Form-extraction pipelines
Primary submission pipeline
Attachment to filed record in 90 seconds. Per-field confidence scoring with a reasoning trace.
Read the caseAll case studies
Case study · Deadline-driven submissions
Meridian Operations · deadline-driven filingsEmail-to-filed declaration: 6-min median
A queue that used to depend on someone manually checking an inbox twice a day now sweeps itself, classifies every attachment by format, and submits validated declarations before the deadline window opens.
0%
Datasheets auto-classified
0 min
Median email-to-filed declaration
0.00
Datasheet-vs-not detection accuracy
Built with
The problem
A coordinator opened the inbox twice a day. Anything else missed the deadline.
Hundreds of pre-arrival declarations a week, all deadline-bound. The whole queue depended on one person opening a shared inbox at 8am and 2pm, downloading attachments, and deciding by eye what each one was. Off-hours arrivals had no path through. When a filing went late, the incident log named the coordinator — but the failure was structural.
- Manual inbox checks twice a dayThe coordinator opened the shared inbox at 8am and 2pm, downloaded attachments one by one, and decided by eye whether each was a datasheet. Anything that arrived at 2:30pm waited until the next morning.
- Five attachment shapes, one fragile processClean PDFs, combined PDFs, partner-system XLSX exports, screenshots of email bodies, faxed scans. The same extraction template was applied to all of them; mis-routing was the most common error and only the downstream reviewer caught it.
- Deadline misses logged as coordinator errorWhen a filing was late, the incident log named the coordinator who missed the email. The actual failure was structural: the queue had no visibility into off-hours arrivals and no automatic catch-up on volume spikes.
- No submission timestamp in the ticketCoordinators wrote the submission ID into a notes field by hand — or forgot. Auditors reviewing late filings couldn't tell when the declaration was actually submitted vs. when the ticket was last updated.
One day's inbox · checks at 8am · 2pm
5 / 7 missed07:42Pre-8am · clean PDFin window
10:15Off-window · combined PDFmissed
12:03Off-window · faxed scanmissed
14:21Post-2pm · partner XLSXin window
17:48After hours · screenshotmissed
22:11Overnight · clean PDFmissed
03:30Overnight · faxed scanmissed
Inbox check at 14:00Next check: 8am
The pipeline
From inbox to verified record in one pass
Every two hours, around the clock. The sweep doesn't care what time the email arrived.
01Trigger
Inbound ticket arrives
New Zoho Desk ticket created by an inbound email; metadata and attachment references written to the local mirror.
Zoho Desk
02Sweep
Mirror scanned every two hours
A timer-driven Azure Function queries the local mirror for tickets in the queue with attachments and no extraction record yet.
Azure Functions
03Classify
Format identified per attachment
Claude evaluates each attachment against a known-format library and emits a confidence score; below 0.80 routes to manual review.
Claude
04Extract
Ten typed fields with source spans
Faxed scans run through CloudConvert OCR; Claude returns every field as a typed value plus the exact page and span it came from.
Claude
05Verify
Second-pass confidence
Claude reads each field against its source span and emits a 0.00–1.00 score plus a reasoning trace.
Claude
06Submit
Declaration filed; ticket annotated
Validated declaration submitted to the regulator's filing portal; ticket annotated with submission ID, timestamp, and per-field confidence.
Zoho Desk
High confidence
Auto-submitted before the deadline window opens
All ten fields at or above 0.70; declaration submitted without coordinator touch; Zoho ticket updated with submission ID and timestamp inside the same two-hour sweep window. 73% of filings now go in pre-deadline, up from 41%.
Low confidence
Held in the reviewer queue with one-field-resolution
Any field below 0.70 — or any attachment the classifier couldn't place above 0.80 — moves to the queue with the score, the failure reason from the Claude reasoning trace, and the source page highlighted. Coordinators correct one field, not the whole form.
Step 04 — Extraction
What Claude saw → what the system received
Scroll triggers the live extraction. Each region Claude reads on the document maps into the structured form on the right.
Source attachment
datasheet-MO-49118.pdfClaude
Target schema
Declaration v2✓ extracted
Hartwell Trading Co.—
✓ extracted
Guangzhou Precision Parts Ltd.—
✓ extracted
Port of Yantian, China—
✓ extracted
Asia Pacific Freight Solutions—
✓ extracted
CN—
0 / 5 fields populated✓ ready for review
Step 06 — Accuracy review
The system tells you when it's unsure
A second Claude pass scores every field against the source spans. High-confidence results post automatically. Anything ambiguous routes to a human with the reasoning attached.
Field-level confidence
Pass 2 — Claude self-review
HighMediumLow
Buyer of recordHartwell Trading Co.
97%High
ManufacturerGuangzhou Precision Parts Ltd.
64%Low
Routed to human review. Two manufacturer fields appear on the datasheet — one labeled 'declared' and one labeled 'actual' — with different values. Conflicting entries; routing to reviewer to choose.
Stuffing locationPort of Yantian, China
93%High
ConsolidatorAsia Pacific Freight Solutions
91%High
Country of originCN
96%High
4 of 5 fields cleared the 0.85 threshold
model: ClaudeThe stack
Boring tech, glued together well
Each vendor handles what it's best at. Aisyst owns the orchestration layer in between.
Zoho Desk
Inbound email queue and ticket mirror; submission annotations written back via API
Azure Functions
Timer-triggered sweep every two hours; hosts classification, extraction, and submission orchestration
Claude
Classifies and extracts, then scores per-field confidence with reasoning
CloudConvert
OCR preprocessing for faxed scans and format normalization before extraction
Postgres + Drizzle ORM
Local Zoho ticket mirror for sweep queries; extraction records and submission audit log
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed for the queue
The sweep fires every two hours, around the clock. The classifier knows what shape every attachment is. The submission timestamp is no longer hand-typed.
0%
Datasheets auto-classified from inbound
0 min
Median email arrival to filed declaration
0.00
Detection accuracy
0%
Filings in before deadline (vs. 41%)
Pre-deadline submission rate is the metric the queue exists to move
73%, up from 41%. The sweep fires every two hours around the clock, which means any email that arrives outside business hours still gets processed before the next deadline window. If pre-deadline rate drops, first check sweep cadence — a missed trigger turns a two-hour gap into a four-hour gap and filings start bunching at the deadline. If manual-review holds spike, a new attachment format is arriving that the classifier hasn't seen.
Related cases