Hi there,
Do you still check what your AI writes?
What I'm Seeing in AI
Verticals Are the Ones Getting Eaten
There's been a running theory in tech called the Saaspocalypse: the idea that AI will kill software companies, because why pay for software when AI can just build you your own?
Everyone (me included) thought the big general tools would die first, since these have the most money. I now think we had it backwards.
We’ll keep using Salesforce, Outlook, Excel and so on. What actually gets replaced is the niche software built for your industry.
Here's why: your individual workflows can now just be skills, little sets of instructions you teach an AI once, and anyone can make them. A workflow built for YOU beats a tool built for people like you, and it costs almost nothing (AI got cheaper again this week, down 43% in ten weeks).
Watch the legal industry as the test case. Legora, a legal AI company, raised at $5.5 billion in March and is already in talks at $10 billion. Meanwhile Google's newest cheap model just beat the specialist tools at their own legal test.
Niche software has to mould itself to each individual customer now, or a custom workflow kills it.
The Models Are Done. We're Not.
My whole belief at the moment: the AI models of the future have basically arrived. If they knew everything about you and you could talk to them anytime, they'd already work like an all-powerful assistant.
What's left is everything around the model: getting it your context, wiring it into your tools, and putting it in a form people actually use.
Look at the week through that lens:
- Anthropic's IPO pitch is a $30 trillion market: roughly 45 cents of every dollar of wages on Earth, with revenue forecast to hit ~$190 billion by 2028.
- The wrapper now matters more than the model. Researchers ran the same AI models on the same tasks inside different setups (the tools, instructions and memory wrapped around a model). Success went from 52% to 76%, and costs swung 40x, just from the wrapper.
- Instinct, a personal AI companion with no revenue and a 23-year-old founder, went from $500 million to $2.5 billion in valuation in three weeks.
- And the public? 52% of Americans are now more concerned than excited about AI, up from 37% in 2021.
The intelligence is here, the money believes it, and most people haven't even started.
What I'm Building
Drinking from a Firehose
Last week Batko surprised me for my birthday by lining up a coffee with a super interesting AI genius - Geoff Huntley.
I learnt more in two hours than in the last month of reading. The main learning was around back pressure.
Here's the thing about AI: it's like a slot machine that makes work cheap. Pull the lever and work comes out, sometimes brilliant, sometimes confidently wrong, and it will happily keep pulling its own lever all day.
So the value doesn't come from the pulling anymore. It comes from the tests, rules and checks that keep the machine right: every piece of work has to pass its checks, and when it fails, the failure flows straight back and the AI has another go until it passes. That's back pressure.
AI made doing the work cheap, so checking the work is now the whole game.
We’re now building a really robust checking layer across everything we make.
Early proof it works: one of our agents kept filing information in the wrong place. We measured it first (62% right), gave it a test to pass, and let it fix itself. It's now at 100%.
Build in Public!
I posted our revenue journey, every month since I quit my job, as one chart on LinkedIn, and wrote the whole story up as an article on X.
It interested a lot of people I think because seeing real numbers is super intriguing.
Thanks to my mate Archie for the idea, you can now like and comment on these newsletters on my website.
See you next Friday,
Finlay
Explore the rest of my writing here → finlayekins.com/writing
Join our monthly company newsletter here → thehourglass.ai/about