Crossing the Chasm

Incomplete thoughts on AI enabled products and build vs. buy

“A software developer looking 10 years to the future over a chasm”

Many people think that at some point in the future - lets say 10 years from now - most software will be fundamentally different to what we know today. Specifically it will have been transformed because of AI. Exactly what will be different is less agreed on, but there’s some recurring ideas: there will be less user interfaces, agents will do the work for you, software will proactively work with you rather than you reactively using it, etc.

These predictions have been made since the current wave of the AI Revolution started in November 2022. And yet in that time there have only been minor advances in the state of the art of AI-empowered software products.

Why is that?

Before I try to answer that, I will clarify my argument a little bit. I’m not saying that no software products have started using AI since 2022. Clearly many have. “Adding AI to your product” just means making an API call with some text and getting some text back. I’d (wildly) guess than more than half of software companies have found a way to do that for some little thing somewhere.¹

I’m not saying there have been no groundbreaking new products since 2022. There’s been some, in some categories. For example in developer tools, Cursor is by far the best example. If you told me in April 2022 that 3 years later, I wouldn’t write code, I’d just type in what I want done and it would write it correctly and use tests to verify it - I would be very very impressed.² So clearly that’s a big breakthrough.

It’s not just Cursor, there’s also Claude Code, Replit Agents, and many other successful and very impressive products that taking new approaches to writing code.

But outside of developer tools, what other categories of software have lots of new products that are doing things that were impossible 3 years ago? I would say even inside the developer tools space, outside of writing code, I struggle to think of products I’ve seen where I wonder “fuck, how do they actually do that”.³

Again, this doesn’t mean that individual products haven’t added features that are made more useful by AI, or are only made possible by AI. But if we are making a 10 year bet about fundamental transformation of the industry, and we’re 3 and a half years in, it doesn’t feel like we have gotten very far down that path.

---

To recap: there is a common perception that pretty soon all software will be exponentially more capable. People want to buy that.

There’s lots of ways to make software products a bit better - like 5%-25% better - using LLMs. People are being sold that.

I think this is what leads to large demand for “in house AI” and large consulting budgets and big projects to “unlock AI capabilities”. And to be clear, in some cases that is warranted. But I think in many it’s not.

In SaaS there is a build vs. buy equation that every sales person loves to flog. And in a sales pitch, the answer is (surprise!) always that you should buy, not build. But that answer has some merits beyond just the cost of doing each thing.

If 20, or 200, or 200000 businesses all use the same software for (say) project management, and you’re deciding if you should buy that software or design and build some system yourself, then one of the things you are also deciding is if you should accept that software’s worldview or create your own. Do you invent your own project management methodology, right down to the nitty gritty details, or do you just adopt the one that is baked into the product.

Of course, not all software is equally opinionated. But you are still buying not just the software, and the hosting, and access to the support team. You are always buying the collective knowledge of all the customers that have come before you - and all those that will come alongside you - and the way in which that knowledge is fed back into the software.

So another variant of the build vs. buy question is: is the collective knowledge of the market more valuable than the in-house knowledge you have?

My opinion is that the collective knowledge is generally more valuable on anything you don’t want to differentiate on. I’m biased, because we sell payroll software, but it seems like a bad idea to me for one to invent a new way of paying their employees that is unique to their business. The downsides are higher than the upside.

By contrast if it is something that you want to do uniquely well then there is more merit to going it alone. At Tanda we’re proud of our product development methodology. We think it’s something we do uniquely well. It’s based on other stuff in the public domain⁴ but is not pegged down to a specific product’s constraints - most of it lives in Google Docs.

I think there will be an equivalent build vs. buy question for AI. And I think the answer will be similar to the one I outlined above.

It’s not about how the costs of building the thing and the maintenance costs of overtime. It seems obvious that it will be cheaper to build software using AI than it has been previously (though it’s not clear by how much).

Instead I think what software companies will figure out is how to bake the collective wisdom of all the market into new kinds of products that can only exist because of large language models. And then you will want to buy those, not because you get to interact with the AI directly, but for the same reason that you buy software today.

---

Does this answer the question of why this hasn’t happened yet?

Not really.

My best answer for that is simply that we haven’t as an industry changed our thinking about what products do enough to work out what the next wave of products will actually do that they couldn’t do before.

Customer jobs don’t change over time, but the tools they can use to solve them do (usually for the better). The internet brought fundamentally new ways to do things using software, but it took a while for people to work out how to connect this to things customers actually wanted to do. The same will happen here. Contrary to apparently almost everyone, I don’t see why it should happen significantly faster this time.⁵

---

I asked Claude to review this article, and part of the feedback was:

Resolution: The ending feels somewhat abrupt. You might consider offering more specific predictions about how you see AI integration evolving in software, especially given your expertise in the SaaS space.

My only surefire prediction is that in 10 years time, articles like this one will have aged poorly. But in the meantime, it was good to get some ideas out of my head.

1. For example, we have a feature where you can build your organisation chart in Tanda by connecting hierarchies between positions (bartender, chef, manager, etc). We added a button that guesses at a correct org chart for you just by feeding the position names into an LLM and asking it what it thinks the org chart is. This is cool and useful but it’s not a product. (We also have much more useful AI-powered features, this is just the simplest to explain.)

2. Even 7 months ago (when I wrote this) I would have been impressed.

3. One category might be meeting summarisation / note taking apps. But Gong started in 2015!

4. Particularly Shape Up.

5. Just because you can write the code faster doesn’t mean you can get to the good ideas faster. Innovation is difficult more than it is expensive.