Marginalia: 2025-07-28

Reading List

Rough Thoughts in Progress

AI Native vs. AI-Grafted-On

What does it look like to reimagine how a company or industry should work with AI at the core of the proposition?

Some examples of what I would consider to be truly AI-native companies: Jack and Jill & OffDeal.

Everyone wants "AI" in their business and in their product but some are able to reinvent, while others feel like they're grafting on the technology to little or no effect.

Something I saw this week that I don't feel got much attention was from Mike Knoop - Zapier has a "sexy" product (AI agent) and a "boring" product (structured workflows). The boring product is crushing it, the sexy product is struggling. Most businesses need predictable results in many to all aspects of their product and processes. A single slip up in a critical process ruins confidence in in an agentic system. So you can put a huge amount of effort in trying to achieve determinism with AI (grounding, RAG, guardrails etc) or you can accept you need to set down your hammer and pick up the right tool for the job.

It's worth thinking about the difference between between Zaps and the AI-native products above. My 2c: don't replace what should be deterministic with something that isn't. Chatting with a recruiter was never going to be deterministic - and nor should it be. The magic is in personalisation, in taking the conversation where it needs to go. Following the "process" is secondary.

So the domain of AI-native businesses are going to be the opportunities that benefit from non-determinism. Perhaps a good heuristic would be to ask whether your competitors are "traditional" software companies or service businesses. In the latter non-determinism is a feature not a bug.

I don't feel this is a particularly novel insight. There are certainly founders attacking service-orientated markets with an agentic thesis. I do however believe many people misunderstand services businesses - you aren't buying the knowledge, or indeed output, of a consulting firm. You're buying the ass covering. Agents don't have a throat to throttle, so you probably should offer up yours. You're also not "capturing" service revenue - people pay what they pay for software, not what they pay for people...

The Offdeal idea is interesting because "OffDeal bankers are supposed to shoulder seven to 10 deals at a time" - this speaks to the idea of melding experts (the throats) with AI (the grunts) to amortize cost. Banking is an interesting industry in this regard because you can take a fee percentage of the work delivered (5% flat success rate with no retainer in Offdeal's case). See also Rocketable Manifesto - buying companies and replacing the org chart with AI agents. It appears to me, the issue with this approach, is actually setting up the coordination to guide a system of agents coherently towards a goal.

As a parting thought, in the beginning of any technical discontinuity, it's natural to see some ugly-duckling products, as we adapt to the new paradigms. These are the Horseless Carriages of the modern day. What businesses will need to figure out is:

  1. Should they react so radically to AI? - if AI is the new internet it's hard to imagine any business being untouched. If it's closer to cloud or mobile, perhaps one can over-rotate. It wasn't existential for Salesforce to have a mobile app (but it was for Facebook).
  2. Can they adapt? - if you're a farrier you could retrain to make alloy wheels. My money would be against your success but godspeed anyway.