AI Agent Development Cost in New Zealand: 2026 Pricing Guide

New Zealand's AI-agent market now stretches from a NZ$70-a-month voice receptionist to a six-figure custom automation system. Both may be sold as an “AI agent”, but they are not the same product. One configures an existing platform around a standard task; the other has to reason over your data, connect to business systems, take controlled actions and survive production.

This guide separates those categories before comparing prices. It uses public New Zealand pricing reviewed in July 2026, then shows where 99 Francs fits: custom production agents priced in NZD at 30% below the comparable public local custom-development range, with no deposit and payment only after each agreed phase works.

The 2026 New Zealand AI-agent price map

Public prices show why a single “average AI agent cost” is misleading. AgentWorks advertises entry-level deployment from NZ$599 plus GST and publishes a broader guide with setup from NZ$2,500 and monthly costs around NZ$150–800. Automate AI offers productised automation from NZ$70 per month, while its public terms put custom development from NZ$2,500. HighPeak Digital publishes NZ$2,500–25,000 for AI agents and reception, NZ$1,000–180,000 for the much broader category of AI systems and automation, and NZ$2,000–15,000 per month for ongoing partnerships.

These are advertised ranges, not quotes for identical briefs. They are useful as market markers, but a NZ$129/month receptionist should not be compared with a system allowed to update customer records, approve a refund or move money. The cost follows the consequences of being wrong.

What 99 Francs charges in New Zealand

Our pricing is for custom work, not a white-labelled agent subscription. Against the public NZ$2,500–25,000 local range for custom voice and workflow agents, our comparable range is NZ$1,750–17,500 — 30% lower. That difference comes from the delivery model, not from removing the work: we are a remote founder-led studio with no New Zealand office, local sales department or default layer of account managers and project managers. You work directly with the senior designers and engineers doing the job, so the budget pays for expertise, testing and the accepted result rather than organisational overhead.

All figures exclude GST where it applies, third-party model/API usage and unusual infrastructure costs. The range becomes a fixed NZD price once we agree the phase, acceptance criteria and dependencies. If a ready-made tool can solve the job for less, we will say so before proposing a custom build.

Why “custom” changes the price

• Actions and consequences. Answering a question is cheaper to validate than changing a booking, issuing a quote, updating a CRM or approving a transaction.
• Integrations. Xero, MYOB, a CRM, an internal database, email, telephony and custom APIs each introduce permissions, failure states and testing work.
• Data and retrieval. Clean source material is inexpensive to connect; fragmented documents, permissions and unreliable records require more engineering.
• Reliability. Provider fallbacks, evaluations, logs, retries, human review and rollback paths cost more than a prompt that works in a demo.
• Interface. Customer-facing and internal agents often need authentication, dashboards, approvals, history and exception handling — effectively a custom web product around the model.
• Risk. The closer an agent gets to money, personal information or an irreversible action, the more sandboxing, limits and evidence it needs before launch.

The payment model matters as much as the price

A lower quote does not remove much risk if half is due before anyone knows whether the agent works. Public NZ terms commonly use advance subscriptions, setup fees or deposits. Automate AI, for example, states that setup fees are paid before activation and custom projects use deposit, progress and completion invoices. Some productised providers let a customer test a standard voice agent before paying, which is useful, but that is different from financing a bespoke development phase.

99 Francs uses an outcome-based model for custom work. We agree what the phase must do and how it will be accepted, build it at our risk, demonstrate it on the agreed cases, revise where needed, and invoice only after you approve the working result. There is no deposit. If it still does not meet the agreed outcome after revisions, you owe nothing for that phase.

Experience where an agent can affect money and risk

A support bot can make an awkward mistake. An automated financial system can create a loss. Our experience includes both ordinary agent surfaces and the harder end of that spectrum: an autonomous prediction-market trading system where software makes and cancels orders around real money, exposure limits and live market risk.

For that system, a strategy does not go directly from backtest to capital. It moves through inert, virtual, mirror and deliberately small real-money stages. Live execution is gated by fresh price data, exposure caps, daily-loss limits, circuit breakers and the continuing profitability of a paper twin. We also moved its latency-critical execution path from Python to Rust without rewriting the strategy layer. The Polymarket trading system case study and the accompanying engineering series show the work in detail.

Beyond financial automation, we have built or prototyped long-memory assistants, Telegram agents, voice interfaces, native mobile AI products and LLM workflows. The common discipline is the same: define what the system may do, test the long tail, make uncertainty visible and keep a human gate where the consequence demands one.

What a New Zealand production build must account for

The technical stack can be remote; the operating context cannot. New Zealand's Privacy Act 2020 applies when an agent collects, uses or shares personal information, and IPP3A added an indirect-collection notification obligation from 1 May 2026. The Privacy Commissioner also recommends a privacy impact assessment, transparency about how AI is used and human review before acting on AI output.

• Price, scope and invoices expressed in NZD, with GST treatment made explicit.
• Human approval before high-impact customer, financial or operational actions.
• Audit logs showing what the agent saw, decided and attempted.
• Data-retention and provider settings checked before personal or confidential information is sent to a model.
• Failure paths for unavailable providers, broken integrations and low-confidence outputs.
• Working-hour overlap with the New Zealand morning for reviews, launches and incident response.

When the agent becomes a web-development project

A production agent rarely lives inside a chat box alone. Once it needs sign-in, roles, an approval queue, customer-facing forms, activity history, reporting or a CMS-managed knowledge base, part of the job is custom web development. That is why the same team should think about the agent logic and the product surface around it. If the agent is part of a wider website, portal or internal tool, see our web development for New Zealand businesses service.

Should you buy, configure or build?

The best-value decision is not always to hire us. AI Forum reporting says most surveyed firms preferred off-the-shelf products to custom systems, and its later reporting found ready-to-use tools had pushed setup costs down. Custom development earns its cost only where the business needs different behaviour, deeper integration or stronger control than a standard product can offer.

Why the timing is good for New Zealand businesses

New Zealand's national AI strategy explicitly prioritises practical adoption over foundational model development. The AI Forum's August 2025 findings say 91% of surveyed adopters saw efficiency gains, while 77% reduced operating costs. MBIE has also expanded the AI Advisory Pilot to 150 businesses, with eligible businesses able to access 50% co-funding up to NZ$15,000 through 31 January 2027. Our guide to AI agent development and the New Zealand AI Advisory Pilot explains where that programme may fit.

A lower-risk way to get a real number

Send us the workflow, the systems it touches and the action you want the agent to complete. We will tell you whether an existing product is enough, whether a short feasibility phase is needed, or whether the scope is ready to price. For custom work, you receive a fixed first-phase price in NZD, acceptance criteria and a timeline before we start — but no invoice until the phase works and you accept it.