Short answer
A custom AI project in New Zealand costs a one-off amount for discovery, build, integration, evaluation and security review, measured in engineer-weeks, plus running costs: model and API fees, hosting, monitoring, staff review time, support and change requests. Scope drives the one-off cost; volume and review rates drive the running cost. Model fees are rarely the largest line, while integration and human review usually are.
- Estimate build effort in engineer-weeks per phase, then apply your supplier’s rate.
- Each extra system, user group or autonomous action adds build and testing effort.
- Model fees per 1,000 eight-page documents range from about NZ$2 to NZ$230.
- Prompt caching cut a worked agent example’s model cost by 55%.
The short answer
A custom AI project has a one-off cost and a running cost. The one-off cost is mostly engineering time: discovery, build, integration with your systems, evaluation and security review. The running cost is model and API fees, hosting, monitoring, the time your staff spend reviewing outputs, support and change requests. Scope sets the first; volume and review rates set the second. We do not publish our rates, so this page expresses effort in engineer-weeks and prices only what has a public price.
The cost model
First-year cost = D + (Wb + Wi + We) × r + S + 12 × (M + H + O + R) + C
| Symbol | Line | What drives it |
|---|---|---|
| D | Discovery | Workflows mapped, people interviewed, data access |
| Wb | Build, in engineer-weeks | Steps, branches, user interface, error handling |
| Wi | Integration, in engineer-weeks | Number of systems, API quality, permissions |
| We | Evaluation, in engineer-weeks | Output types, accuracy bar, test-set size |
| r | Weekly rate | Your supplier’s quote |
| S | Security and privacy review | Data sensitivity, autonomy, customer exposure |
| M | Model and API fees a month | Volume, tokens per task, model tier |
| H | Hosting a month | Infrastructure, region, uptime needs |
| O | Monitoring and support a month | Hours of cover, release frequency |
| R | Staff review a month | Share of outputs checked, minutes each, internal cost |
| C | Change requests | Upstream changes, model migrations, new cases |
Year two drops D, Wb, Wi and most of We and S, but keeps M, H, O, R and C, plus at least one model migration as providers retire versions.
How scope moves the effort
| Scope dial | Narrow end | Broad end |
|---|---|---|
| Workflows | One step of one workflow | Several workflows across teams |
| Systems | Reads one source | Reads and writes to several systems |
| Users | One team, same permissions | Many groups with different access |
| Autonomy | Drafts for a person to send | Acts without approval |
| Exposure | Internal staff | Customers or the public |
| Inputs | One document type, one language | Many formats, channels and languages |
Each move toward the broad end adds build, integration and, above all, testing effort. The cheapest way to control cost is to start at the narrow end and expand with evidence, which is how an AI proof of concept works.
Model and API costs, worked
Formula: (input tokens × input price + output tokens × output price) ÷ 1,000,000, using US-dollar list prices on 19 October 2026 and NZ$1.7656 per US dollar, the European Central Bank reference rate on 16 October 2026 (Frankfurter).
Per 1,000 documents of about eight pages (8,000 input tokens, 1,000 output tokens):
| Model (standard tier) | Per 1,000 documents (US$) | Per 1,000 documents (NZ$) |
|---|---|---|
| Claude Haiku 5.5 or GPT-6 Luna | 1.30 | 2.30 |
| Gemini 3.8 Flash (introductory price to 31 December 2026) | 9.75 | 17.21 |
| Claude Sonnet 5.5 or GPT-6.1 Sol | 26.00 | 45.91 |
| Claude Opus 5.5 | 52.00 | 91.81 |
| Claude Fable 5.1 or GPT-6 Astra | 130.00 | 229.53 |
Prices from Anthropic, OpenAI and Google; batch processing halves them.
Agents cost more per task because they call the model repeatedly. An agent task of 10 calls on Claude Sonnet 5.5, each with 6,000 input tokens (5,000 of them a stable prefix of instructions and tool definitions) and 300 output tokens, costs US$0.150 uncached. With prompt caching (a US$2.50 per million write once, then US$0.10 per million reads) it costs US$0.067, 55% less. Research agents go further: Lucent’s whitepaper describes processing “1-5 million tokens per analysis” (Lucent); 3 million input tokens alone is US$12 on Opus 5.5.
Three worked projects
Illustrative only. Every figure is an assumption; one-off effort is left as engineer-weeks because it depends on your systems and supplier.
Project 1: internal knowledge assistant over SharePoint. 120 active staff asking 5 questions a working day, built as described under RAG development.
| Line | Assumption | NZ$ a month |
|---|---|---|
| Model fees | 12,600 questions, 4,000 input and 400 output tokens each, Claude Sonnet 5.5: US$151.20 | 266.96 |
| Indexing | 50,000 documents of 1,500 tokens with text-embedding-3-small: US$1.50 once | 2.65 once |
| Monitoring | 2 hours a week at NZ$70 an hour | 606.67 |
| Content fixes | 4 hours a month by document owners at NZ$70 | 280.00 |
| Hosting | Index and app in an Auckland cloud region | H |
| One-off | D, plus Wb, Wi for SharePoint permissions, We for a 150-question benchmark | engineer-weeks |
Project 2: contract clause review with citations. A law firm reviewing 120 contracts a month.
| Line | Assumption | NZ$ a month |
|---|---|---|
| Model fees | 30,000 input and 3,000 output tokens a contract, Claude Opus 5.5: US$0.18 each | 38.14 |
| Lawyer review | Every output checked, 20 minutes a contract at NZ$150 an hour internal cost | 6,000.00 |
| Manual today | 60 minutes a contract at NZ$150 | 18,000.00 |
| One-off | D, Wb, Wi for the document system, We with lawyers labelling clauses, S | engineer-weeks |
Review is not optional: the Law Society says a lawyer “is not absolved from responsibility” for work derived from generative AI (NZLS). Net benefit before H, O and C is about NZ$11,962 a month, and it depends entirely on the 20-minute review assumption.
Project 3: customer-facing booking agent with tools. 8,000 conversations a month.
| Line | Assumption | NZ$ a month |
|---|---|---|
| Model fees | 6 calls a conversation, each 6,000 input tokens (5,000 cached) and 300 output, Claude Sonnet 5.5: US$0.045 each | 635.62 |
| Escalations | 15% to staff, 6 minutes each, at NZ$45 an hour | 5,400.00 |
| Monitoring | 4 hours a week at NZ$70 an hour | 1,213.33 |
| One-off | D, Wb, Wi for booking and CRM systems, We, and S before launch | engineer-weeks |
How we scope and estimate
- Discover: fix the scope. A workflow assessment sets the scope dials, lists the systems and measures volumes, so the estimate rests on facts.
- Prove: price the narrow version. The first build is the narrowest system that can move the operating number, quoted as a fixed piece of work.
- Scale: itemise running costs. Before launch you receive monthly running costs by line, with token assumptions, so you can watch actuals against them.
- Effort by phase
- Engineer-weeks for discovery, build, integration, evaluation and review.
- Systems list
- Every system touched, with access method and known limits.
- Running-cost table
- Each line above, with volume and token assumptions.
- Currency basis
- Exchange rates and dates used.
- Change terms
- How new requirements and forced upstream changes are priced.
Checking a quote
Ask whether it states engineer-weeks per phase, names every system to be integrated, includes an evaluation set, states token and volume assumptions, counts staff review time, says who pays model fees and in which currency, and explains how model migrations are handled. New Zealand specifics belong too: integration fees such as Xero’s developer plans priced in Australian dollars, Xero’s per-organisation API limits, and roughly 10% more for regional processing endpoints. Our AI cost calculator and the companion page on AI automation pricing use the same lines.
Work this draws on
- Lucent: a Levy Street venture that runs several frontier models together for investment research; at its token volumes, model choice and caching are budget decisions.
- Property management agent for APL Property: an internal agent handling routine administration for a regional NZ property firm, with staff keeping decisions.
Sources
- Pricing (Claude API), Anthropic, accessed 2026-10-19
- API pricing, OpenAI, accessed 2026-10-19
- Gemini Developer API pricing, Google, updated 2026-10-16
- Reference exchange rates, 16 October 2026 (European Central Bank data), Frankfurter, 2026-10-16
- LUCENT: The AI Consortium for Institutional-Grade Equity Analysis (whitepaper), Lucent, 2025-08
- Xero API limits, Xero, accessed 2026-10-19
- Xero developer pricing, Xero, accessed 2026-10-19
- Generative AI guidance for lawyers, New Zealand Law Society, updated 2026-02-24
Written by the Levy Street team in Wellington. Last updated 19 October 2026.