Evaluated · recommended against

AI Credit-Card Advisor Passed

An assistant that tells you which card to use for each purchase, when to redeem, and when to enroll in offers. I wrote the proposal, I ran the numbers, and I passed on it.

90%of revenue per user comes from affiliate · $20 of $22
$17ka year for the trust-pure model, with no affiliate and no bank credentials
1testable dependency the whole proposition reduces to

What it is

The target user is a US points maximizer carrying three or more rewards cards. The job isn't information, because that already exists everywhere. The job is effortless and trustworthy optimization at the moment you're paying.

This one never got built. What it got was a proposal, a business case, and a recommendation not to build it.

The decision

I made the tension arithmetic instead of adjectival.

The pitch writes itself: an advisor you can trust precisely because it doesn't take affiliate money. The trouble is that affiliate money is where the revenue is, so when you strip it out, the only real differentiator earns $17k a year, which is a hobby rather than a business.

And the six-figure version reintroduces exactly the conflict of interest the pitch is built on rejecting, which is why I ended up writing that there's no version of the numbers where strong trust positioning and a real revenue line coexist comfortably.

Once that was written as arithmetic rather than as a concern, the decision stopped being a matter of taste. Everything else reduces to one testable dependency, which is whether you can get on-device purchase context without bank credentials, and if that fails the product is just a worse CardPointers.

What it cost

The recommendation was to pass, with at most a two-week time-boxed explore of a small-business niche. Passing means giving up a real market and a genuinely interesting problem, and it means this particular piece of work produced a memo instead of an app.

I've kept it on this site deliberately, because the projects that didn't get built are the ones where you can check the reasoning most easily: there's no shipped product sitting in front of it.

Status & next

Closed. I'd revisit it only if that on-device context dependency gets cheap, which would change the arithmetic rather than the argument.