Crossary – AI-assisted field mapping that outputs signed Excel files
Crossary is an AI-powered field mapping tool for integration engineers, consultants, and data professionals. It uses a five-stage pipeline to extract fields from source and target specs, propose mappings with evidence, and export signed Excel workbooks. It emphasizes honesty, determinism, and data privacy.
Who it's for
If you hand-build the mapping sheet someone else implements — this is for you.
01
Integration & interface engineers
Mapping heterogeneous specs — PDF, XML, EDI-style guides, Excel — to a target schema.
02
Implementation & onboarding consultants
Turning a client's messy spec into a reviewed workbook on day one — not week three.
03
Data & migration engineers
Building the field inventory and design before any records move.
04
EDI & interface analysts
Turning implementation-guide PDFs into evidence-backed rows where the “why” has to be defensible later.
How it works
A strict five-stage pipeline. Then a round-trip.
Upload to signed workbook in five steps — and back again, without losing a note.
01
Artifacts
Drop in your source + target specs. If a file can't be fully read, it tells you exactly how much it dropped.
xlsx · pdf · csv · json · xml · xsd · sql · yaml
02
Fields
A field inventory is extracted from both sides — every path, every column — so you map against the real surface.
03
Mapping
For each target: a proposed source, type, verbatim evidence, reasoning, and confidence. Abstains when unsure.
04
Validation
A deterministic check for structural & cardinality blockers. Zero AI, zero spend.
05
Export
A signed .xlsx your team can open anywhere — that reconciles cleanly when it comes back.
Anatomy of a reviewed row
Six things every row tells you. Nothing taken on faith.
Open any proposed mapping and you see exactly why the AI suggested it — and how sure it says it is. The evidence, not the confidence, is the product.
target field ①
order_date *
confidence ④
high
source field ②
S_SHIP_DATE
mapping type ③
transformation
to_date(S_SHIP_DATE)
evidence — verbatim from the source ⑤
— ship date, ISO-8601 string. Target is a date type.
See it actually happen
The honest gap it won't guess — and the round-trip that holds.
The real sample, end to end: where the AI withdraws a tempting guess and asks instead, and where your edits survive a trip through Excel and back.
the sample
hubspot → salesforce
target fieldlow confidenceno confidence
Region
←Countryno clear source
No HubSpot field tracks sales territory, and it must not be guessed from Country.
?Region (sales territory) isn't in the HubSpot export. Assign it in Salesforce, or point me at a source?
When it isn't sure, it says so.
confidence: none · no guess
Mapping memory
Every mapping you approve makes the next one faster.
Sign off an export and your approved field-pairs become a private, workspace-scoped library. On the next run it gap-fills only the targets the AI abstained on — inserted as suggestions you still review, never auto-applied.
Scoped to your workspace — it never crosses to another.
Captures the decision only — no evidence, values, or client data.
It never trains a shared model.
next run · a gap, pre-filled
Industry
library
Trust & data
Trust isn't asked for here — it's enforced in code, and written down.
Guaranteed in code
Honest about what it read
If more than ~10% of a source is dropped during ingestion, it stays flagged — in the app and on the export cover sheet. A partial read is never reported as complete.
“Validated” means one thing
Validation is deterministic and checks structure and cardinality only. It does not claim semantic correctness — and the UI says exactly that.
Never loses your note
On re-import it applies what matched, skips what changed underneath you, and turns every reviewer note into a tracked question. Nothing is silently overwritten.
Schema-validated, or rejected
Every AI response is checked against a strict schema before it can touch your data. A malformed result is thrown away, never persisted.
Your data
Specs in, not your records
Crossary maps from your source and target specs — schemas, dictionaries, guides. It's built to work from the spec, not your data, so in most cases there's no need to upload production records or PII.
Pricing
Start free. Pay for AI runs, never for reviewing.
Reviewing, validating, exporting, and round-trip re-import don't call the AI — so they're always free, on every plan.
Free
$0
For your first real integration.
Start free
3 integration credits
Free review, validate & export
Round-trip re-import
Mapping Memory included
1 workspace · up to 2 members
Most teams start here
Pro
$99/ month
~20 integrations a month.
Start with Pro
Everything in Free
~20 integration credits / mo
Higher-accuracy pass
Unused credits roll over
3 workspaces · up to 3 members
Team
$399/ month
~75 integrations a month.
Start with Team
Everything in Pro
~75 integration credits / mo
Unlimited workspaces · up to 10 members
Shared mapping library
Plus applicable taxes.
Run out? Existing work is never blocked — review, validate, export & re-import stay free. Unused credits roll over, you can buy more anytime, or upgrade. Only new AI runs pause.
Questions, answered straight
The honest answers, before you sign up.
What's a credit?
Roughly one standard integration. Higher Accuracy and unusually large or heavily re-run jobs use more — and you can see what each integration cost.
What's always free?
Reviewing, validating, exporting, and round-trip re-import. They don't call the AI, so they never use a credit.
Is accuracy gated to higher plans?
Every plan gets the same core mapping quality. Higher Accuracy — an optional, heavier pass — is on paid plans; Mapping Memory is included on every plan.
What happens when I run out?
Existing work is never blocked — review, validate, export and re-import stay free. Unused credits roll over, you can buy more anytime, or upgrade. Only new AI runs pause.
How do I upgrade?
Click subscribe on the plan you want — secure checkout, and your plan activates immediately.
Does my data train your model?
No. Mapping Memory is scoped to your workspace and never crosses it; no shared model learns from your mappings.
Map your first integration in the next ten minutes.
Upload a source and a target spec. Get back a reviewed, evidence-backed workbook — with every honest gap flagged for you.
Start free
3 integration credits free · no card · no lock-in.