Skip to content
Guide

AI Document Chasing for NZ Accounting Firms

By Nic Fouhy••13 min read
AI Document Chasing for NZ Accounting Firms

The annual accounts for a Levin building company are finished, near enough. The coding is done, the workpapers are tidy and the provisional tax is worked out. One thing is missing: the February statement for a business savings account the client opened last summer and forgot to mention.

The accountant emailed about it three weeks ago and followed up last Friday. The builder has seen both emails. He read them on his phone at smoko, meant to sort it that night and then spent the evening pricing a deck. Nobody in this story is doing anything wrong. The job still cannot be signed off. The invoice still cannot go out.

Every practice has a shelf of jobs like this one. Each is waiting on a statement, a logbook, an interest certificate or a signature and each takes a few minutes of somebody's attention every week it sits there.

We build AI agents that take over the asking. The agent picks up the query when the accountant raises it, sends the client a short text with a secure link, checks that what comes back is the thing that was requested, files it against the job and tells the accountant it is ready. Here we cover how that works, what your practice software already does, where the agent stops and a person takes over and how to keep it inside the Privacy Act.

AI document chasing flow for an accounting firm, from a query raised in practice software through a text with a secure link, client upload from a phone, the agent checking and filing the document, to the job marked ready for review
Query raised, text sent, document uploaded, checked, filed. The accountant sees the job again when it is ready.

Why do finished jobs wait so long on one document?

Because the request arrives by email, at a desk the client is rarely sitting at, about a deadline that feels a long way off. Most small business owners do their admin on a phone in the gaps of the day. An email with four queries and a PDF checklist asks for a block of time they do not have at that moment.

The calendar plays a part too. Returns are due on 7 July for people filing their own and clients of a tax agent with an extension of time generally have until 31 March of the following year. That extension is a real benefit to the client. It also means the missing statement has no urgency for them until well after it has become urgent for you.

What does the waiting cost a practice?

Three things. You can put your own numbers on each. The first is time: every stalled job gets opened, checked and closed again by someone, often more than once. Count the jobs sitting in your "waiting on client" state today and multiply by ten minutes a week.

The second is lockup. The work is nearly all done and paid for in wages. It cannot be billed. The third is the hardest to measure. Graduates and seniors trained for years to do accounting and the follow-up email is the part of the week few of them enjoy.

Why does a text work where an email stalls?

A text reaches the client where they already are and it asks for one thing. "Hi Dave, we need the February statement for the savings account to finish your accounts. You can upload it here" can be dealt with in the ute in two minutes, because the banking app is on the same phone.

The channel matters less than the size of the ask. One item, described in the client's words, with a link that opens a camera or a file picker. Firms that already send requests this way by hand will recognise the effect.

What does an AI document chaser do?

It runs the request from start to finish. The accountant flags what is missing in the practice software, in a sentence. The agent turns that into a client-friendly text, sends it with a single-use link, watches for the upload, reads the document to check it matches the request, names and files it and moves the job to ready for review.

Phone screen showing a client upload page for an accounting firm with a short checklist of three requested documents, one ticked as received and a button to take a photo or choose a file
What the client sees: a short list, in plain words, with one button

Language models suit this work because both ends of it are unstructured. The accountant's note says "need Feb stmt for the 02 savings a/c, also FBT logbook". The client needs to read "the February statement for your business savings account" and "your vehicle logbook for the Hilux". Translating between those two is work a language model does well. It is the idea behind our view that AI made language programmable.

What starts a request?

A change your team already makes. In most NZ practices that is a job moving into a "query with client" state in Xero Practice Manager, FYI or Karbon, with a note saying what is needed. The agent reads the note and the client's mobile number from the same record. Nobody fills in a second form.

For the first few weeks we have the agent draft the text and wait for the accountant to approve it. Once the wording is reliably right for the common requests, those go out on their own and only the unusual ones wait for a person.

Can the agent tell whether the right document came back?

For common documents, yes, well enough to be useful. A vision model can read an uploaded PDF or phone photo and tell you that it is a bank statement, which account it is for and which dates it covers. If the client was asked for February and has uploaded March, the agent replies straight away, thanks them and asks for the right month.

That quick correction saves the most time. Without it, the wrong statement sits in the folder until the accountant next opens the job and the cycle starts again. The agent also renames the file to your convention and saves it to the right place.

We keep one boundary firm. The agent checks that a document is the one requested. It does not code transactions, judge whether an expense is deductible or change anything in the ledger. Uploads it cannot read with confidence go to a person, labelled as unchecked.

Doesn't our practice software already do this?

Partly. Use what you have first. Karbon has client requests with automatic reminders built in. FYI integrates with Content Snare for collecting documents. If your firm has these available and has not switched them on, that is this month's job. It needs no custom build.

A custom agent earns its keep in the gaps those tools leave. Three come up most often: sending by text where the built-in tool only emails, checking the content of what was uploaded and turning an accountant's shorthand into a request the client understands without a call. If none of those is your bottleneck, the off-the-shelf route is the right one and we will say so.

Which systems does the agent connect to?

The ones your jobs already live in. Xero Practice Manager has an API covering jobs and clients. FYI syncs job states from XPM and has its own automation features. Karbon has an API too. Which of them the agent reads from and writes to depends on where your team raises queries today. We confirm what each connection allows before quoting a build.

We log everything in our builds. Every text, every upload and every reminder is written back to the client record. A partner should be able to see what the agent said to their client and when, in the same place they see everything else.

How long does a build take?

For a practice on a common stack, a few weeks from first conversation to a pilot. The first week goes on mapping how queries are raised today and choosing the job states that act as triggers. Then comes the build, a test run against dummy jobs and a pilot with a small group of clients the partners choose.

Take the pilot slowly. Pick clients who know the firm well, tell them it is coming and read every message for the first fortnight. Costs depend on the integrations involved. We have laid out typical ranges in what AI implementation costs an NZ SME.

How do you chase without annoying clients?

By being specific, brief and easy to stop. A client who gets one clear request in working hours, signed with the name of the person they deal with, tends to be relieved that the task has shrunk to one photo. Trouble starts when messages are vague, frequent or sent at nine at night.

Reminder timeline for an AI document chaser showing a first text on day one, a gentle reminder on day four and a handover to a person at the firm on day ten
Two messages, then a person. The agent never sends a third.

We write the limits into the agent as hard rules. Messages go out on weekdays within set hours. A request gets one reminder, a few days after the first text. Every message says who it is from and how to reach that person.

When does the agent hand over to a person?

In three situations. When the client replies with anything other than the document, such as a question, a "can I call you about this" or a note that the account was closed. When the reminder has gone unanswered. When the request involves something sensitive, such as a query about drawings or an overdue tax payment, which should come from the accountant from the start.

On handover the agent stops sending, puts the client's reply in front of the accountant and says so on the job. The relationship belongs to the firm. We have written more about getting that moment right in the escalation experience.

How does this sit with the Privacy Act 2020?

Comfortably, if the text carries nothing sensitive and the link does the work. A text message can appear on a locked screen and can be read by whoever picks up the phone. So the message names the document in general terms and holds no IRD number, no balance and no figures. Everything sensitive stays behind the link.

Split diagram for an accounting document request: the text message contains only a first name, the document name and a link, while IRD numbers, bank statements and figures stay behind the secure upload link
What goes in the text and what stays behind the link

The link is single-use and expires. Once the document is uploaded, or after a set number of days, it stops working and shows a short note saying the request is closed. Uploads go straight to your document system and are not kept anywhere else.

Where is client data processed?

Wherever you agree it will be, in writing, before the build. The Privacy Act 2020 requires reasonable security safeguards for personal information (principle 5). The Privacy Commissioner's guidance on sending information overseas says that using an offshore cloud provider to store or process data on your behalf is not treated as a disclosure under principle 12, provided the provider is not using it for its own purposes. Responsibility for the information stays with the firm.

For an AI build, that means choosing a model provider whose terms rule out training on your data, knowing which region it runs in and deleting uploads from any intermediate storage once they are filed. We covered the wider version of this for another regulated sector in the Privacy Act and AI for insurance claims. Your own professional obligations on confidentiality sit on top and the partners should sign off the data flow as they would for any new supplier.

Practical Takeaway

Begin with a count. Run a report of every job in a waiting-on-client state and for each one note what is missing. In most practices a handful of document types make up the bulk of the list: bank statements, interest and dividend certificates, vehicle logbooks, loan summaries and signatures.

Next, check what your current software can already send and remind for and turn it on. Then try the text approach by hand for a fortnight on ten stalled jobs. One item, one message, sent mid-morning. You will learn whether the channel is your bottleneck before spending anything.

If it is, the build is contained: a trigger from your job states, a text with a secure link, a check on what comes back and a full log on the client record. If you would like to talk through how that would sit in your practice, get in touch.

Frequently asked questions

Can AI chase clients for missing tax documents?
Yes. An AI agent can pick up a query raised in your practice software, send the client a short text with a secure upload link, check that the document matches the request, file it against the job and tell the accountant it is ready. Questions from the client and unanswered reminders go to a person.
Does this work with Xero Practice Manager, FYI and Karbon?
Yes. Xero Practice Manager and Karbon both have APIs and FYI syncs job states from XPM. The agent uses a job state change as its trigger and writes every message and upload back to the client record. Karbon and FYI also have built-in or integrated document request tools that are worth switching on first.
Is it safe to request tax documents by text message?
It is, provided the text contains nothing sensitive. The message names the document in general terms and carries a single-use, expiring link. IRD numbers, balances and the documents themselves stay behind that link and uploads go straight into the firm's document system.
Will automated reminders annoy our clients?
Not when they are specific, brief and limited. We set the agent to send one request and one reminder, on weekdays within set hours, signed with the name of the accountant. Any reply other than the document stops the sequence and goes to a person.
What happens if a client uploads the wrong document?
The agent reads the upload, sees that it does not match the request, such as the wrong month or the wrong account and replies straight away asking for the right one. Uploads it cannot read with confidence are passed to a person and labelled as unchecked.

Relatable? Get in touch for a chat.

EmbedAI builds these systems for NZ businesses. Ready to talk if it applies to yours.

Thanks, . I'll be in touch.