AI as another team member, not another tab
The goal is not one more tool to open, log into and copy data between. It is AI working in the in-between, quietly taking the busywork off your plate so the team gets back to the skilled, people-facing work. We split the opportunities into three layers and tackle them by pain and impact, not strictly top to bottom.
The repeatable jobs
IRD mail · disbursements · document combiningThe high-frequency, judgement-heavy or fiddly tasks that eat hours every week. These are where a custom assistant or a bit of automation pays back fastest, so they lead.
The connections
Xero · FYI · Outlook / SharePoint · IRDWe do not want to replace the tools you live in. Where they expose an API or an export, we connect AI through them so it can search and summarise across your emails, documents and records.
The firm's brain
Context that compounds over timeThe longer-term prize: your calls, emails and documents feeding a shared context so AI understands True North - your clients, your tone, the way you work - rather than starting cold each time.
We mix a fast, visible win with the slower background build, so value lands early while the bigger picture comes together.
Opportunities, in order
Ranked by impact against effort. The structured, well-defined jobs are the cleanest quick wins; the judgement-heavy ones are higher value but need a little more care. Tap any item to expand the detail.
IRD mail triage
The most tedious recurring job - reviewing every IRD letter and deciding what reaches the clientThe pain
- Every IRD letter has to be read and a call made - forward to the client, or hold it back because IRD does not yet know what the firm knows.
- Volume spikes at certain times of the month - some days a handful, other days pages of letters.
- It is judgement-heavy: a client may look overdue, but only because a return is yet to be filed. Get it wrong and clients ring in a panic about penalties.
- It currently sits with one person at a time and is described as genuinely tedious.
What we would build
- An assistant that reads each letter, classifies it, and proposes the call - send on, hold, or no action - with the reason, for a human to approve.
- Rules captured from how the team already decides, including "we have already sent this three times, the client knows" patterns.
- A simple review queue, so the person reviewing confirms rather than starts from scratch on every letter.
- Built in phases - start by summarising and grouping the day's letters, then layer in the suppress-or-send judgement as trust grows.
Xero disbursements reconciliation
The monthly spreadsheet shuffle - reorder columns, apply per-client rules, send to XPMThe pain
- A CSV comes out of Xero in the wrong shape. Columns are copied across into a template in the right order, by hand.
- Per-client discount rules (some clients are not charged full price) live in a separate spreadsheet that has to be referred to.
- Several files open at once, easy to get wrong, and done every month.
What we would build
- A custom assistant (a saved, reusable workflow) where you drop in the monthly Xero export and it returns the template, correctly ordered.
- The per-client discount and GST rules held once in a maintained reference file, applied automatically rather than looked up by hand.
- Structured in, structured out - the cleanest kind of task to automate, and a fast, satisfying first win to show the team what is possible.
Email & record retrieval
Finding "that email where I told the client how much tax to pay" without the digThe pain
- FYI threads collapse sent items into the most recent message, so a specific email you sent is hard to find later.
- Outlook archives or clears older mail after a period, scattering things across folders and years.
- Andy flagged searching the delegate box and archives for the context of a request as a recurring chase.
What we would build
- Connect AI to Outlook (and ideally FYI) so you can ask in plain language - "find the email where we told this client their tax to pay" - and get the thread back.
- Plain-language search and summarisation across mail and documents, rather than exact-match folder hunting.
- Feasibility hinges on what FYI exposes - see the open question on the FYI API.
Combining group papers
Merging group papers and supporting documents, currently a manual Adobe jobThe pain
- Group papers and their supporting documents are combined by hand through Adobe.
- It can take a while, especially for the bigger clients with more documents.
What we would build
- An automated assembly step that merges the papers and supporting documents in the right order, ready to review.
- Lower frequency than the jobs above, so it sits a little lower - but mechanically simple, so it could be bundled in as an easy add-on.
IRD vs internal record matching
Reconciling your records against IRD where the data cannot be bulk-pulledThe pain
Checking two databases against each other - your records versus IRD - to spot where a return or notice does not match. The catch is that IRD will not let you download, say, a CSV of every client texted on a date; you have to open each one.
Why it is parked for now
Without a bulk export or API from IRD, any tool still has to gather the IRD side record by record. We would revisit once the higher-value jobs are landing, or if a workable access route appears.
Pain against ease of build
Disbursements is the cleanest quick win - structured data in and out - so it lands fast and visibly. IRD mail triage is the bigger prize but needs phasing, because it carries real judgement. Email search waits on what FYI exposes; IRD matching waits on access.
The systems, and where AI plugs in
We extend what you already use rather than replace it. The big variable is connectivity - some New Zealand tools expose an API or export, and some do not. That is what decides how cleanly AI can reach into each system.
Xero
The core ledger. Data is pulled out and reshaped in Excel to get what you need. We have built on the Xero API repeatedly, so the disbursements flow is well within reach.
FYI
Document management for the firm. Strong for filing, but its email threading collapses sent items into the latest message, which is what makes finding a specific past email hard.
Microsoft 365
Outlook for email, SharePoint for files. Copilot is available on the licence, and Copilot Studio (same licence) can run agents behind the scenes - triage mail, draft replies, create tasks.
IRD portal
Where the mail and tax notices come from. No bulk export and no API - you have to open each client individually. This is the main limit on the IRD-matching work.
Adobe
Used to combine group papers with their supporting documents, by hand. Slow for the bigger clients - a clear candidate for an automated assembly step.
XPM / XBM
The destination the disbursements template feeds into. Once the reconciliation output is correct and consistent, this stage becomes a clean hand-off rather than a manual reshape.
Open questions
A handful of answers would sharpen the estimates and settle the order. These came straight out of the workshop.
Connectivity
The repeatable jobs
Platform & data
Approach & security
A few principles underpin everything above. The first is the one the team rightly raised first - what happens to your data.
Paid plans keep your data yours
On free plans, data can be used to train the model - think of it as going back into the lake. On paid plans it stays in your own environment and is not retrained on. For an accounting firm handling client and tax data, that is the line: paid only, and the same care you already apply to spreadsheets of client data.
Structured work automates; unstructured work suits AI
A spreadsheet to a Xero invoice is structured in, structured out - ideal for automation (disbursements). An hour of email or a messy thread distilled into a clear summary is unstructured - that is where AI earns its keep (IRD mail, search).
AI as a team member - keep a human in the loop
Treat it like a sharp intern: clear instructions and context, checked work, more responsibility as it earns trust. On tax-sensitive calls, it proposes and a person approves. Nothing goes to a client unchecked.
Reusable assistants beat re-prompting
For anything you do over and over, we set up a saved assistant with the instructions and reference files baked in - shared by a link so the team uses it without rebuilding the prompt each time.
Tool choice - practical over hype
We moved to Claude for its data analysis and practical, useful releases. Copilot is fine if it is already working for you, and Copilot Studio (on your existing licence) is a real advantage for running agents behind the scenes. We would settle on one to commit to rather than half-use several.