AI for Accountants and Advisors: Where the Time Really Goes

Ask a practice principal where their week went and the answer is rarely the technical work. It went on emails asking a client for the same bank statement a fourth time, and on the onboarding paperwork for a new client that has been sitting half finished for a fortnight.

That is the honest starting point for AI for accountants in NZ. The compliance work is the part you charge for and the part that requires your licence. The communication around it is the part that quietly consumes the hours, and it is the part that automates cleanly.

Before any of that, though, there is a question this audience always asks first, and it deserves a straight answer rather than a paragraph near the bottom.

Confidentiality comes first, because it should#

Client financial information is the most sensitive material most practices hold, and your professional obligations around it do not soften because a tool is convenient. So the vendor conversation happens before the pilot, not after.

Four things need answers in writing. Whether your data is used to train the vendor's models, and how you switch that off. Where the data physically sits, since most of these services run offshore. How long conversations and uploads are retained, and whether you can force deletion. And who inside the vendor can read a conversation, under what circumstances.

Most reputable business-tier products answer these well. Consumer-tier free plans frequently do not, and the difference between the free tier and the paid tier of the same product is often exactly this. A staff member pasting a client's financials into a free chat tool is the realistic risk in most practices, not a vendor breach, which means an internal policy about which tools are approved matters more than the vendor's security page.

The New Zealand privacy principles apply to all of this, particularly around disclosure to an overseas provider and how long you keep information. The plain-language version is in the article on the Privacy Act and AI, and the practical vendor questions are set out in AI, customer data and privacy. Neither is legal advice and neither replaces a conversation with your own adviser or your professional body.

There is a segmentation approach that works well in practice. Decide which categories of information may go near an AI tool at all. Deadlines and general process questions sit in one bucket. Client names attached to financial positions sit in another, and the second bucket either stays out of the tools entirely or goes only into a system where you have the vendor answers in writing.

Chasing documents is the highest-volume job in the practice#

Every practice runs the same loop from April onward. Send the request, wait, send the reminder, wait, send the slightly firmer reminder, eventually ring. Multiply by a few hundred clients.

The loop is entirely rules-based, which is what makes it a good automation candidate. A system sends the initial request with a clear list of what is needed, follows up on a schedule you set, escalates to a staff member when a client goes silent past a threshold, and stops the moment the documents arrive. Nobody on your team has to hold a mental list of who has not replied.

Two design details make the difference between a system that works and one clients ignore. The request has to be specific to that client, listing the four documents you actually need from them rather than a generic checklist of twenty. And there has to be a single, obvious way to send things back, because half the delay in most practices is a client who does not know whether to email, post or upload.

The escalation rule matters too. A client who has ignored three reminders does not need a fourth automated message, they need someone to ring them. Set the threshold where the automation stops and a human starts, and hold to it. The general shape of that transition is covered in the handoff.

The same twenty questions, answered instantly#

Every practice has a set of questions that arrive constantly. When is my terminal tax due. What do I need to bring to the annual meeting. Has my GST return been filed. How do I add a new employee in payroll. Am I able to claim this vehicle expense in principle.

These are process questions rather than advice questions, and an AI agent trained on your firm's own material handles them well. It needs your service list, your deadlines calendar, your onboarding process and your fee structure, plus a firm set of rules about what it must never attempt. What it does not need is your entire website dumped into it, which tends to make the answers worse rather than better. The specifics of what to feed a system are in how to train an AI agent on your business.

The boundary has to be drawn hard and early. An AI agent must not answer a question that requires professional judgement about a specific client's circumstances. "When is the GST return due for a six-monthly filer" is a process question. "Should I register for GST" is advice, and it goes to a person every time, without exception and without the system offering a partial view first.

That distinction is not a nicety. A partial answer that a client acts on creates a problem that will land on your desk with your name attached, and no efficiency gain covers that.

Onboarding is where new clients form their opinion of you#

The gap between a client saying yes and the first piece of real work is the most fragile period in the relationship. It is also full of admin, engagement letters and identity verification sitting between the two.

Sequencing that automatically shortens the gap considerably. The client gets a clear list of what happens next, each item arrives when it is needed rather than all at once, and the practice sees at a glance which step each new client is stuck on.

Identity verification requirements deserve care here. Automate the request and the reminder, keep the actual verification decision with a person, and store what you collect the way your obligations require. This is one of the places where a tidy automated process and a compliance failure can look very similar from the outside, so document what the system does.

Where AI for accountants in NZ should not go near the work#

I would keep it away from anything that touches a professional opinion. Drafting a summary of a client meeting for a person to check is fine. Producing advice that reaches a client without a qualified person reading every line is not, regardless of how good the draft looks.

Reconciliation and coding sit in a middle zone. Your ledger software's own automation is generally better suited to that than a general AI tool bolted on, and it already sits inside a system with an audit trail. Adding a second layer usually creates reconciliation work rather than removing it.

Anything client-facing that expresses certainty should be reviewed. A confident sentence about a deadline that turns out to be wrong costs a client money and costs you a relationship.

A sensible first quarter#

Pick one workflow and instrument it before you change anything. Count how many document-chase emails your team sent last month and how long the average client took to respond. That is your baseline, and without it you will not be able to tell whether the automation helped.

Then automate that single loop, run it for a filing period, and measure the same two numbers again. If the response time drops and your team stops holding the list in their heads, extend to the routine question load. Cost expectations across the local market are set out in the guide to AI solutions for Auckland businesses, and if you are weighing this project against others competing for the same attention, the five-question test will settle it faster than a vendor demo.

The work stays yours. Getting the twelve emails a day off your senior staff so they can do that work is the whole point.