Looking ahead in AI
Hot (and cold?) takes on the future of AI
Programming note: We’re taking a (delayed) summer/fall vacation for the next couple of weeks, so we won’t have any new posts. We’ll use the opportunity to re-send some of our favorite posts from the last couple of years. See you in October!
This is our 100th post on the AI Frontier; last week we used the milestone to look back on what we’d learned over the last two years of writing this blog. A 100 posts feels like a big milestone — and many of you haven’t been reading us for that long — so we thought we’d let the takes fly today. Feedback (or angry an rebuttal?) is, as always, welcome. 🙂
Enterprise AI adoption is coming (still)
One of the lessons that we left on the cutting room floor last week was that enterprise adoption was faster than before but still slow. While that’s true — enterprises really still are figuring out what to do with AI — the intent to adopt is certainly there, and we think that intent is going to convert into action soon. How and when that happens depends on where in the stack you are today, and outcomes are likely to vary wildly.
One Fortune 500 company we spoke with recently told us that they were close to wrapping up an 18-month (not a typo) POC with Glean. At the end of this POC, they were planning on building a variety of applications on top of Glean. What were those applications going to do? Basically everything at the company. How well were those applications going to work? Well, at the end of this period, they hadn’t started building many of them!
Doing something is of course better than doing nothing, and if you’re a horizontal player — especially one with the stellar branding of a Glean or a Sierra — you’re likely already seeing enterprise traction (more on platforms below). Vertical applications will likely take a little bit longer to see real enterprise traction, but we’re seeing the green chutes from teams that realize that horizontal solutions won’t solve all their problems. It’s not that verticalization doesn’t add value, but enterprises will pick the low-hanging fruit first. Once the learn that enterprise search isn’t going to solve all their problems, they’ll look for what’s next.
Value will be increasingly decoupled from chat
We’ve written about chat as an interface many times before, so we’ll try not to repeat ourselves too much. Chat was a great starting point for language models, but the fact that so many products and customers are anchored to chat is a problem for the industry writ large. Chat as an interface is simply too limited for many of the kinds of work that people regularly do. Even within the product that’s the posterchild for chat — ChatGPT — we increasingly find ourselves spinning off a piece of work and letting it run in the background while we do other things. (The product itself hasn’t caught up to that type of usage.)
As we get pulled further into use cases like AI SRE, we’re becoming increasingly confident that background work with interactive follow-ups is the right mode of interaction — think Deep Research, Claude Code, etc. — and that products that are all-chat or all-background work are going to be too limited in scope. You’ll build one then the other, in whatever order happens to work best for you.
Note that we said decoupled from chat rather than suggesting chat will be replaced altogether. Chat still plays a valuable role for digging deeper, clarifying, retrieving more data, sharing corrections, etc. Chat isn’t going anywhere, but it doesn’t need to be the headline for every application.
Of course, this is from the perspective of someone who’s thinking a lot about AI. We still occasionally talk to customers who are surprised that LLMs can be configured to do more than chat. We believe it’s going to be a key evolution point for AI applications to convince people that more-than-chat is possible and necessary, and we’re certainly on the warpath. It might happen slower than we’d expect, but it’ll happen.
Everyone’s trying to become a platform
We’ve noticed that many of the well-funded players in the space are trying to platformize. This started with Glean going from enterprise search to an app development platform, and Sierra announced an agents SDK for along with their most recent round of funding. We can understand the motivation: Customers increasingly want configurability, so you might as well take the building blocks you’re working on and package it into something composable. It naturally increases your TAM.
What we’re struggling to reconcile though is how many customers are willing to build genuinely complex workflows with these platforms. It’s one thing to take the call transcript of a sales call, summarize it, and maybe change a deal stage based on the results. (You can do this in Zapier too.) Generating cold outbound email at scale, investigating production errors & alerts, or debugging customers issues is a completely different task.
The operative question for those complex tasks isn’t whether you’re building on Langchain vs. Sierra; it’s whether your team understands enough about the internals of all the requisite integrations and has enough expertise in LLMs to make them work. At the extreme, some enterprises will invest in building the expertise, but it will be slow and painful. We fully believe simple applications (especially ones that look like chat) will be commodified, but these platforms won’t generalize to more complex apps.
Competition → Consolidation
If you’re building an AI application today, your list of competitors is probably at least 10 companies long at this point. As YC churns out more companies than ever, that’s likely to grow in the short term. There’s plenty of smart and motivated prospective founders out there who are looking for interesting problems to solve, and most markets are early enough where there isn’t a clear winner that you should be afraid of (yet).
That level of competition is likely unsustainable however. In the short run, it’s good in the sense that it generates creativity and forces everyone to put their best product forward. But at some point — whether because of macroeconomic changes or because we decide we don’t in fact need the 87th AI SDR product — there won’t be funding available for the Nth startup in a space to get off the ground.
While that might feel like a cynical thought, it’s a natural settling down after any hype cycle. Most of the nascent startups in the space will be bought either by larger incumbents attempting to build out an AI team or by more established startups in the market. It’s always difficult (read: impossible) to predict the market, so whether this is a 2026 trend or a 2030 trend is not clear — but our bet is that it’ll happen sooner rather than later.
Who knows what will happen with foundation models?
It would be a strange blog post about what’s coming in AI without a take on foundation models. We tried to thread the needle on this prediction — something that would account for a trajectory that is slowing but include the necessary humility — and it turned into gibberish. The honest truth is that we really don’t know where models are going to go.
If we had to predict today, we’d say that the rate of changes of existing models has been coming down, and that means more and more innovation is going to happen at the application level. That’s certainly how we felt about GPT-5. At the same time, this whole hype cycle was kicked off by a series of unexpected improvements in the technology that didn’t necessarily reflect the previous rate of progress.
We’re bearish on outlier outcomes like ASI or one model that rules them all. So for the time being, we’re going to keep our heads down and focus on building the best applications we can.
Having written all of this, we’re almost obliged to add: Who knows what the hell is going to happen in general? If the last two years have taught us anything, it’s that our expectations about speed and change have been completely off. We’ve learned enough such that we think we can make somewhat coherent predictions without embarrassing ourselves too much, but there’s always a chance we’re horribly off.
Either way, there will be plenty to learn!