Post - RFI

The real bottleneck in getting business logic right was never writing the rules. It's getting the requirements out of someone's head in the first place.

That used to require a business analyst trained in a very specific technique — repeatedly asking "what is X?" until every term bottoms out at something the system actually stores. It's how rules adoption worked long before natural language made direct capture possible. It still works. It's just slow, and it needs a skill that's gotten rarer.

So we shipped a feature that lets AI run that interview itself.

We call it RFI — Requirements From Interview. Instead of needing a pre-written prompt to start a project, you can just say "let's discuss," and the AI runs a structured, BA-style interview: what are the constants, what are the lookups and foreign keys, where are the genuine judgment calls versus the straightforward derivations, what does the type hierarchy actually look like. It synthesizes all of that into a real requirements document — not a transcript, a build-ready spec.

It's live and documented here: [link to https://apilogicserver.github.io/Docs/Exec-Reqmts/#requirements-from-interview]

The validation I like best didn't come from us. A senior IT executive at a major logistics company — someone who's spent decades running large-scale systems, and isn't easily impressed — used it, unprompted, on a real customs-declaration domain: the rules for what has to happen when a shipment crosses into Canada. He didn't write it a prompt. He just started talking to it.

The AI kept surfacing requirements he hadn't stated yet, revised its own design as it learned more, and self-corrected through several rounds of live testing. In under an hour it had a working draft — about a dozen tables, three dozen rules, covering a genuinely complex regulatory domain — ready for review and sign-off, not a rough sketch.

His reaction: "blown away."

The goal underneath this — the goal underneath everything I work on — is letting people express logic in the language they already think in, without losing the rigor of governed, readable rules. RFI pushes that one step further: you don't even need to know how to write the prompt. You just need to know your domain. (I'll be writing more here about how I got to that goal in the first place — decades before AI made it possible.)

[DRAFT — not yet posted. Notes for Val below this line, delete before publishing.]

  • Resequenced to run FIRST per your call, so Post 1 and Post 2 can reference it. Rewrote the opening (no longer assumes prior posts exist) and the closing (now points forward to the origin-story posts instead of backward). Post 1 and Post 2 each got a bracketed link back to this one — fill in once this is live.
  • Built from your project_ai_ba_design notes: RFI = "Requirements From Interview," shipped and publicly documented at the link above, grew out of the historical BA "stepwise definition of terms" technique.
  • The validation story is the CBSA customs example — I kept it fully anonymized (no name, no company, "major logistics company," "senior IT executive") per the standing rule that this person's real identity and employer can't appear in public material. That's the same pattern your "Enterprise AI: The Quiet Failure" Medium piece already uses (the "W. Ries" pseudonym + unnamed "major logistics company"), so this stays consistent with what you've published before.
  • I only used "blown away" as color. He reportedly also said "unbelievable" and made a half-joking "lots of folks are going to be out of a job" comment — left both out since the second one especially could read as a loaded claim out of context on LinkedIn. Add back if you want a punchier close and are comfortable with it.
  • Numbers (about a dozen tables, three dozen rules, under an hour) match your notes (~12 tables, ~36 rules) but I rounded the language rather than stating exact figures — felt more natural for a post; swap in the precise numbers if you'd rather be specific.
  • This reads fine as a standalone third post, or as a closer to the two-part Post 1/Post 2 series if you want to run all three close together — your call on sequencing.
  • No markdown will render on LinkedIn — post as plain text.