Business fraud · handwritten checks
9,343 checks
~12 checks to a page. Their OCR platform could only read one. We turned the set around in three days. 100% after audit.
Forensic document extraction
From messy documents to a court-ready database.
You were retained to find the story in the numbers, not to type bank statements into a spreadsheet. We turn unstructured financial evidence into one structured, searchable database. Every figure links back to the original page. A US-based team reviews the deliverable, and samples the extraction to AICPA standards.
- US-based expert review
- Every figure linked to its page
- Never used to train a model
STMT_2024-04_4187.PDF · p.12 / 24reconciled
The proof
We win head-to-head.
Recent engagements, measured against the current best alternatives.
01 / 09
100%
Standing guarantee
of figures link back to the original page, highlighted where they were printed.
Why manual extraction fails cases
The ledger is the case. Build it like evidence.
Defensibility
If the extraction is sloppy, people will ask whether the ledger is complete and where each number came from. Our detection suite is deterministic. Benford analysis, duplicate detection, and reconciliation are pure, reproducible functions, not model output.
Traceability
When a number can't be traced back to the original page, checking it gets slow and expensive. Click any figure in the ledger and we take you to that page, with the amount highlighted where it was printed.
Throughput
Manual data entry across tens of thousands of documents drains expert time. Our layout-parsing pipeline reads a statement segment in seconds, and falls back automatically when a document won't parse.
How it works
Five steps from banker's box to database.
One document, five stages.
Tap a step to jump.
Pipeline · artifact 0011 / 5 · Secure ingestion
Secure ingestion
Evidence arrives by upload or from a connected Drive folder. Every file becomes a logged artifact with custody-aware metadata. Nothing enters the case outside the record.
Segmentation & classification
One scanned PDF often bundles many documents. A triage pass splits each upload into typed segments (statements, checks, invoices, wires, ledgers), each with page ranges and a confidence score.
Structured extraction
Specialist extractors read each segment: multi-page statements, dense check-image grids anchored by check number, and combined statements split into their sub-accounts.
Confidence-gated review
A US-based team samples the extraction under AICPA standards, and any field that comes in below confidence still goes to a named reviewer. They see the original page with the figure highlighted where it was printed, and every decision is recorded.
Reconcile, analyze, package
Each statement must balance forward to the penny before analysis begins. Then we run detection across the ledger. Corroboration decides which make the report. Deliverables ship with a .
Forensic analysis
Deterministic detection. Corroborated leads.
Analysis your expert can defend under Daubert. Reproducible statistical methods over a reconciled ledger, and a corroboration rule that keeps one-off anomalies out of the deliverable.
Benford's Law & statistical tests
First-digit conformity with chi-squared significance, duplicate detection, round-number analysis, and temporal patterns like weekend or after-hours activity. All of it is deterministic. Run it twice and you get the same answer.
Corroboration, not speculation
A flag from one method never becomes a forensic lead on its own. We only elevate a finding when two or more independent methods point at the same subject, and the report records which ones agreed and why.
Reconciliation to the penny
Opening balance plus signed activity has to equal closing within $0.01 for every account, including each sub-account on a combined statement. That status is saved and reportable across the whole set.
Meet Clive
An investigator's copilot that shows its work.
Once the evidence is structured, Clive sits on top of the case. Ask a question and you get tables and KPIs, with page citations. Clive will not guess at a number.
- Ask a question across the whole case ("who received the most round-number payments?") and get an answer from deterministic tools, not a model guessing at the numbers.
- Document Q&A cites its sources. Every answer links back to the original page, highlighted where the figure was printed.
- Strictly case-scoped. Clive never mixes data between engagements, and your evidence is never used to train a model.
Clive · case 2024-CV-8841case-scoped
Which payees received round-number payments across more than one account?
3 payees corroborated by 2+ methods
| PAYEE | METHODS | ACCT | SOURCES |
|---|---|---|---|
| MARMOT RIDGE | round+dup | 3 | stmt p.12 |
| CASCADE SUPPLY | round+temp | 2 | checks p.4 |
| ALDER HOLDINGS | dup+benford | 2 | wire p.2 |
deterministic toolspage-citedno free-text numbers
Our standing pledge
Your evidence is never used to train a model.
Not ours, not a vendor's, not “anonymized and aggregated.” Case data is scoped to a single engagement, used only to produce that engagement's deliverables, and never fed back into training, fine-tuning, or evaluation. See how we handle evidence →
- Used only for your deliverables
- Scoped to one case, by construction
- No training. No fine-tuning. No eval sets.
- No cross-case reuse, anonymized or not
On the record
What people on the case said.
“They saved me weeks on extraction and analysis. I spend that time on the financial story now.”
“Opposing counsel asked where a number came from. I showed them the original page, the figure highlighted, and who reviewed it.”
Deliverables
Built for analysis. Packaged for testimony.
Four artifacts leave every engagement, plus the structured database behind them.
XLSX
Master ledger workbook
A branded, filterable workbook of every extracted transaction: statements ledger, checks register, invoices register. Each row opens the original page, with that figure highlighted where it was printed.
PDF
Forensic leads report
Corroborated forensic leads only, with the plain-language basis for each lead and the detection methods that agreed.
DOCX
Methodology & Integrity Report
A court-facing document recording the extraction pipeline, models and validators used, review thresholds, reconciliation results, and the full human review record.
LOG
Tamper-evident execution log
A hash-verified log of every processing step, supporting a reproducible, audit-ready account of how the evidence set was handled.
FAQ
Common questions.
Is this a self-serve SaaS product?
No. This is a service, not a self-serve product. We run the extraction engine, the review, and the analysis, and we stay on the case with you. A US-based team reviews the deliverable, and samples the extraction using stratified sampling under AICPA standards.
Does a person review every extracted field?
No. A US-based team reviews the deliverable, and samples the extraction using stratified sampling under AICPA standards. Fields that come in below confidence still go to a named reviewer. Every sample decision is recorded.
How do you preserve evidentiary integrity?
Every value links back to the original page, highlighted where it was printed. Human review decisions are saved with who made them. The processing run is captured in a hash-verified execution log. The analysis methods are deterministic, and we document them in the methodology report.
What kinds of documents can be processed?
Bank, credit-card, and combined statements (including multiple sub-accounts under one account number), check images, invoices, receipts, wire transfers, deposit slips, ledgers, and correspondence. Bundled scans, where many documents live in a single PDF, are fine too.
What does the final delivery look like?
A master ledger workbook, a forensic leads report gated by multi-method corroboration, and a Methodology & Integrity Report suitable for disclosure, plus the structured database behind them.
Send us a live case.
We'll process your pilot at a steep discount, and you'll get back a structured, source-linked database instead of a stack of PDFs.
Request intake review
Typical response within one business day.
james@accordai.net