The next threat isn't the chatbot
A lot has been said about the SaaSpocalypse. What's already clear is that LLMs can outperform a lot of SaaS applications at one-off tasks, given enough context.
What I think is the next threat, and the one that needs a real strategic response from SaaS companies, is what I'd call synthetic, on-demand UIs. Jonathan von Rüden, Chief AI Officer at SAP, put it this way: we're heading into an era where user interfaces are created dynamically by AI.
What Claude can't do (yet)
Claude and others are already eating into generic, multi-purpose software, like building slides and spreadsheets.
What you can't do with Claude et al yet, at least not conveniently and without clunky setup, is run multi-step, multi-hand-off business workflows that rely on deep business logic grown over a long period of time.
But fast-forward a little. What if the models get fast and capable enough to mimic pretty much any UI on demand? They infer from the underlying data and the people using it how to make each step convenient for one user, or several. At that point your screens stop being your product.
So what's left as a moat? Here's my take on a drawbridge blueprint for B2B SaaS.
1. Own the deterministic core
Your SaaS should own core, deterministic operational business logic.
Why? Companies are terrified of their key workflows breaking. It can mean huge losses and real reputational damage. They'd rather pay a bit more for a license, knowing things will keep humming, than go on an adventure, hand that logic to an all-purpose AI and hope for the best.
So AI can sit on top, to help understand and change the system and to run transactional requests. But the system of record underneath needs to do exactly the same thing every single time.
2. Stay away from the mainstream
Anything with very broad application is at risk of being swallowed by the LLMs, because that's where monetization is easiest for them.
So I recommend deep specialization, ideally with context that's out of reach for LLMs: in the physical world, or in technical nooks that don't waft across the internet.
3. Build UIs that are genuinely hard to recreate
Not filters and dropdowns. LLMs already knock those out while you grab coffee.
I'm talking about complex visualizations, and multi-step hand-offs between several humans and agents. The kind of interface where the structure is the value, because it's how a team agrees on what's happening and what happens next.
4. Handle more deterministically
The more your app can handle deterministically, the better, because it saves tokens.
LLMs have to solve everything with tokens, unless they build massive deterministic architectures of their own, which would defeat their purpose. Sure, tokens could become dirt cheap one day. But right now the trend points the other way, because the model makers need to recoup their gigantic infrastructure investments.
Holding ourselves to it
We try to live by this blueprint at Splotch. A deterministic map of what a NetSuite account actually contains. Deep specialization in the quirks of one system, not a mile-wide tool for every ERP. A visual layer for the hard parts, like the blast radius of a change and who signed off on it. And agents that read from the map instead of burning tokens rediscovering it.
Winter is coming. The castles that make it through won't be the ones with the prettiest screens. They'll be the ones with something underneath that a model can't conjure on demand.
From Kevin Lorenz
← Back to all posts