The technical co-founder I spent years looking for
Every idea I had needed someone else to build it. So none of them got built. Then that stopped being true, and nobody sent a circular.
For most of my working life the answer to “why haven’t you built that yet” was the same sentence: I’d need someone technical.
It was not an excuse. It was true. I could specify a product in detail — what it should do, what the numbers should say, where it would break — and then the specification would sit in a notes app because the distance between a specification and a working thing required a person I did not have.
So the ideas accumulated. Not bad ideas. Unbuilt ones.
What actually changed
The useful way to describe what changed is not “AI got good at coding.” It’s narrower than that, and more specific.
What changed is that the translation layer stopped being a person.
The scarce skill was never the domain knowledge. I already knew what a clean reconciliation looks like, what a forensic analyst flags in an annual report, what an Indian bank statement does that no Western parser expects. That knowledge was the hard part and I already had it. The missing piece was purely the translation from that knowledge into code — and that translation is now something I can do myself, out loud, in English.
This is a smaller claim than “anyone can build anything now.” It is also a much more useful one. If you have no domain knowledge, AI gives you a fast route to a generic thing that already exists. If you have deep domain knowledge and no code, it removes the one dependency that was actually blocking you.
Why finance people are unusually well-placed
Three habits transfer almost directly:
You already think in systems with rules. Double-entry, accrual, materiality, reconciliation. You are used to a domain where the rules are explicit and the edge cases are where the work is. That is the same shape as writing a spec.
You are trained to distrust output. An auditor’s reflex is to ask what would have to be true for this number to be wrong. Applied to a model’s output, that reflex is worth more than any prompt technique. The failure mode of people new to AI is accepting confident output. That is exactly the failure mode a CA is trained out of.
You know what the answer should roughly be. This is the underrated one. When a tool returns a categorised ledger, I know within seconds whether it is plausible. Someone without the domain would accept the wrong answer and never know.
The part nobody mentions
It broke constantly.
Not in an interesting way. In a boring, four-hours-on-a-Sunday way, where the thing worked on one file and fell over on the next because a bank formats its statement differently. The demo is thirty seconds. The work is the long tail of formats, edge cases and silent failures behind it.
That is not a complaint. It is the actual job now, and it is a job I can do, because it is far closer to reconciliation than to computer science.
If your list is still sitting there
The specification is the asset. If you have been carrying around a list of things you would build if only you had someone technical, the honest update is that the list was never blocked on a person. It was blocked on a translation step that has become something you can do yourself.
Start with the smallest item on that list. Not the ambitious one — the small one you understand completely. The point of the first build is not the tool. It is finding out that the dependency is gone.