Using AI in a migration practice: workflows, review, and data controls

Choose the workflow before the tool

An AI feature is useful only when the practice can define an acceptable result and detect a bad one. Begin with the work: who supplies the input, what output is needed, who checks it, and what happens when the tool cannot complete the task reliably.

Good early candidates include rewriting approved public service information, formatting a generic preparation note, or turning a documented internal process into a checklist. Each has a source against which the result can be checked. Work involving a client’s circumstances requires additional data-handling and professional controls before live use.

Match the review to the consequence

A summary can sound accurate while omitting a detail that changes its meaning. Supplying a source document or connecting the tool to a knowledge library can improve context; it does not establish accuracy or currency.

Run the fictional example

Use this generic preparation note to test a narrow drafting task without client information:

The consultation is an online meeting. The practice sends a meeting link after the booking is confirmed. The person should prepare a short list of questions. Identity documents should not be emailed in response to the booking confirmation. If documents are needed, the practice will explain the approved upload method. The booking confirmation states the fee and cancellation terms.

Use a constrained prompt:

Using only the preparation note above, draft a short consultation-preparation email and a separate internal checklist. Preserve every instruction. Do not add fees, deadlines, document requirements, legal advice, or promises about outcomes. Mark information missing from the source as “to confirm”. Keep the email in plain language. List any source instruction you could not include.

The prompt improves the specification; it does not enforce it. Compare the output with the source rather than accepting the tool’s own statement that it followed the instructions.

Review against explicit acceptance criteria

For this exercise, accept the result only if it preserves the online meeting, the condition for sending the link, the questions to prepare, the restriction on emailing identity documents, the separate upload instructions, and the reference to booking terms. It must not invent the fee or cancellation conditions.

An invented sentence such as “email your passport within 24 hours” would fail three checks: an unsupported document request, an invented deadline, and a contradiction of the source. This is a constructed teaching example, not a reported output from a tested product.

Record the source version, tool or feature, test date, errors found, edits required, reviewer, and final decision. A corrected output is not evidence that the workflow will behave reliably on every later input. Test missing information, conflicting instructions, and a longer example before expanding the task.

Check data handling before connecting client records

OMARA’s AI guidance states that RMAs remain responsible for the assistance they give. It also calls for explaining AI use in giving immigration assistance in advance and obtaining written client consent before entering personal details into an AI platform or system.

That consent does not establish that the product or workflow is suitable. The OAIC recommends against entering personal information into publicly available generative AI tools as a matter of best practice, and sets out product assessment and oversight considerations.

Before considering a live-data workflow, establish:

• Access: which files, matters, and connected accounts the feature can read, and whether it respects existing permissions.
• Use and retention: whether inputs or outputs are retained, used for training, or reviewed by the provider; which terms apply to the actual plan.
• Location and subcontractors: where information is processed and which other organisations can access it.
• Control: whether administrators can restrict the feature, inspect actions, revoke access, and manage deletion.
• Responsibility: the applicable confidentiality and privacy requirements, required explanation and consent, and the named professional reviewer.

An “AI assistant” embedded in existing software still needs this assessment. Brand familiarity or a paid subscription does not answer the questions. Removing a name from a document may also leave a person identifiable through dates, family relationships, employer details, or the history of the matter.

Separate drafting from taking action

Give a trial workflow only the access it needs. Keep draft generation separate from sending messages, changing client records, lodging material, or deleting files. Make the approval step explicit and ensure staff can see the source and proposed action before authorising it.

Treat client documents and retrieved pages as material to analyse, not instructions granting the tool permission to act. Test how the workflow behaves when source material contains an instruction to ignore the task or disclose other information. Content supplied by another party should not be able to expand the tool’s authority. OWASP documents this risk as indirect prompt injection.

Measure the complete task

Count setup, prompting, source preparation, checking, correction, and recordkeeping. For a fictional example, a manual task takes 20 minutes. AI drafting takes 2 minutes, preparation 3, checking 8, and correction 4: the assisted task takes 17 minutes, saving 3 rather than 18. Include setup costs when judging whether repeated use is worthwhile.

Also compare completeness and error rates using the same acceptance criteria. Stop or narrow the workflow if verification takes longer than doing the work directly or if consequential errors are difficult to detect.

Keep reusable approved wording in the practice’s template library. When a fixed template solves the task, use it directly. Reserve generation for work where variation provides a benefit you can demonstrate.