Form-extraction pipelines
Deadline-driven declarations
Datasheet detection in the queue and ten-field extraction in under a minute, before the deadline.
Read the caseAll case studies
Case study · Form extraction
Meridian Operations · high-volume document submissionField re-keying: 22× faster, 94% match
A team that spent its days copying numbers from PDFs into a system of record now reviews exceptions instead — and the extraction gets more accurate every week without a single prompt change.
0%
Extraction match vs. reviewer truth
0x
Faster than manual re-keying
0.00
Reviewer agreement on auto-posted records
Built with
The problem
Every packet meant a person reading a PDF and retyping it three times
Coordinators triaged thousands of submission packets a month. The shape was always the same: an inbound document attached to a ticket, fields copy-pasted into one system, then retyped into a third. Most of the cognitive load was visual pattern-matching plus catalog lookups by memory — exactly the part you shouldn't be paying humans to do, and exactly where the same client kept hitting the same error week after week.
- Re-keying the same number three timesEvery inbound PDF was opened in a viewer, copy-pasted into the ticket, then retyped a third time into the submission form. The third one was always the one that was wrong.
- Classification codes entered by memoryCoordinators looked up codes in a separate tab and typed them by hand. A transposed digit only surfaced days later in review, well after the packet had moved downstream.
- No audit trail to the sourceWhen a reviewer questioned a declared value, there was no way to point at which line of which page it came from. The manual PDF search took as long as the original entry.
- Per-account corrections that didn't stickA reviewer's correction applied to that one packet. The next packet from the same account had the same error. No feedback loop from correction back to extraction.
Repeat-error log · last 6 weeks
recurringHartwell Trading Co.Code 8471.40 vs. 8471.304w running
Westridge ImportsTotal $/qty mismatch3w running
Cascade LogisticsOrigin missing6w running
Northbay GroupDate format flipped2w running
Olmsted & SonsCode 9018 vs. 90195w running
Same accounts. Same errors.No feedback loop
The pipeline
From inbox to verified record in one pass
Six orchestrated steps. Two Claude calls. Reviewer corrections become enrichment suggestions for the next packet from the same account.
01Trigger
Packet arrives
Zoho Desk webhook fires on a new ticket; attachment URLs and metadata land in the processing queue.
Zoho Desk
02Detect
File type routed
PDF, DOCX, XLSX, and email-body attachments each route through CloudConvert into a searchable PDF.
CloudConvert
03Extract
Typed schema with source spans
Claude returns every field as a typed value plus the exact page and span it came from.
Claude
04Enrich
Parties matched, codes validated
Submitting accounts matched against the parties reference table; classification codes validated against the code list.
05Verify
Second-pass confidence
Claude emits a 0.00–1.00 score per field with a one-sentence reasoning trace.
Claude
06Learn
Corrections re-applied
Reviewer corrections are stored as enrichment suggestions and auto-applied to the next packet from the same account.
High confidence
Auto-post to the system of record
All fields at or above 0.70 with reviewer agreement of 0.97; the record is written to the submission form and the Zoho ticket is annotated with the extraction summary and source page references. No human touches it.
Low confidence
Routed to the reviewer queue with reasoning trace
Any field below 0.70 — or any packet flagged with a structural anomaly — lands in the queue with the score, the source span highlighted in the PDF, and the reasoning visible inline. Reviewers correct the specific 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
packet-MO-48217.pdfClaude
Target schema
Submission record✓ extracted
Hartwell Trading Co.—
✓ extracted
2026-03-14—
✓ extracted
8471.30.0100—
✓ extracted
$142,380.00—
✓ extracted
M. Okafor—
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
Submitting accountHartwell Trading Co.
98%High
Effective date2026-03-14
96%High
Classification code8471.30.0100
91%High
Total declared value$142,380.00
61%Low
Routed to human review. Line-item subtotals sum to $138,240.00, not the stated $142,380.00; one row is partially obscured by a red stamp on page 4. Routed to reviewer for manual check.
Party signatureM. Okafor
88%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 ticket queue + webhook source; extraction summary written back as ticket annotation
CloudConvert
File-type routing and OCR preprocessing for scans, DOCX, XLSX, and email-body attachments
Claude
Extracts typed fields with source spans, then scores confidence with reasoning
Azure
Functions hosting the extraction pipeline, suggestion re-application, and reviewer routing
Postgres + Drizzle ORM
Parties reference table, enrichment suggestions, per-packet extraction records
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed for the team
The pipeline runs every minute against new tickets. Coordinators graduated from re-keying every field to validating the rare flagged ones — and the system gets sharper week over week as their corrections feed back in.
0%
Extraction match vs. ground truth
0x
Faster than manual re-keying
0.00
Reviewer agreement on auto-posted records
0%
Week-12 error reduction on enriched fields
Watch the week-12 error reduction, not the day-one accuracy
Day-one accuracy says how well the prompt is written. Week-12 reduction says how well the loop is working. When a reviewer corrects a product field, that correction is stored and re-applied automatically to the next packet from the same account — so the per-account error rate falls each week without any prompt change. Flatten before week 8 and the suggestions are saturating a narrow catalog; climb after week 12 and the catalog is drifting. Either signal calls the curator in.
Related cases