FOUNDER NOTE
What I actually use AI for as a technical founder
A practical list, including the parts where it does not help and the places I have decided not to let it near.
Most writing on this topic is either "AI does everything now" or "it is overhated". Here is the specific version, from someone running client work, products and a small team.
Where it genuinely helps
Exploring an unfamiliar area. When I need to understand a protocol, a library or a domain I have not worked in, having something to interrogate is faster than reading documentation linearly. I verify anything I am going to rely on, but the orientation phase is meaningfully quicker.
First drafts of code I know how to review. This is the important qualifier. It is useful where I can immediately tell whether the output is right. Where I cannot, it is a liability wearing a helpful expression.
Debugging as rubber duck. Explaining a problem in enough detail to ask about it solves it a reasonable fraction of the time, which is an old effect with a new interlocutor.
Boilerplate. Migrations, types from a schema, test scaffolding, repetitive transformations. Genuinely tedious work with an obviously checkable result.
Documentation drafts. Getting from nothing to a rough structure is the part I procrastinate over. Editing something mediocre is much easier than starting.
Where it does not help me
Architecture decisions. It will produce a plausible architecture. Plausible is not the bar — the bar is knowing which constraint matters most for this client, and that comes from having sat in the meeting.
Anything requiring current facts. Pricing, regulation, current model capabilities, what a vendor's API does today. This is exactly where confident wrongness is most likely and most costly, so I check primary sources.
Understanding what a client actually needs. The stated requirement is usually a solution somebody already chose. Getting past that is a conversation, and it involves noticing what someone is uncomfortable saying.
Where I deliberately keep it out
Client-confidential material, unless it is running locally on hardware I control. This is most of why I run agents on a Mac Studio rather than sending everything to a hosted model.
Final legal or compliance language. Related to the whole "no source, no claim" position — I would not let a model be the last reader of something with legal consequence.
Anything published under my name without me reading it properly. Including this. The failure mode is not obvious errors, it is smooth text that says slightly less than I meant, and the only defence is reading it.
The honest summary
It has made me faster at the parts of my work that were already mechanical, and roughly as fast at the parts that were hard. The hard parts were always judgement — what to build, which constraint dominates, whether the client's stated problem is the real one — and none of that has moved.
Which is a less exciting claim than the ones in circulation, and closer to what I actually observe.