The bottleneck is intake, not judgment
A visa or work-permit filing arrives as a pile of mismatched artefacts: passport scans, employer letters, degree certificates, payslips, prior filings, each in a different format and often photographed rather than scanned. Before anyone can practise law, somebody has to identify what every document is, pull the fields out of it, and notice what is missing.
That identification and extraction layer is where the hours go, and it is genuinely automatable. Classification, field extraction, and completeness checking against a required-document list are pattern work. The legal reasoning that follows is not.
Treat extraction accuracy as a distribution, not a number
Vendors quote a single OCR accuracy figure. It is close to meaningless on its own, because errors are not evenly distributed. A clean employer letter extracts near-perfectly. A phone photograph of a laminated document at an angle, in a non-Latin script, with a handwritten amendment, does not.
The useful question is not the average but what happens at the tail. Does the system know when it is unsure? A pipeline that flags low-confidence fields into a review queue is far safer than one with a higher headline accuracy that fails silently. Ask any supplier to show you their confidence thresholds and what happens below them.
Deadlines make the escalation path the real design work
Immigration work is deadline-governed, and a missed filing window is not recoverable. That changes the failure model. An automation that stalls quietly is worse than no automation, because the team stops watching the thing they used to watch manually.
So the escalation path gets designed first: what triggers a human, how fast, and to whom. Anything touching eligibility, legal strategy, or a client representation stays with a qualified person. The system prepares, orders, and flags. It does not decide, and it does not file.
Use this before you buy or build.
- Every required document type is enumerated before the build, not discovered during it.
- Low-confidence extractions route to a named reviewer instead of passing through.
- Deadline tracking has an alert path that a stalled workflow cannot suppress.
- Eligibility calls, strategy, and filings remain with a qualified practitioner.
Automate the sorting, extraction, and completeness checking. Keep the legal judgment and the filing decision human, and design the escalation path before the happy path.
