit. There’s a whole lot packed in that last statement: you’ve heard about product-market fit. But if founder vision/motivations aren’t really in sync with product-market learnings, a startup begins to lose steam. It is indeed incentives that makes the world go around.
5. 99 Problems but $$$ ain’t one
Of all the above, this is my favorite learning. Simply because it is the most counterintuitive. None of the above issues(or most of others) can be solved by throwing massive amounts of $$$ or data scientists at these problems. You are simply stretching out and even amplifying your pain. Perhaps the only thing money can help with is surviving the enterprise catch 22.