Airclerk puts Claude to work inside New Zealand accounting practices, on real client engagements and on the systems the firm already runs. We built two things for it. The NZ Accounting Pack is twenty-five skills for the work that fills compliance season, written to the Income Tax Act 2007, the GST Act 1985, the Tax Administration Act 1994 and IRD's own Tax Technical material, and kept verified against those sources rather than remembered. The second is a Xero connection that assumes a client list rather than a single business, and that covers every operation Xero publishes for accounting, fixed assets and New Zealand payroll. Anthropic's nearest accounting release is written for a US business owner; this is the New Zealand practice version. Behind it is a team that spent years implementing systems inside regulated New Zealand firms, including years at Trineo, one of the country's earliest Salesforce partners.
Here for your own finance function rather than client engagements? See finance teams →
The imputation credit account is a memorandum account, so it gets reconstructed from IRD's transaction records, the dividend statements, the share register and last year's IR4J. A company with no dividends and no refunds is a short job. This one has both, which is the version that eats an afternoon. One skill of the twenty-five, and the clearest illustration of what tying a figure to evidence actually costs. Figures are illustrative.
| Movement | Credit / (debit) | Tied to |
|---|---|---|
| Opening balance | 41,200 | Last year's filed IR4J, agreed to myIR |
| Provisional tax, three instalments | 36,000 | IRD transaction records. Paid, not accrued. |
| RWT on interest received | 410 | IRD transaction records |
| Credits attached to the September interim dividend | (19,600) | Dividend statement. $50,400 cash plus $19,600 credits, so $70,000 gross: 28 cents per dollar, the maximum. The year's first imputed dividend, so it sets the benchmark ratio. |
| February refund | (4,100) | IRD transaction records |
| Closing balance, 31 March | 53,910 |
Five sources agreed. The sixth, the one place the account appears to live already, was out by exactly the refund. That is the ordinary outcome of this job, and it is the reason the schedule has an evidence column at all.
Two things this workflow will not do. It will not roll the account to the company's balance date, because the imputation year ends 31 March whatever the balance date is, and that is where most of these go wrong. And it will not file: a debit balance at 31 March means further income tax equal to the debit, plus a 10% imputation penalty tax, so it quantifies both, flags them, and hands the IR4J notes to the CA.
New Zealand accounting firms use Anthropic's Claude models across the bounded, reconciled work that fills compliance season: year-end workpaper assembly tied to the trial balance, provisional tax schedules and client letters, FBT workings, payday-filing reconciliations, IRD correspondence triage and client questionnaires. In public practice Claude needs to be connected to the ledger and practice systems, with engagement isolation so one client's records are walled off from another's, and a reviewer's sign-off before anything leaves the building.
Engagement isolation between clients is a hard rule in every skill rather than a setting someone remembers to switch on. Each output names the client and the period at the top. The job is sorted for materiality before it is worked, so a $500 coding query does not come back as a research memo. Consequential steps stop for the named reviewer, and a schedule that will not tie to the ledger states its reconciling difference rather than carrying a plug figure.
Provisional tax with its four options and use-of-money interest. GST, including the cases the ledger software does not finish: zero-rated land, registration boundaries, change of use. Payday filing and schedular payments, FBT, the bright-line and land rules, and the XRB reporting tiers. Every skill cites to the Income Tax Act 2007, the GST Act 1985, the Tax Administration Act 1994, IRD's Tax Technical material or the XRB, with a last-verified date attached. Where a rate cannot be verified, the skill declines to state the number instead of reaching for one.
Eighteen carry a deliverable end to end: year-end workpaper packs that tie to the trial balance, provisional tax schedules on the IRD calendar, FBT workings, payday-filing reconciliations, GST exception cases, IRD correspondence triage, AML/CFT onboarding support, management accounts with a 12-month rolling cash-flow forecast, and five added in August 2026 for the balances that fill real NZ client files: imputation credit accounts, offshore shares and funds under the FIF rules, livestock on the year's determination values, dairy income tied to the processor statements, and construction contracts with their retentions. Five are tax research skills, covered below. One runs the practice's monitoring rig across IRD, tax policy and XRB sources. One interviews the firm on the way in and writes the practice profile every other skill reads, so the pack knows your client mix, balance dates, systems and reviewers before it does any work.
The pack is an accelerator, not a finished product: it is where an implementation starts, and each firm's version diverges from there as skills are adapted to the firm's own workpapers, templates and risk settings. We provide the skills index behind it: each skill, the work it carries, the sources it verifies against. Ask us about it.
Claude can be given a Xero connection, and Xero ships one itself. For a single business it does the job. A practice needs two harder things: reach across every client, and reach into every part of the ledger the work touches. Ours covers 312 operations, which is every one Xero publishes for its Accounting, Assets and Payroll NZ APIs, generated from a checksum-pinned copy of Xero's own specification.
Raise a $4,200 consulting invoice for Acme.
Two of your clients match Acme: Acme Holdings Ltd and Acme Trading Ltd. Which one?
Trading.
Drafted in Acme Trading Ltd, $4,200 to consulting income. Ready to post?
This is what joins the pack to the ledger. The payday-filing review can reach payroll instead of waiting for an export, and the year-end workpaper pack can read the asset register it depreciates. Completeness here means every operation in Xero's published specification for those three APIs, generated from a pinned copy and held to it by the test suite; running each one against a live organisation is a separate exercise. It is Airclerk's own tool rather than a Xero product, and not affiliated with or endorsed by Xero. We install and configure it as part of an implementation, and which ledgers it is pointed at stays the firm's call.
New Zealand tax settings move constantly, which is why nothing in the pack is stored. The first half of 2026 alone, every date below verified against the primary source on 11 July 2026:
The pack does not memorise any of this. Every skill verifies rates and dates against IRD and legislation sources when it runs and carries the date it was last checked, and a monitoring skill sweeps the Tax Information Bulletin, taxpolicy.ird.govt.nz and the XRB alerts monthly, holding the line between announced, enacted and in force. Last full re-verification across the original twenty skills: 9 August 2026; the five added in August were verified on 16 August 2026, as they were built.
A tax question that ends in a memo fails differently from a workpaper pack. Not a number that will not tie, but a confident answer resting on a search result nobody opened. Five of the twenty-five skills do that work, and they run one method between them.
Every run is routed Lookup, Quick, Standard or Deep from the complexity of the question and how far the firm intends to rely on the answer, and the output names the depth it ran at. A $500 coding query does not get a research memo.
A search result is not a source. Each run separates what it relied on from what it reviewed and set aside, and discloses the leads it never opened at all. legislation.govt.nz refuses a good share of automated requests, so there is an order of fallbacks behind it, and the output names which one supplied the text rather than implying it read the section.
Every research output goes looking for what cuts against the position it is building, and records what it found. A question about an earlier income year either confirms the version of the law in force at the time or says it could not, instead of quietly answering with current law. IRD material is classified by whether it actually binds, and a Bill or a consultation is never described as law.
Nothing is computed in a research note. A figure that matters is produced by the skill that owns that deliverable, where it has to tie to the ledger or carry a stated reconciling difference. Public case coverage in New Zealand is incomplete and the outputs say so rather than implying a citator. No client name, figure or identifying detail goes into a public search query. Each run ends in a stated result state, so a partial or blocked run cannot be mistaken for a finished one.
The five split by the job in front of you: a question to answer, a single provision or ruling to status-check before it is relied on, client documents to work through before any conclusion is formed, someone else's draft to check, or a date-bounded sweep of what changed in a window. Every output is draft research for a Chartered Accountant to review. None of it is advice.
The patterns we've seen produce the cleanest path from pilot to production. The implementation anchors on year-end workpapers; most of the rest of the pack feeds that file or follows it.
Lead schedules per balance-sheet line, tied to the trial balance, with queries listed and reviewer notes flagged by severity.
The cases Xero doesn't finish: zero-rated land, registration boundaries, adjustments and change of use, flagged before filing.
Instalment schedules by balance date and option, use-of-money interest exposure shown, client letter drafted.
IRD correspondence from the firm's approved feeds, classified by client, deadline and severity; drafts for routine items, escalation for the rest.
Agency-list filing progress, missing-information status, 31 March risk flags and terminal-tax reminders, with questionnaires and chasers drafted in the firm's tone.
Vehicle, loan and unclassified-benefit calculations with the rate options laid out for the reviewer's call.
Payday-filing totals against the ledger, ESCT bands checked, exceptions tabled for review.
Captured-activity checks, CDD level and beneficial-ownership maps under the firm's existing programme.
A monthly practice briefing from IRD, taxpolicy and XRB sources, filtered to what touches your client base.
P&L against budget, working capital, a 12-month three-way forecast with stated assumptions, and owner commentary for the reviewer's sign-off.
A substantive question worked at a stated depth, every claim cited to a source that was opened, with a contrary-authority pass and a result state on the front page.
Someone else's memo or letter checked claim by claim against official sources, cited IRD material status-checked. It corrects and never rewrites.
Accruals, prepayments, depreciation and reclassifications drafted from the workpaper that computed them, posted on the reviewer's approval with the supporting schedule attached to the journal.
Aged receivables, coding exceptions or a trial balance pulled straight from Xero for a named client, with writes gated behind the organisation they land in.
Livestock movements reconciled by class on the year's determination values, and dairy income tied to the processor statements with deferred milk income at balance date.
Retention schedules both directions and WIP reconciled to progress claims, with the Construction Contracts Act retention checks tabled for the reviewer.
The ICA rolled on its 31 March year and dividends tested against the benchmark and maximum ratio rules, with a debit closing balance flagged for the reviewer.
Offshore shares and funds gated through the exemptions first, then a method proposed with the workings tied back to the investment statements.
Behind every workflow sits AI Process Assurance: the reasoning, sources and sign-off retained as an AI review record, the AI Workpaper, so the firm can show what the AI did, how it was checked and who signed off.
The implementation is built around year-end workpaper preparation: the highest-volume work in the practice, with a trial balance to tie back to and a review a partner already knows how to run. Most of the other skills in the pack support that job, from the GST and payroll reconciliations that feed the file to the chasers that unblock it. Final numbers, tax positions, structuring advice and anything filed stay behind stricter gates, with the accountant.
Each workflow defines who reviews what: partner, manager or the engagement's responsible reviewer. The AI Workpaper records which sources were attached, what changed in review and whether any required gate was missed.
We run the workflow on held-out client jobs before an accountant relies on it: does every schedule tie to the ledger or state its difference, does it verify rates and dates against current sources, does it respect engagement isolation, does it route consequential outputs to the right reviewer. Prompt-injection and cross-client-leakage checks are part of the production gate.
The configured workflow on your systems, a Workflow Charter setting out stages, evidence and approval gates, the evaluation results from the held-out tests, a runbook with training run separately for accountants and support staff, a retained AI Workpaper for each material run, and reviews at 30, 60 and 90 days to check the workflow is being used, not just installed.
Fit qualified before any engagement: systems, client base, obligations, balance-date load.
The pack configured to the firm, first governed workflow into production. Fixed fee, quoted after the scoping call.
Controls, approvals, audit trails.
Monitor, improve, expand.
Everything is deployed to your own Claude instance, under your own agreement with Anthropic.
The argument for partners is about margins, not technology. Compliance work is priced against a market that keeps compressing, so the hours a job absorbs decide its realisation. When the workpaper pack arrives assembled and tied, the reviewing accountant starts at review instead of assembly, and the same partner can stand behind more jobs without reading any less carefully. We won't invent a percentage for a practice we haven't met; the implementation puts numbers against your own jobs instead. The quieter effect is retention. The graduates a firm wants to keep did not join to key numbers into workpapers.
Download the partner briefing (PDF) →
One page to circulate before the scoping call: what this is, where client data goes, the practice-review answer, and what the implementation asks of the firm.
A short call to qualify the fit: your systems, your client base, your balance-date load. Then six weeks to year-end workpapers your partners can stand behind, at a fixed fee quoted after the call.
New Zealand accounting firms use Claude across the bounded, reconciled work that fills compliance season: year-end workpaper assembly tied to the trial balance, provisional tax schedules and client letters, FBT workings, payday-filing reconciliations, IRD correspondence triage, client questionnaires and chasers, AML/CFT onboarding support, and the GST cases the ledger software doesn't finish: zero-rated land, registration boundaries and adjustments. Airclerk designs and implements these as governed workflows, connected to the firm's ledger and practice systems, whether that is Xero, Xero Practice Manager, FYI, Karbon, MYOB or another stack, with engagement isolation between clients and a reviewer's sign-off on every output that leaves the building. For Xero we built the connection ourselves, because the ones that exist assume one business rather than a practice carrying hundreds of client organisations.
Airclerk's NZ Accounting Pack is the New Zealand answer to that question: 25 skills for public practice, covering year-end workpapers, GST, provisional tax, FBT, payroll, IRD correspondence, reporting tiers, AML/CFT onboarding, management accounts with cash-flow forecasting, imputation credit accounts, offshore shares and funds under the FIF rules, livestock and dairy clients, construction contracts and NZ tax research, each source-linked to current IRD, legislation and XRB materials with verification dates. Anthropic's nearest release, Claude for Small Business, is written for a US business owner rather than a practice. The pack is built to the shape of public practice instead: engagement isolation between clients as a hard rule, every output naming its client and period, rates and dates verified against the source when a skill runs rather than stated from memory, and a named reviewer gate before anything leaves the building. Eighteen skills carry a deliverable end to end. Five are tax research skills that route a question by depth, cite only sources the run actually opened, run a contrary-authority pass against their own conclusion and hand any arithmetic to the skill that owns that deliverable, where a schedule has to tie to the ledger or state its difference. One runs the practice's monitoring of IRD, tax policy and XRB sources, and one interviews the firm on the way in and writes the practice profile every other skill reads. The pack is an accelerator rather than a product: it is where an implementation starts, and each firm's version diverges from there as skills are adapted to the firm's own workpapers, templates and risk settings.
No. Every output is a draft for the firm's own accountants to review, with sources cited, rates and dates carrying verification tags, and uncertainty flagged. The reviewing accountant takes professional responsibility for anything that leaves the building. The workflow is supervised like delegated work, with added AI-specific controls: schedules must tie to the ledger or state their reconciling difference, proposed tax positions carry their statutory basis, and consequential actions stop for human sign-off. Airclerk is an AI implementation consultancy, not a CA firm and not a tax agent: nothing we build files with Inland Revenue, operates myIR or sends anything to a client except by the firm's own people. Where the firm's practitioners are Chartered Accountants, workflows are designed to support their obligations under the NZICA Rules and Code of Ethics: professional competence and due care, which the Code's technology revisions (effective December 2024) extend to AI-assisted output, confidentiality, terms of engagement, client monies where relevant, and the profession's ethics and tax-compliance guidance for tax practice.
Two protections come before anything else. Engagement isolation: one client's records are walled off from another client's work. Permission-aware access: the workflow sees only what the person running it is allowed to see in the firm's existing systems. Before any client records are connected, we check the vendor terms: whether inputs are used for training, what is retained and for how long, who the subprocessors are, and where data goes offshore (IPP 12). Everything runs on the firm's own Claude instance, under the firm's own agreement with Anthropic; client records reach it through connections the firm approves and scopes, and nothing routes through Airclerk. Under the Privacy Act's section 11 agent rule, a provider processing solely on the firm's behalf under no-training terms holds the information as the firm's agent rather than receiving a disclosure, and security responsibility under IPP 5 stays with the firm. Some steps never move into the workflow at all: filing returns or employment information, taking final tax positions, moving client money, AML/CFT compliance-officer decisions and suspicious activity reports, and signing anything stay with the firm's own people. And where people at the firm are already pasting client figures into personal AI accounts, we treat that as the first thing to fix, not a fact of life: an engagement starts by finding where AI is already in use and moving that work onto governed rails.
The AI arriving inside Xero, Karbon and MYOB is genuinely useful, and it belongs to their platforms: it works where their product works, on their roadmap, in their shape. What we implement is different in kind. Claude runs on the firm's own instance and works across systems: the ledger, the practice manager, the document store and the email trail in one workflow, encoded with the firm's own review standards and templates. The ledger connection is our own build, and it treats the client list as the normal case rather than the exception. The firm controls its configured workflows and prompts, and keeps the record of what the AI did. For most practices the two are complements, and the honest comparison is platform features against a build the firm keeps.
Yes, through a Xero connection Airclerk built for public practice. The connections that already exist assume one business with one ledger; a practice has hundreds, and the failure mode is an invoice or payment landing in the wrong client's file. Ours selects a client by name, alias or enough of the name to be unambiguous, and refuses with the candidates listed rather than guessing when two clients match. Reads are open; every write has to name the organisation it is landing in. Practices past Xero's 25-connection cap for an uncertified app run a second app as a second credential profile, so the client list is not the limit. It runs on the firm's own machine, so Xero credentials and refresh tokens stay there rather than in a hosted middleman. Coverage is the whole published surface for three Xero APIs: 235 accounting operations across 31 route families, 6 fixed-asset operations and 71 for New Zealand payroll, 312 in total, generated from a checksum-pinned copy of Xero's own specification. That includes manual journals end to end, twelve operations covering the journal itself, the supporting attachment and its history record, so a year-end adjusting entry can be drafted from the workpaper that computed it and posted once a reviewer approves. Payroll AU, Payroll UK, Projects and Files are deliberately out of scope. Payroll scopes are opt-in and can be granted read-only, payroll request and response bodies are never written to logs, and creating a posted pay run needs a second explicit confirmation. Every create and update carries a stable idempotency key, so a retry cannot leave a client with two of the same invoice. It is Airclerk's own tool, not a Xero product and not affiliated with or endorsed by Xero, and we install and configure it as part of an implementation.
They do different jobs, and Xero's is very good at the one it was built for. Its connector in Claude's directory is read-only by design, which Xero states plainly in its own listing, and its seven tools all return summaries: financial position, cash position, profit and loss, top customers by revenue, contacts and receivables, and the organisation's details and financial year. For an owner asking what their profit looks like this quarter, that is the right tool. Public practice needs two things it does not set out to offer. The first is depth: a year-end workpaper pack is built from a trial balance, aged receivables and payables and transaction-level listings of invoices, bank transactions and payments, and a summary figure will not reconcile to anything. The second is the ability to act, because a practice does not only ask questions about a ledger, it raises the invoice, records the payment, fixes the contact and posts the year-end journals. Ours covers 312 operations, every one Xero publishes for its Accounting, Assets and Payroll NZ APIs, so a workpaper pack can pull a trial balance, the fixed-asset register behind the depreciation schedule and the payday-filing detail behind the payroll reconciliation. Writes name the organisation they land in, every time. It also picks a client by name out of the whole client list, which is a problem a tool built for one business has no reason to solve.
Yes, and the Xero connection is often the faster win in-house. The pack is written for public practice, which is why so much of this page talks about engagement isolation and reviewer gates, but most of the work underneath is the same: GST returns, FBT, payday filing, provisional tax, year-end preparation for your auditor or accountant, monthly management accounts with a rolling cash-flow forecast, and keeping up with IRD and XRB changes. The skills that assume a client list do not apply, AML/CFT client onboarding and client chasers among them, and engagement isolation becomes entity isolation instead. The connection needs no adaptation at all. A finance team gets the same 312 operations, which means month-end accruals, prepayments, depreciation and reclassifications posted as manual journals from the workpaper that computed them, the fixed-asset register behind the depreciation, payday-filing detail, aged receivables and payables, and the standard reports. If the group runs several entities in separate Xero organisations, the multi-organisation design does the same job it does in a practice: a consolidation journal landing in the wrong subsidiary is the same error as one landing in the wrong client, and the same guard stops it. What does not change in either setting is the review boundary. Outputs are drafts, and whoever signs the return or the accounts still signs them.
Because the pack is the easy part. What is left is the work that decides whether it sticks: drawing out the firm's own review standards and encoding them, connecting the ledger and practice systems with the right scopes, designing the reviewer gates, testing the workflow on held-out jobs before an accountant relies on it, and training people so adoption is deliberate rather than informal. A firm with engineering capacity can do this itself. Most firms would rather buy the weeks back, especially the weeks either side of 31 March.
A practice review samples engagement files and tests the firm's system of quality management: PS 1 for compliance and advisory engagements, PES 3 where the firm does assurance work, and the NZICA Code of Ethics over all of it. An AI-assisted file has to show a reviewer the same things any file shows. What the work relied on, what was checked, who reviewed it, who signed it off. That is the discipline the pack is built around: outputs are drafts for a reviewing accountant, rates and dates carry verification tags, and every schedule ties to the ledger or states its difference, so the file a reviewer samples reads the way a well-run file reads. The boundary is worth stating plainly. Airclerk does not make a file compliant; the firm's quality management system does that. Our work is making sure AI-assisted output leaves the evidence that system needs.