AI client document collection.

Why document collection is a good AI workflow

Client document collection is the slowest part of year-end work and one of the least skilled. Someone sends a list, clients send some of it in the wrong format, someone opens every file to see what it is, and someone chases the rest, often three times. AI is useful here because most of the checking is mechanical: whose name is on the statement, which dates it covers, whether page 3 of 4 is missing. The judgement about what the documents mean stays with the accountant.

The design choice that matters is the request list. A generic checklist produces generic chasing. A list built from last year’s workpapers and this year’s ledger (two rental properties, a Kiwibank home loan, a new term deposit) lets the system check against what this client actually needs.

How the workflow runs

Inputs

Last year’s workpapers and document index, ledger accounts that imply documents (loans, rental income, investments), client uploads from the portal, email attachments and phone photos.

Systems

The practice management system and client portal, the document store, email, and model APIs for classification and extraction, configured not to retain or train on client data.

Steps

Build each client’s request list; send it; classify each arriving file; check entity, period, page sequence and legibility; tick off the list; draft a reminder naming what is missing; escalate after the second unanswered reminder.

Human review

An administrator reviews low-confidence classifications daily and approves reminder batches. The client manager decides when the file is complete enough to start, and handles anything sensitive.

Outputs

A live checklist per client, filed and named documents, drafted reminders, and a list of files ready to start.

Measures

Days from first request to complete file, reminders per client, documents needing manual classification, and files started with items still missing.

Risks

Documents for the wrong entity, partial statements, automated messages to clients who should not get them, and identity documents ending up in general prompts.

Illustrative example: a Palmerston North practice at year end

Illustrative example, synthetic numbers: a practice with 600 individual, trust and small company clients expects about 1,800 documents after the 31 March balance date. Today an administrator spends about 6 minutes per document opening, naming, filing and ticking it off, plus two rounds of chasing at 4 minutes per client per round.

Model cost assumption, using OpenAI prices as listed on 19 October 2026: classification of every document with gpt-6-luna (US$0.10 per million input tokens, US$0.50 per million output) at 3,000 input and 200 output tokens costs about US$0.40 per 1,000 documents. Period and page checks on multi-page statements with gpt-6.1-sol (US$2 and US$10) at 8,000 input and 500 output tokens cost about US$21 per 1,000. Running both on all 1,800 documents comes to roughly US$39 for the season.

NZ rules that shape collection

• Keep records properly. Year-end documents are business records. Inland Revenue asks for them to be kept for 7 tax years or more, and if a portal or document store holds them outside New Zealand, that storage needs IRD’s approval, whether held by the practice or its cloud provider. Know where each system keeps files.
• AML/CFT documents. Accounting practices are covered by the AML/CFT Act 2009 when they carry out listed activities such as forming companies, acting as trustee or nominee, or managing client funds. Identity documents for customer due diligence need restricted access and a trained reviewer.
• Information about other people. Clients send documents about tenants, trust beneficiaries and family members. Since 1 May 2026, IPP3A requires agencies that collect personal information indirectly to tell the person concerned unless an exception applies. Check how it applies to your practice before widening what you collect.
• Security and retention. The Privacy Act 2020 requires reasonable safeguards and no longer retention than needed. Pull documents out of mailboxes into one controlled store. Our lesson on protecting customer information covers the basics.

What staff keep

Staff decide when a file is complete enough to start, whether a missing document can be estimated or must be chased, and how to handle clients in difficult circumstances. They approve reminder templates and any message that asks for new categories of information. The accountant, not the system, interprets what the documents mean for the return.

Failure modes to design for

• Right document, wrong entity. Personal statements uploaded for the company, or one spouse’s statement for the other. Match the account holder name to the entity on the request list.
• Partial and repeated files. Page 2 of 4, or last year’s statement uploaded again. Check page sequences and period end dates, not just document type.
• Screenshots instead of statements. A banking app screenshot shows a balance but not the period. Treat it as a pointer, not evidence.
• Messages that look like phishing. Automated reminders with links train clients to click. Send only portal links clients recognise, never ask for passwords, and see our guide to drafting emails with human review.
• Double chasing. A reminder goes out an hour after the client dropped the file at reception. Make every intake channel update the same checklist before reminders run.

Collected documents feed invoice coding and exception review and month end reporting preparation. The accounting sector guide maps the full set, and our AI document processing service covers classification and extraction builds.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.