We kept shipping wrong data mappings to clients. So Team built an AI that's allowed to say "I don't know.
A few years ago I was doing implementation work — the kind where you inherit a client's messy export from HubSpot, Salesforce, some legacy ERP, whatever — and you have to map every field to a target schema before anything can go live.
Everyone who's done this work knows the drill: you open two spreadsheets side by side, squint at column names that don't match, guess at what "ACCT_NM_2" probably means, and hope you got it right. Get it wrong and it's not a bug — it's someone's account name silently mapped to the wrong field in prod, discovered three weeks later by an angry customer.
When AI mapping tools started showing up, I was hopeful. Then I tried a few. They had the exact same failure mode as a rushed human — just faster. They'd confidently map "Region" to "Country" because the words felt adjacent, with zero indication that it was a guess. A wrong mapping delivered with total confidence is more dangerous than an honest gap, because nobody double-checks the confident stuff.
That's the whole reason Crossary exists.
The idea was simple: what if the AI was allowed to abstain?
Instead of forcing a mapping for every target field, Crossary proposes one only when it has evidence — and shows you that evidence. Every row comes with the verbatim quote from the source spec it based the decision on, plus its own confidence rating (high, medium, low, none). If nothing in the source actually supports a mapping, it says so instead of guessing from a similar-sounding name.
That "no clear source" case turned out to be the whole product, honestly. The moment I saw it flag a gap instead of confidently botching it, I knew this was the trust mechanic worth building the entire tool around.
What it actually does now:
A strict 5-stage pipeline — Artifacts → Fields → Mapping → Validation → Export. You drop in a source and target spec (Excel, JSON, XML, PDF, CSV, XSD, SQL, YAML), it extracts the full field inventory from both sides, proposes mappings with evidence and reasoning, runs a deterministic (non-AI, free) validation pass for structural issues, and exports a signed .xlsx.
The export was the other thing I obsessed over. It's not a locked-in app view — it's a real spreadsheet. Someone can open it in Excel, edit it, argue with a reviewer over it, add notes, and re-import it. Crossary applies what matched, skips anything that changed underneath, and turns every note into a tracked question instead of silently overwriting it. Nothing gets lost in that back-and-forth, which is where these projects usually die in practice.
There's also a "Mapping Memory" — every approved mapping becomes a private, workspace-scoped library, so the next similar integration gets pre-filled suggestions on the fields it previously abstained on. Still reviewed, never auto-applied.
Where it's at now:
Just shipped the landing page: https://www.crossary.com/
Free tier is 3 integration credits, no card. Review, validate, and export never cost a credit — they don't touch the AI, so they're free forever, even on the free plan. You only spend credits on the actual mapping run.
What I'm still figuring out:
Whether "field mapping that shows its work" resonates outside people who've lived the pain of this work firsthand
Pricing — credits map to roughly one integration, but I'm watching whether that mental model is intuitive or needs simplifying
How much of the audience is solo integration/EDI consultants vs. teams
If you've ever hand-built a mapping sheet that someone else had to implement off of — I'd genuinely love to hear how you did it, and where it broke.
The part that resonated with me wasn't the mapping itself, it was treating uncertainty as information instead of something the product has to hide.
Sometimes a system becomes more trustworthy not because it knows more, but because it makes the limits of what it knows visible.