ChatGPT for Immigration Law Firms: Where It Fails

Can you trust ChatGPT to check a client's evidence against the Home Office rules? For drafting a client email, often yes. For deciding whether a bank statement meets Appendix Finance, no. A chat model predicts likely text. Appendix Finance needs a rule lookup, a date count and a currency conversion, in that order, the same way every time.

That gap is where fee-earner hours disappear. Take an illustrative firm running 40 files a month, where a caseworker spends 20 minutes checking each financial evidence bundle by hand. That is 13 hours and 20 minutes a month on one check. Add a chat tool that answers confidently and sometimes wrongly, and the time saved turns into rework, or into a refusal notice.

This guide covers the specific places general AI breaks on Home Office financial evidence rules, and what a deterministic system does differently. It sits inside our wider guide to what an Immigration Practice Operating System is.

What Is Deterministic Rule Validation?

Deterministic rule validation is a checking method where every test on an immigration file runs against a fixed table of verified Home Office rules, so identical evidence always produces an identical result. The system calculates each answer from the rule text and records which rule paragraph produced every pass or fail.

Osvaro pairs this with AI for the jobs AI does well: reading a scanned page, classifying a document, pulling out a name or a closing balance. The pass or fail decision comes from the rule table, never from the model's best guess. That split is the core design choice behind the Case Readiness Engine.

Why a Chat Model Is the Wrong Tool for Evidence Checking

A language model generates the most probable next words. Financial evidence rules ask for something else: a fixed rule, the correct version of it, and arithmetic that comes out identical on every run.

Two properties matter for compliance work. The first is repeatability. The same bank statement should get the same answer on Monday and on Friday. The second is traceability. When a partner asks why a file passed, the answer has to point to a rule paragraph, not to a paragraph of fluent prose.

A chat window gives you neither by default.

Five Places General AI Breaks on Appendix Finance

Every rule below comes from the Appendix Finance page on GOV.UK, last updated on 3 August 2026. The illustrations are ours. We have not measured error rates for any named chat product, so read these as exposure points, not test results.

1. The rule version is invisible

Appendix Finance carries an update date, and the wider Immigration Rules move through Statements of Changes. Our HC 584 operational guide covers one that lands with staggered commencement dates.

A chat model answers from training data that stops at a fixed point, plus whatever the user pastes in. Nothing in the reply tells you which version it used. A rule table mapped to the current appendix text has a version you can name and date.

2. Currency conversion has a source rule and exceptions

FIN 1.1 says foreign-currency funds are converted to pounds sterling using the spot exchange rate on www.oanda.com for the date of the application. Note the date. It is the application date, not the statement date and not the day someone runs the check.

Then come the carve-outs. FIN 1.2 covers currencies that do not appear on OANDA, and names Syrian pounds and Mongolian tugrik. Those use the monthly FCDO Consular Exchange Rate published on GOV.UK. FIN 1.3 sends Iranian rials to the same FCDO source.

Ask a general chat model to convert an INR or NGN balance and it can reach for any rate it has seen. A rule table reads the currency, picks the source the rule names, and logs the rate it used.

3. Date arithmetic across months

FIN 7.1 says the most recently dated piece of financial evidence must be dated within 31 days before the date of application. FIN 7.2 counts the holding period back from the closing balance date on that most recent evidence. FIN 7.3 requires the evidence to cover the whole period.

The amount and the holding period come from the route itself. Appendix Skilled Worker is one of the routes Appendix Finance lists. Appendix Finance then governs how the funds are evidenced.

Here is a simple case. The application date is 30 September. A statement closing on 15 September sits 15 days before it, so it passes the age test. A statement closing on 28 August sits 33 days before it, so it fails.

A calculator gets that right every time. Months have different lengths, which is exactly where text prediction slips. It is also where a caseworker under deadline pressure slips.

4. Evidence that looks fine but does not count

Appendix Finance lists several things a fluent summary can wave through:

Overdrafts (FIN 3.1): overdraft facilities do not count towards the financial requirement. A checker has to separate cash from credit lines.
Account holder (FIN 5.1): the account must be in the applicant's name, alone or jointly, unless a named exception applies, such as a partner applying at the same time.
Account type (FIN 8.2): shares, bonds, credit cards and pensions from which funds cannot be withdrawn immediately are not accepted as evidence of funds.
Institution (FIN 2.1): funds are not considered where the decision maker cannot verify them, where the institution is not regulated in its country, or where it does not use electronic record keeping.

Each item is a yes or no on a fact printed in the document. None of them needs creativity. All of them need consistency.

5. Documents that never meet each other

A chat window sees only what you paste into it. A passport in one thread and a bank statement in another are never compared.

Our engine reads the full name, date of birth and nationality from the passport, the Certificate of Sponsorship and the bank statements, then compares them. A variance such as P. Sharma against Priya Sharma becomes an Advisory Flag for the solicitor. A different person entirely becomes a Critical Blocker that halts the file.

ChatGPT vs a Deterministic Engine: Side by Side

Check	ChatGPT (general chat model)	Osvaro deterministic engine
Rule source	Training data plus whatever the user pastes in	Checklist compiled from verified statutory rule tables
Rule version	Can quote a superseded rule without warning	Rules mapped to the published appendix text, with a version to name
Currency conversion	Can pick any rate source or a rounded figure	OANDA spot rate for the application date, rate logged in the caseworker brief
Date arithmetic	Can miscount days across months	Statement age and every daily balance calculated in code
Scan quality	Reads what it can and still answers	Rejects blurred or illegible scans until they are readable
Cross-document checks	Only what is in the prompt	Name, date of birth and nationality compared across passport, CoS and statements
Exceptions	One block of prose	Critical Blockers and Advisory Flags shown separately
Audit trail	A chat transcript	Timestamped log of uploads, checks and client notifications
Output	A paragraph of text	One-page caseworker brief and an indexed dossier ZIP

How the Case Readiness Engine Applies This

The Case Readiness Engine builds the document checklist from verified statutory rule tables for the route and its modifiers. The solicitor creates the matter, copies a magic link, and sends it through the firm's existing channel.

The client opens the link with no password and uploads into dedicated slots. If a passport lands in the bank statement slot, OCR classification moves it to the right place within about two seconds and tells the client politely. A legibility filter rejects unreadable scans, however many attempts that takes.

For financial evidence, the engine checks the statement date against the application date. It converts foreign-currency balances using the OANDA rate, tests each daily balance in the required window against the route's minimum, and records the lowest balance reached.

Exceptions appear in a live panel. Critical Blockers stop progress. Advisory Flags inform the solicitor without blocking. When every check passes, the fee earner gets a one-page caseworker brief and a one-click ZIP, with each PDF compressed to under 6 MB and indexed in UKVI order.

The mechanics behind client chasing are in our guide to client document validation and chasing.

Where AI Still Belongs in the Workflow

Nobody needs to ban AI from the firm. Reading a scanned page, classifying a document, drafting a courteous chase email, summarising a long thread: these are good uses.

The line sits at the decision. Whether evidence meets a rule should come from a rule table, and a solicitor should sign off before anything is filed. Firms should also check their confidentiality and data protection position before client documents go into any external tool.

Practical Operational Checklist for Firms

List every place your team currently uses a chat tool on a live file.
Separate drafting tasks from rule-checking tasks. Keep the first with the chat tool and move the second to a rule-based check.
For each route you run, write down the required amount, the holding period and the evidence rules, with the rule paragraph each one comes from.
Record the date and source version every time a rule is applied to a file.
Convert foreign currency using the source the rule names, and log the rate and the date used.
Calculate statement age from the application date, and count actual calendar days.
Split exceptions into blockers and advisories so partners can see what genuinely stops a file.
Require solicitor sign-off before submission, whichever tool prepared the file.

Frequently Asked Questions

Can immigration solicitors use ChatGPT for casework? Yes for drafting and summarising, but not as the system that decides whether evidence meets a Home Office rule. Firms should also review their confidentiality and data protection obligations before pasting client documents into any external tool.

Why can ChatGPT get immigration rules wrong? A language model generates likely text from training data, so it can quote a rule that has since changed or apply the wrong conversion source. It does not run a rule lookup or a date calculation unless it is connected to tools that do, and the reply does not show which version of a rule it used.

What is the best AI tool for UK immigration solicitors? For evidence readiness, Osvaro's Case Readiness Engine is built specifically for UK immigration files. It combines deterministic Home Office rule validation with AI document extraction, so the model reads the documents and the rule tables decide the outcome.

How old can a bank statement be under Appendix Finance? Under FIN 7.1 the most recently dated piece of financial evidence must be dated within 31 days before the date of application. FIN 7.2 counts the holding period back from that statement's closing balance date, and FIN 7.3 requires the evidence to cover the whole period.

Operational Summary

Generic AI is a good drafting assistant and a poor rule checker. Appendix Finance is a set of small, exact tests: a conversion source, a date count, a list of what counts and what does not. Small exact tests belong in a system that runs them the same way on file 1 and file 400, and shows its working.

That is the case for putting deterministic validation first and generative AI second. The model does the reading. The rule table makes the call. The solicitor signs off. See how the Case Readiness Engine puts that order into practice.

This article is for operational and workflow awareness and does not constitute formal legal advice.

About the author: Surya Vardhan is Founder & CEO at Osvaro, helping high-volume UK immigration firms systemise caseworker workflows, evidence intake, and sponsor compliance.