Legal Services
Meridian & Low
Client documents that file themselves — with a human still holding the pen
Meridian & Low, a 12-lawyer firm, was losing two paralegal-days a week to re-typing client documents from email attachments into their practice management system. We built an n8n pipeline that classifies incoming documents, extracts the key fields with frontier LLMs, and routes anything uncertain to a human review queue — nothing enters the system unreviewed.
Every client document arrived as an email attachment and left as typing
Meridian & Low's intake worked the way most 12-lawyer firms' intake works: clients emailed PDFs, scans, and photos of documents — engagement letters, IDs, financial statements, court filings — and paralegals re-typed the relevant details into the practice management system. Around 14 hours of skilled paralegal time went into transcription every week, and the errors that slipped through were the expensive kind: a mistyped date on a limitation period is not a typo, it's a liability.
The partners had already looked at "AI document processing" tools and bounced off, for a good reason: none of them would say what happened when the AI was wrong. In a law firm, that is the only question that matters.
We shadowed the intake desk for a week before writing anything
Before building, we sat with the intake paralegals for a week and logged what actually arrived: which document types, in what volumes, how messy the scans were, and which fields the practice system genuinely needed. The distribution was lopsided — five document types made up over 80% of volume — which told us where accuracy effort would pay and where a simple "route to human" was the right answer.
That week also set the design principle for everything after it: the paralegals weren't the problem to automate away, they were the quality bar to automate up to. The system's job was to do the typing; theirs was to keep the judgment.
An n8n pipeline from inbox to practice system, with checkpoints
An n8n workflow watches the intake mailbox, splits attachments, and runs OCR on scans. A classification step identifies the document type, then an extraction step — using the latest OpenAI and Claude models, selected per document type by which performed better in testing — pulls the fields that document type requires into a strict schema. A Python validation layer cross-checks the output against the practice management system: does this client exist, is this matter number real, is this date plausible for this document type.
Only after validation does anything get written to the practice system, via its API, with the source document attached to the record. Everything runs in Docker on the firm's own infrastructure, with PostgreSQL holding the queue and the audit trail — client documents never leave an environment the firm controls.
Every extraction shows its work, and doubt routes to a person
Every extracted field carries a confidence score and a pointer to the exact source passage it came from. High-confidence extractions queue for a one-glance approval; anything below threshold routes to a paralegal review queue that shows the field, the model's answer, and the highlighted source side by side — confirming or correcting takes seconds, not re-reading the document. Nothing enters the practice system unreviewed, full stop.
Thresholds are per-field, not global: a client's postcode can tolerate more model confidence than a limitation date, so dates and monetary amounts run stricter cutoffs. Every decision — model output, confidence, who approved it, what they changed — lands in an append-only audit log in PostgreSQL. During the four-week tuning period, each paralegal correction became a test case, which is how field-level accuracy climbed to 96% on the document types that matter.
Fourteen hours a week back, and the errors now get caught on the way in
Transcription work dropped from roughly 14 hours of paralegal time a week to under two hours of review-queue time — the queue itself surfaces only what needs human eyes, and the one-glance approvals move at reading speed. That returned time went where the partners wanted it: client contact and matter preparation, the work paralegals were hired for.
The quality change was quieter but bigger. Validation against the practice system now catches mismatched matter numbers and implausible dates at intake, before they propagate into filings. In the first quarter it flagged a scanned statement attributed to the wrong client — the kind of error that used to be found late, by someone billing at partner rates.
The office manager tunes it now, which was the design goal
Because the pipeline is n8n rather than custom code, the firm's office manager — trained over two sessions — adds document types, adjusts routing, and reads the audit log without touching a terminal. We handed over the workflows, the validation code, runbooks for the failure modes worth knowing about, and a monthly accuracy report the partners actually read.
Meridian & Low owns the entire stack and every credential; the models are swappable behind one configuration step as better ones ship. We check in quarterly to review accuracy drift. So far the trend line only goes one direction, which makes it a short meeting.
Built With
Every other vendor demoed the happy path. Axomble opened with what happens when the model is wrong — the review queue, the audit log, the source passage next to every answer. That is why the partners signed off.Managing Partner · Meridian & Low
More Case Studies
Want Results Like These?
Tell us what you're building — or what's breaking. We'll come back with a concrete plan and a fixed price.
Book a Free Consultation
At Axomble, our expertise lies in creating brand loyalty. We are dedicated to forging connections that lift your name and fame, making you the world's leading brand.
© Copyrights 2023-2026 | Axomble | All Rights Reserved