AI lets you scalably do things that don't scale

You’re probably familiar with YCombinator’s advice to startups looking for product-market fit: “Do things that don’t scale.” The canonical example of this is the Airbnb founders going in person to apartments to take high-quality photos in order to make their listings look more professional. Just last week, we found ourselves reading through hundreds of test questions from a potential customer to help explain to them why RunLLM’s answers were high quality — that’s certainly not something we could do for every one of our customers even as a startup, much less as a much larger company!

Source: DALL-E 3.

That level of customer service and care would of course be ideal if you could do it, but with a human in the loop, there’s simply no way it would be affordable to engage at the level with every customer. We’ve come to realize that the power of LLMs is to bring to every task the level of personalization (or customization or attention to detail) that a startup founder would bring to the table. With inference costs dropping rapidly, this is going to be more obvious than ever. This is a natural complement to last week’s post on changing user behavior — in fact we wish we mentioned it there as well.

Let’s start by getting more specific about where this manifests with a few examples:

• Sales development. AI SDRs have been all the rage, especially with the rumors that 11x.ai raised a $50MM Series B. We’ve been exploring similar tools, and while the technology has a long way to go, the promise of AI-powered SDRs is obvious. A well-built AI SDR pipeline can take in a variety of signals — prospect background, company news, open job postings, etc. — and generate customized email copy within a few second. That’s the type of work that founders will do for early customers and enterprise-focused SDRs do for high-value leads, but this technique can now be applied across every prospect in your TAM. Customized outreach at scale wasn’t possible before, but it is now.

• Technical support. We of course know our market at RunLLM the best. We touched on this point briefly in last week’s post about user behavior, but an AI technical support engineer can customize code, translate between programming languages, and provide detailed explanations that a person would find out of scope and tedious. However, when you can do this reliably (which LLMs now can!), the customer experience becomes dramatically better than it was before because you can get started more quickly and have your issues resolved instantly. In other words, AI technical support allows a level of customization that wasn’t previously possible.

• Software engineering. There are fun parts of software engineering, and there are tedious parts of software engineering. For most people, adjusting border widths on a text input or checking whether your new API’s parameters are consistent with what already exists is not the fun part — but doing this well is what leads to high-quality software. Until now, this has meant that only the teams who are willing to put the time into these details win. While the products aren’t quite there yet, developer assistants have the capability of automating the tedious parts of these tasks and letting engineers focus on what’s more fun — and thereby again enabling excellence at scale.

It’s outside of our area of expertise, but what seems to make Character AI and the like so appealing (and perhaps scary) as well is the level of customization of experience you can achieve — in a way that you can’t customize other humans!

This is only a handful of examples that are top-of-mind for us, but the real list is much, much longer than this. Hopefully this is at least illustrative of what we mean by “things that don’t scale.” Realistically, we’re so early in discovering all the uses of this technology (e.g,, customized healthcare, more detaield interviewing) that we haven’t even begun to scratch the surface of what the implications are.

Something that’s extremely top of mind for us in running a startup, however, is that this allows startups to scale more quickly and while staying more lean than before, and that’s something that’s very exciting. A product leader at one of our customers shared with us that its their team who’s responsible for frontline support, and while they know they’ve reached enough scale to hire in technical support, no one at the company has done it before. Rather than figuring out how to do it — and potentially making a mistake — they’re excited about using RunLLM to handle 80%+ of questions that can be automated, and relying on the existing support rotation to handle the other 20%. This allows them to scale much further with their existing headcount while they’re on a limited budget.

Of course, as we’ve said many times, we don’t think that AI tools are going to fully replace humans in any job function in the near future. You’re not going to reach true scale without a support team (or without human SDRs, without software engineers, etc.) — the difference is just that you can get more done with your existing headcount while you’re small and scrappy, and you can focus on hiring key roles later on. When you do start scaling, the team you hire will be correspondingly more productive.

Being small and scrappy is what doing things that don’t scale is all about! Until the most recent wave of AI tools, you grew out of small and scrappy by finding ways to automate the things you were doing by hand. AI is a form of automation, but it’s one that allows hypercustomization and specificity, which means that you still get much of the benefit of what you did in the non-scalable phase.

We’ll close with a small disclaimer: This probably doesn’t apply to super early stage companies. While you’re still defining your product and understanding your customer, you don’t want to have an AI in the way — you need to be in every conversation in order to develop an intuition that allows you to make smarter decisions. Once you get past that stage, though, everything’s fair game!