Short answer
AI automation in New Zealand costs a one-off amount to design, build, connect and test the workflow, plus running costs: platform fees, model fees per run, integration fees, monitoring, staff review time and changes. For typical business automations, model fees are often the smallest line, a few NZ dollars per thousand runs on low-cost models, while staff review time and build effort dominate.
- Model fees at 2,000 input and 300 output tokens range from NZ$0.62 to NZ$61.80 per 1,000 runs.
- Staff review time is usually the largest running cost, not the AI model.
- Platforms bill differently: n8n per execution, Make per module action, Zapier per task.
- Vendors bill in USD, EUR or AUD, so exchange rates move your costs.
The short version
An AI automation has a one-off cost (designing, building, connecting and testing the workflow, mostly engineering time) and a running cost: platform, model fees per run, integration fees, hosting, monitoring, staff time checking outputs and handling exceptions, and changes when something upstream moves. For most business automations model fees are the smallest line; review time and build effort are the largest.
We do not publish our rates; this page gives the structure and public prices so you can test any quote, ours included.
Every cost line, and what drives it
| Cost line | One-off or running | What drives it | How to estimate |
|---|---|---|---|
| Discovery | One-off | Number of steps, exceptions and people involved | Days of workflow mapping |
| Build | One-off | Steps, branches, approvals, error handling | W engineer-weeks times the weekly rate you are quoted |
| Integration | One-off | Number of systems, API quality, permissions | Part of W; more for systems without good APIs |
| Testing and evaluation | One-off, then per change | Number of input types and the accuracy bar | Labelled test cases, plus engineer time |
| Automation platform | Running | Billing unit and volume | Published plan prices (below) |
| Model fees | Running | Runs, tokens per run, model tier | Formula below |
| Integration fees | Running | Some vendors charge for API access | Published developer plans |
| Hosting | Running | Self-hosted platform or custom code | Your cloud provider’s quote (H) |
| Monitoring | Running | Weekly sample checks and alerts | Hours a month |
| Staff review | Running | Share of outputs checked, minutes per check | Volume times minutes times your hourly cost |
| Change requests | Running | Upstream changes, new cases, new teams | Hours a quarter |
Platform fees, as listed on 19 October 2026
| Platform | Billing unit | Entry paid plan | NZD at 16 October 2026 rates |
|---|---|---|---|
| n8n Cloud | Execution: one full workflow run, any number of steps | Starter 20€ a month billed annually, 2,500 executions | about NZ$39.74 |
| n8n self-hosted | Free Community Edition for internal use | Hosting only (H) | H |
| Make | Credit: each module action, per bundle of data | Core $9 a month billed annually, 10,000 credits (currency not stated) | about NZ$15.89 if US dollars |
| Zapier | Task: each successful action; AI steps 1, 3 or 5 tasks | Professional US$19.99 a month billed yearly, 750 tasks | about NZ$35.29 |
| Power Automate | Per user or per bot licence | Premium NZ$24.30 per user a month, yearly, excluding tax | NZ$24.30 |
Sources: n8n, Make, Zapier and its AI step rates, Microsoft. Billing units matter more than headline prices: Make counts every module once per bundle, so 10 items through 3 modules uses 31 operations (Make). Microsoft 365’s included automation rights cover standard connectors only, not premium or custom ones (Microsoft Learn). Our comparisons of n8n and Make and Zapier and n8n go further.
Model fees per run: the arithmetic
Cost per run = (input tokens × input price + output tokens × output price) ÷ 1,000,000. Prices are US dollars per million tokens; we convert at the European Central Bank reference rate of NZ$1.7656 per US dollar on 16 October 2026 (Frankfurter).
For a typical automation step reading 2,000 tokens (a long email plus instructions) and writing 300 (a classification and short draft), standard-tier list prices on 19 October 2026 give:
| Model | US$ per million (in / out) | Per 1,000 runs (US$) | Per 1,000 runs (NZ$) |
|---|---|---|---|
| GPT-6 Luna or Claude Haiku 5.5 | 0.10 / 0.50 | 0.35 | 0.62 |
| Gemini 3.5 Flash-Lite | 0.30 / 2.50 | 1.35 | 2.38 |
| Gemini 3.8 Flash (introductory to 31 December 2026) | 0.75 / 3.75 | 2.63 | 4.63 |
| GPT-6.1 Sol, Claude Sonnet 5.5 or Gemini 4 Argon (introductory to 30 November 2026) | 2.00 / 10.00 | 7.00 | 12.36 |
| Claude Opus 5.5 | 4.00 / 20.00 | 14.00 | 24.72 |
| GPT-6 Astra or Claude Fable 5.1 | 10.00 / 50.00 | 35.00 | 61.80 |
Prices from OpenAI, Anthropic and Google. Batch processing, for work that can wait, halves these figures at all three providers.
Three worked scenarios
All three are illustrative; every figure is a stated assumption. W (build engineer-weeks), H (hosting) and C (change hours) stay as variables because they depend on your systems and supplier.
Scenario A: supplier invoices from a shared inbox into Xero. A 15-person trades business, 400 invoices a month.
| Line | Assumption | Monthly NZ$ |
|---|---|---|
| Platform | n8n Cloud Starter, 400 executions | 39.74 |
| Model fees | GPT-6.1 Sol, 3,000 input and 400 output tokens an invoice: US$0.01 each | 7.06 |
| Xero API | Developer Starter plan, free for up to 5 connections | 0.00 |
| Staff check | 30 seconds an invoice at NZ$60 an hour | 200.00 |
| Exceptions | 10% handled by hand, 5 minutes each, at NZ$60 an hour | 200.00 |
| Known running total | 446.80 | |
| Manual today | 4 minutes an invoice at NZ$60 an hour | 1,600.00 |
| One-off | Discovery plus W engineer-weeks; testing on 200 past invoices | W |
Net monthly benefit before C is about NZ$1,153. Using Xero’s remote MCP server instead would need its Core developer plan, A$35 a month (about NZ$43.00), for a real organisation (Xero).
Scenario B: customer email triage and drafting in Microsoft 365. A 40-person distributor, 3,000 emails a month.
| Line | Assumption | Monthly NZ$ |
|---|---|---|
| Platform | Power Automate Premium, one service user | 24.30 |
| Classification | GPT-6 Luna on every email, 1,500 input and 50 output tokens: US$0.53 a month | 0.93 |
| Drafting | Claude Sonnet 5.5 on 1,800 emails, 4,000 input and 400 output tokens: US$21.60 a month | 38.14 |
| Monitoring | 1 hour a week of sample checks at NZ$55 an hour | 238.33 |
| Time saved | 3 minutes net per drafted email, after review, at NZ$55 an hour | 4,950.00 benefit |
| One-off | Discovery plus W engineer-weeks; a test set of 300 labelled emails | W |
Scenario C: claims document classification and extraction. A larger organisation, 20,000 documents a month, processed overnight.
| Line | Assumption | Monthly NZ$ |
|---|---|---|
| Model fees | GPT-6.1 Sol via batch, 5,000 input and 600 output tokens a document: US$0.008 each | 282.50 |
| Same at standard tier | For comparison, if results are needed immediately | 564.99 |
| Hosting | Custom pipeline on your cloud | H |
| Staff review | 8% of documents routed to people, 4 minutes each, at NZ$65 an hour | 6,933.33 |
| Monitoring | 4 hours a week at NZ$65 an hour | 1,126.67 |
| One-off | Discovery plus W engineer-weeks, with integration to the claims system and a large labelled test set | W |
In Scenario C, review time costs about 24 times the model fees. Cutting the review rate from 8% to 6% saves more than switching to a model at half the price.
New Zealand factors that change the numbers
- Currency. Model APIs bill in US dollars, n8n in euros and Xero’s developer plans in Australian dollars. We used 16 October 2026 rates: NZ$1.7656, NZ$1.9868 and NZ$1.2286 respectively.
- GST. The API price lists are before tax. Claude’s pricing page, viewed from New Zealand, says “Prices include 15% GST” while showing the same figures as Anthropic’s US-dollar docs (Anthropic). Confirm on a real invoice.
- Data location. OpenAI charges 10% more for data-residency endpoints on models released from 5 March 2026, and its Australian region stores data without processing it there. n8n Cloud stores data in Frankfurt; self-hosting is the route to local storage.
How we estimate an automation
- Discover: measure the workflow. Volumes, exception rates and the systems involved, from real data.
- Prove: price the narrow build. The smallest automation that moves the operating number, quoted as a fixed piece of work in engineer-weeks.
- Scale: itemise running costs. Each running line, with volume and token assumptions to check monthly.
- One-off effort
- Engineer-weeks for discovery, build, integration and testing.
- Running lines
- Platform, model, integration, hosting, monitoring, review and changes.
- Assumptions
- Volumes, tokens per run, review rates and exchange rates, with dates.
Estimates fail when they count model fees and ignore people, measure exception rates on clean samples, misread billing units, or forget upstream change: Xero moved developers to tiered, usage-based plans on 2 March 2026, and providers retire models on fixed dates. Our AI cost calculator uses the same lines. For larger builds see what custom AI projects cost; for delivery, AI workflow automation.
Work this draws on
- Property management agent for APL Property: an internal agent that handles routine administration, including email triage, records, orders and follow-up, for a regional NZ property firm, with staff keeping the decisions. Agents of this kind carry the same cost lines as Scenario B.
Sources
- API pricing, OpenAI, accessed 2026-10-19
- Pricing (Claude API), Anthropic, accessed 2026-10-19
- Gemini Developer API pricing, Google, updated 2026-10-16
- n8n pricing, n8n GmbH, accessed 2026-10-19
- Make pricing, Make (Celonis), accessed 2026-10-19
- Operations, Make Help Centre, updated 2026-10-16
- Zapier pricing, Zapier, accessed 2026-10-19
- Zapier pricing rates, Zapier, accessed 2026-10-19
- Power Automate pricing (New Zealand), Microsoft, accessed 2026-10-19
- Power Automate licence deep dive, Microsoft Learn, updated 2026-09-22
- Xero developer pricing, Xero, accessed 2026-10-19
- Xero remote MCP FAQ, Xero, accessed 2026-10-19
- Claude pricing (viewed from New Zealand), Anthropic, accessed 2026-10-19
- Reference exchange rates, 16 October 2026 (European Central Bank data), Frankfurter, 2026-10-16
Written by the Levy Street team in Wellington. Last updated 19 October 2026.