AI could cut the back-and-forth that slows most cases down, take routine checking off our reviewers, and get applications to INZ sooner. This page covers what we could bring in, what each piece gains us, and what to watch out for. Everything runs on Microsoft Azure, which we already use, with personal data processed in Australia.
Much of the time lost today goes on rounds of back-and-forth. A candidate uploads something unusable, it waits in the queue, a reviewer spots the problem days later and writes a note, and the candidate starts again. AI helps most by catching problems earlier and taking over the routine reading, so our people spend their time on judgement.
Problems are caught while the candidate is still on the page. Every rejection we avoid saves a review, a note, a follow-up and days of waiting.
AI does the proofreading: matching numbers, dates, names and page counts. Reviewers start from a list of findings and decide.
Fewer rounds plus faster reviews means applications reach INZ sooner. That's what clients and employers notice most.
Expired certificates, name mismatches and typos in passport numbers get caught before lodgement, not after an INZ request.
The same team handles more cases. New staff get up to speed faster with case summaries and a searchable assistant.
Instant, clear feedback in plain English, which is often their second language, instead of a confusing rejection email days later.
≈ 3 minutes
A typical candidate costs about US$1.40 in AI fees. If staff time costs around US$30 an hour (an example figure; we'd use our own), that's under 3 minutes of staff time. Avoiding a single rejection round takes more than that. Several of the first features cost nothing per use at all.
We already run on Azure, so there's no new vendor or contract. The upload process already has a step reserved for reading documents that was never switched on, and documents already have a place to store extracted details. The first features are mostly about connecting, not rebuilding.
Each feature runs quietly in the background first, so we measure what it would have caught before anyone relies on it. People make every decision, and each feature has its own off switch.
Catch problems at the moment of upload, while the candidate is still there to fix them, instead of days later in review. Every warning is a heads-up the candidate can override, never a hard block.
When a photo is picked, we check in the browser whether it's blurry, too dark or too small. If it is, the candidate is asked to retake it. This happens on their own device, so nothing is sent anywhere.
"This photo looks blurry. Please retake it in good light, holding the camera steady."
Limits: catching glare, cut-off corners or missing pages is harder. We'd add those only once we've tested how often they give false alarms.
What we gain: fewer "unreadable, please re-upload" rejections. The fix takes the candidate seconds instead of days, at no running cost.
Azure can learn to tell document types apart: passport, birth certificate, bank statement and so on. If an upload doesn't look like what we asked for, the candidate sees a gentle warning.
"This looks like a driver licence, but we asked for your passport photo page. Upload anyway?"
Limits: it has to be trained with real examples of each type. We'd start with the five or six most common documents and say "not sure" rather than guess.
What we gain: wrong files are sorted out before they take up a reviewer's time.
We read the passport number, full name, date of birth, nationality and expiry date. The two machine-readable lines at the bottom of a passport contain check numbers, which catch some reading mistakes in the passport number and dates.
Passport uploaded → number, name, DOB and expiry appear pre-filled for the candidate to confirm.
Limits: the check numbers don't cover the name, and they don't prove the passport is genuine; a forged passport can still have valid check numbers. The candidate always confirms the values.
What we gain: no retyping, and fewer typos in passport numbers and dates, which otherwise can carry all the way into the application.
The same idea works for birth and marriage certificates, police certificates, bank statements and English test results. We pull out the facts each one needs, such as names, issue dates, statement period and test scores.
Police certificate uploaded → issue date 14 Feb 2026 detected and compared with the validity period set for this visa type.
Limits: each new document type needs setup and testing, and custom extraction costs more per page than passports. We'd add types one at a time, starting with the ones rejected most often.
What we gain: reviewers stop hunting through pages for dates and periods; the key facts are already laid out.
Details already read from documents fill in the matching questions on forms, so candidates don't type the same thing five times. They always see and confirm the values before anything is saved.
The partner's passport details are already filled in on the family form.
What we gain: forms get finished faster, and answers match the documents because they come from them.
Before a candidate hits Submit, they get a short list of things to double-check: a name that doesn't match their profile, a document past its validity date, a missing page. The rules come from each visa type's settings, maintained by our advisers.
"Your passport expires in 4 months. This visa type is set to require at least 6 months. Please check with your adviser."
What we gain: candidates fix their own mistakes before anything reaches our queue.
AI does the first read, so reviewers start from a list of findings instead of a blank page. The reviewer still makes every accept or reject decision.
Next to each submitted file, the reviewer sees what AI found, with the exact text and page each finding is based on. It shows a clear status: no issues found, issues found, not sure, or couldn't check. That way "no issues" never just means the check didn't run.
"Passport number on the document (LA1234567) differs from the profile (LA1234576)."
"Bank statements cover 3 months; this document requires 12."
Limits: results follow the same access rules as the file itself, so reviewers only see findings on files they can already see. As soon as a file changes, its findings are hidden until it's checked again.
What we gain: faster reviews, the same checks applied by every reviewer, and new reviewers productive sooner.
We flag passports that expire too soon and police or medical certificates that may be too old. These are plain rules, so the result is always the same and easy to explain. The limits are set per visa type and document, and measured from the right date (for some documents, when the application is made; for others, when INZ receives it), with exceptions for documents already given to INZ.
"Medical certificate dated 2 Jan 2026 may be more than 3 months old by the time INZ receives it. Please check."
What we gain: out-of-date documents never slip through to lodgement, with fully predictable results and no AI fees.
We compare names across all of a candidate's documents and spot differences a simple rule would miss, such as different spellings, name order, a missing middle name, or a married name.
"Birth certificate says 'Nguyen Thi Mai'; passport says 'Mai Thi Nguyen'. Likely the same person (name order). Please confirm."
What we gain: name mismatches get dealt with at the start of a case, not discovered at the end.
Once the reviewer picks a rejection reason, AI drafts a clear, polite note to the candidate based on the findings. The reviewer edits or replaces it.
"Your bank statements only cover January to March. Please upload statements for the last 12 months, April 2025 to March 2026."
Why the reviewer picks first: if AI pre-selects the decision, people tend to just accept it. Letting the reviewer choose keeps the judgement with them.
What we gain: clear, specific notes mean candidates get it right on the second try, and reviewers stop writing the same note from scratch.
For documents not in English, reviewers get a machine translation beside the original so they can understand the content. It's only an internal reading aid and is never submitted. Translation requirements for INZ depend on the visa and document type.
Chinese household register uploaded → English working translation shown beside the original.
Limits: Azure's standard Translator service processes Asia-Pacific requests in Japan or Southeast Asia. To keep this data in Australia, we'd translate with the in-region Azure OpenAI models instead.
What we gain: any reviewer can work on any language straight away, without waiting for a colleague who reads it.
The queue shows which files have AI findings, so reviewers can spot clean files and problem files at a glance. This sits alongside our existing deadline-based priority and doesn't override it, and we'd keep an eye on overdue cases.
Queue shows "3 issues found" on one file and "No issues found" on another.
What we gain: managers can see where the problems are piling up and plan the team's day around it.
For any service request, we produce a one-page summary: what's in, what's missing, risks found, and the latest conversation. The system calculates all the facts and numbers; AI only turns them into readable text. It's useful before a call, at handover, or when someone is on leave.
"Family of 4. 18 of 22 documents accepted. Missing: partner's police certificate. Risk: child's passport expires before the requested visa end date."
What we gain: anyone can pick up a case in a minute instead of reading its whole history. Handovers and leave cover stop being painful.
When a candidate sends a message, AI drafts a reply using their case status and our approved guidance. It only uses information the adviser can already see, and internal notes are never included. The adviser checks, edits and sends it. Nothing goes out automatically.
Candidate: "Why is my application still pending?" → draft explains which 2 documents are still waiting and what to do next.
What we gain: quicker replies and consistent answers, so candidates feel looked after and chase us less.
Incoming messages are tagged as a status question, upload problem, visa question or complaint, and marked by urgency. Visa questions are routed to a licensed adviser.
"Urgent · Complaint" goes to the top; "Status question" can be answered from a draft.
What we gain: urgent issues and complaints don't sit in an inbox, and each message reaches the right person first time.
We check INZ and Home Affairs updates every day. When something changes, AI summarises it in plain English and suggests which of our visa types and checklists might be affected. An adviser confirms before anything is changed; the system never edits requirements on its own.
"INZ updated AEWV partner work rights on 1 Nov. Possibly affected: AEWV Partner and AEWV Dependent Child checklists."
What we gain: we hear about rule changes the same day, and checklists don't quietly go out of date.
Paste an INZ requirements page and AI drafts the checklist and document requirements for that visa type: which documents, what details to collect, and how long each one is valid. An admin reviews every line before saving.
A first draft of a new visa type in minutes, instead of an afternoon of copying from the INZ site.
What we gain: new visa types go live faster, so we can take on new kinds of work sooner.
AI helps write and polish email templates and document instructions so they're clear and consistent, and can rewrite them in simpler English.
"Make this instruction easier to understand for someone whose first language isn't English."
What we gain: clearer instructions mean fewer "what does this mean?" messages and fewer wrong uploads.
There are three separate kinds, for different people. Every answer links back to its source, and the assistant is designed and tested to say "I don't know" rather than guess.
We look at which documents get rejected most and why, how many times they're re-uploaded, and how long they take. Then we rewrite the instructions for the worst ones, with AI's help, and watch whether rejections drop. The first part is plain reporting, so we can start straight away with no AI risk.
"Bank statements: 34% rejected, mostly 'not enough months' and 'no bank stamp'. Suggested new instruction attached."
What we gain: we fix the cause of rejections once, for every future candidate, and can show the improvement in numbers.
Staff search INZ guidance, Home Affairs pages and our own procedures in plain English. We'd start with a smart search that returns the official text and its effective date. Once that's reliable, we add written answers with sources. Content is kept separate by country and visa type.
"Can a student visa holder's partner work full time?" → the exact manual section, then a short answer citing it.
What we gain: answers in seconds with the source attached; senior advisers get interrupted less and new staff can look things up themselves.
Candidates can ask how to use the portal, what's still outstanding, and where to find our instructions. Questions about how the rules apply to their own situation go to their adviser. Today's help chatbot only matches fixed questions and answers, so this would be new work.
"How do I upload a multi-page PDF?" "What's still missing for my partner?"
What we gain: candidates get help at any hour, and advisers get fewer "how do I…" messages.
We warn employers ahead of time when workers' visas are due to expire, and turn their sponsor obligations into a simple to-do list.
What we gain: a reason for employers to keep coming back, and a feature competitors may not offer.
Agencies upload a CV and the candidate profile fills itself in. We then suggest which candidates fit a job, with the reasons. We'd need to check matching for bias before using it.
What we gain: agencies skip the data entry and fill roles faster.
Matching a passport photo to a live selfie gives evidence that it's the same person. Checking whether the passport itself is genuine is a separate problem and needs a specialist tool.
Limits: Microsoft must approve us for face matching, and that isn't guaranteed. Biometric data also needs its own consent and deletion rules, plus a manual alternative.
What we gain: stronger protection against someone applying with another person's documents.
All of these are under our existing Microsoft account, so there's no new vendor. But being under the same agreement doesn't remove the privacy review above, and "created in Australia East" doesn't always mean "processed in Australia". The table shows the difference.
| Service | What it's for | Where data is processed | Indicative price |
|---|---|---|---|
| Azure Document Intelligence | Reading documents, spotting document type, pulling out passport and certificate details | Australia East (the resource's region). Results kept 24 hours by default | Read $1.50 · Classify $3 Passport $10 · Custom $30 per 1,000 pages |
| Azure OpenAI, regional standard | Checks that need judgement, summaries, drafts, translation, the knowledge assistant | Australia East. For text generation on pay-as-you-go, only GPT-4o and GPT-4.1 mini are offered this way | Cents per case |
| Azure OpenAI, regional provisioned | The same, if we need newer models (GPT-4.1, GPT-5 family) | Australia East, on dedicated capacity | Capacity charge |
| Azure AI Search | Searching guidance for the staff assistant (or use our existing Postgres to start) | Australia East | Monthly, by size |
| Azure AI Face | Passport-to-selfie matching (later, needs approval) | To confirm | Liveness + verify ~$15.50 per 1,000 |
Not for personal data: Azure's "Global" and "Data Zone" options. Data Zone covers the whole Asia-Pacific region; there's no Australia-only zone. The same goes for the standard Translator service, which processes in Japan or Southeast Asia. These are fine for public content such as INZ web pages.
Azure charges per page and per amount of text processed, not per document, so cost depends on how many pages a candidate has and which documents need custom reading.
| Scenario | How it adds up | AI service fees per candidate |
|---|---|---|
| Light | 20 pages, all standard reading: classify $0.06 + read $0.20 + AI checks ~$0.05 | ≈ US$0.30 |
| Typical | 60 pages: classify $0.18 + 40 standard pages $0.40 + 20 custom pages $0.60 + AI checks ~$0.20 | ≈ US$1.40 |
| Heavy | 200 pages (e.g. long bank statements): classify $0.60 + 20 standard pages $0.20 + 180 custom pages $5.40 + AI checks ~$0.50 | ≈ US$6.70 |
"AI checks" assumes mostly GPT-4.1 mini with some GPT-4o. These are rough estimates from published list prices and cover AI service fees only. They don't include a monthly search service, reserved model capacity, engineering time, or re-uploads. Real figures would need measuring on our own documents, and prices confirming with the Azure calculator for our subscription.
Rough estimates in engineer-weeks, for one full-stack engineer who knows the platform, including tests. Two engineers would roughly halve the calendar time. These are planning ranges, not quotes.
| Piece | Engineer-weeks | What's involved |
|---|---|---|
| Shared foundation | 3–4 | Background job runner, Azure connections, storing results, audit record, per-feature off switches and cost caps, and the trial-run measurement set-up. Built once, used by everything after. |
| Blurry-photo warning | 1 | Browser-only check and the warning message. |
| Rejection analysis | 1–2 | Reports on rejection reasons, re-uploads and time taken. |
| Passport reading | 2–3 | Reading, check-number validation, and the confirm screen for candidates. |
| Expiry and date rules | 2–3 | Per-visa-type settings for admins and the checks themselves. |
| Case summary | 1–2 | Gathering the case facts and the summary panel. |
| Step 1 total | 10–15 | |
| AI first-pass review | 3–4 | Findings with evidence in the review screen, the four statuses, hiding results when a file changes. |
| Wrong-document warning | 2–3 | Training and testing the classifier. Also needs reviewers to label a few hundred sample documents. |
| Drafted rejection notes | 1 | Draft into the existing rejection dialog. |
| Drafted replies | 2 | Draft in the messaging screen, with internal notes excluded. |
| Staff search assistant | 3–4 | Loading and refreshing INZ and Home Affairs content, search screen. Written answers add 2–3 more. |
| INZ rule-change alerts | 2–3 | Daily check, change summary, adviser confirmation. |
| More document types | 1–2 each | Two types assumed in the total. |
| Smaller add-ons | 5–8 | Name consistency, working translation, message sorting, checklist set-up help, email writing (about 1 week each). |
| Step 2 total | 20–29 | |
| Auto-fill forms | 3–4 | Linking form questions to extracted details; this link doesn't exist today. |
| Pre-submit check | 2 | Check list shown before Submit. |
| Candidate help | 4–6 | New assistant to replace the fixed-answer chatbot, with limits on what it can discuss. |
| Step 3 total | 9–12 |
Outside engineering: the legal and privacy review; a few days of reviewer time to label sample documents for testing; and admin time to set the expiry rules for each visa type.