Upsides and Downsides
Every startup founder knows about Geoffrey Moore's concept of "crossing the chasm"–that you have to change your marketing and sales approach as you gain marketshare fit a more conservative buyer. But most fail to internalize what crossing the chasm means when it comes to their product.
I recently stumbled upon Adam Mastroanni's post on strong-link problems, and realized that it's the perfect framework for thinking about this shift.
In essence, Adam says there are two types of problems: strong-link and weak-link problems. Strong-link problems are solved by looking for excellence in a single dimension. When building a startup or doing drug discovery, it doesn't matter how many times you are wrong, you just have to be right once. Weak-link problems are solved by eliminating failure in all dimensions–it's why the FDA puts standards on the internal temperature of your meat or why you might study the p95 latency of an endpoint.
When you have a strong-link problem, you increase variance because you benefit from outliers. You focus on upsides. When you have a weak-link problem, you decrease variance because outliers will destroy you. You focus on downsides.
Early stage startups tend to benefit primarily from solving problems of upside. Early adopters choose a startup because it provides some quantum of utility that no other product does. It doesn't really matter how much downside that startup might have, customers pick it because there is effectively no replacement.
But as a startup matures, new problems come into focus: uptime requirements, security and access controls, audit logging, cost + performance, etc etc. The downside problems.
Many startup teams fail to realize that as they gain marketshare, the problems of their customers shift. The early adopters who once valued variance are replaced by late adopters who care about minimizing risk.
Making this change can be really difficult–revenue stalls at 5-10m, product velocity falls off a cliff, churn is up. All because the new breed of customers care more about the downsides than the upsides.
This model also helps explain why founders from big companies often fail to get off the ground. They don't understand that the early phase of a startup is all about exploration. The downsides don't matter until they begin to find product market fit. The only things to do are actively ship and cut scope.
Balancing upsides and downsides
Where this gets really tricky when a company has matured but still needs to launch another product. Founders need to simultaneously balance the "production-grade" infrastructure needs of their existing customers while incubating the high-variance labs to launch the next big product.
Doing this well takes skill and practice. You can't just ignore your customers when they say "the product doesn't work". But you should still be taking new bets, even as a mature company.
At Segment, we'd think about balancing these needs via the McKinsey horizon's framework:
- 60% core (existing product)
- 30% emerging bets (products which would generate revenue in the next year)
- 10% new bets (more speculative bets)
We'd typically start new bets with a very small percentage of customers who were likely to pay a premium (e.g. only enterprise). They weren't asked to be built with rock-solid infrastructure from day one. It was only once they had some measure of product-market fit that they would start the transition.
There were times where we'd focus a lot more on downsides as well. After a particularly bad few months of uptime, we did a "reliability reset" to double down on improving our core data pipelines. We had some big security up-levels after a handful of scary incidents. But all of those came years down the line.
LLMs and the "upside" phase
The second reason this idea top-of-mind today is because I've been trying a lot of AI-powered products out there.
When it comes to the models themselves, my belief is that most models are still optimizing for the "upside" phase. They are impressive, but don't handle cases of nuanced judgement, adversarial inputs, or uncertainty well enough for a mature business to trust them entirely. Many models will crush various advanced reasoning evals, but you probably wouldn't trust them with a credit card.
The same goes for AI products. In a strange twist, most AI demos nail the "upside" phase, but that's where they get stuck: just a demo, not a product.
There are a handful of products which really manage to make the leap (Copilot, Cursor, Midjourney, etc). My belief is that these tools can cross the chasm primarily because they aren't flakey.
To get a toe-hold, the product needs to be amazing. But to get to mass adoption, it needs to just work.
As your product matures, it's worth asking every few months: does my customer care more about the upside or the downside?
If it's the latter, it might be time to shift focus.