AI in immigration practice: what works, what doesn't
A candid look at where AI is genuinely useful in a regulated practice, and where it's still a liability waiting to happen.
Every week we get the same question from prospective customers: 'Can your AI write the whole submission letter?' The honest answer is: it can, and you shouldn't let it. Not because the output is bad — modern models produce remarkably fluent, legally-plausible prose — but because your CICC licence is not backed by a language model's confidence.
So where does AI actually earn its keep in an immigration practice? In our experience, three places: structured extraction, drafting scaffolds, and decision support.
Structured extraction is the highest-value, lowest-risk use. Point a well-tuned model at a passport, ECA, IMM 5257, or reference letter and it will pull out the fields you need with accuracy that matches or beats a rushed human. When it's wrong, it's usually wrong in obvious, easy-to-catch ways. Combined with a mandatory human review step, this alone can cut intake time by 70%.
Drafting scaffolds is where most firms start and where most firms get burned. The right pattern is: AI produces a first draft against a firm-approved template, populated from structured case data, with citations back to the source facts. The consultant edits, signs off, and files. The wrong pattern is: paralegal pastes 'write me a PR submission letter for this client' into ChatGPT and ships what comes back. The first is a productivity multiplier. The second is a College complaint waiting to happen.
Decision support — 'is this client eligible for CEC?', 'which PNP fits best?', 'what's the risk in this file?' — is the frontier. Done well, it's a superpower for junior staff. Done badly, it's automated malpractice. The critical guardrails are: only reason over verified case data, always cite the specific IRCC program version being used, always show the reasoning chain, and never let the AI answer directly to the client.
The uncomfortable truth is that AI in immigration is not a product feature — it's an operating discipline. The firms that win with it are the ones who treat every AI output as a junior articling student's first draft: useful, fast, and requiring supervision.