AI AutomationPricing

How Much Does AI Automation Cost for a Small Business in New Zealand?

Kage Works1 September 20267 min read

Search "AI automation cost" and you'll land on a dozen pages quoting ranges like US$2,000 to US$150,000. None name a survey, a methodology, or a publication date behind the number. Most are marketing pages for AI consultancies, and the range is wide enough to be true of almost anything, which makes it useless for deciding whether your project is worth doing.

Two numbers actually matter, and they come from different places. The build is a quote from whoever does the work, worth interrogating like any quote. The monthly running cost is arithmetic, because the companies pricing the models publish their rates. This article works through both, with the actual numbers behind each.

Why most of what you'll read about this doesn't hold up

New Zealand has decent data on AI adoption and close to none on AI automation cost specifically. The AI Forum's biannual "AI in Action" survey found 82% of NZ organisations now use AI in some form, up sharply from its previous survey six months earlier. 56% of respondents report a positive financial impact from that use, and 71% cite operational cost savings as a benefit (AI Forum New Zealand). A Deloitte Access Economics study commissioned by 2degrees put a number on that impact for SMEs specifically: businesses using AI earned around NZ$400,000 more in FY25 than comparable non-adopters, based on survey data collected in early 2026 (itbrief.co.nz reporting on the 2degrees/Deloitte study).

Both of those are real, sourced, and about adoption and outcomes. Neither tells you what a workflow automation project costs to build, because most of that 82% is a person pasting text into ChatGPT, not a system running unattended against a live queue. The AI Forum's own breakdown shows why: 72% of organisations use off-the-shelf tools like ChatGPT or Copilot, and only 13% have built anything custom. The custom 13% is the group whose cost structure looks like what the rest of this article covers, and it's small enough that nobody has published a proper NZ cost survey of it yet. If you see a specific dollar figure for "AI automation cost" with no source attached, that's most likely a made-up number.

Our breakdown of which tasks are worth automating covers the adoption side in more detail. This one is about the two costs you can actually plan a budget around.

Running cost, per item

API pricing is published and public, so this is the part of the estimate with real numbers behind it. Anthropic and OpenAI both list current per-token rates for their models, and the cost of running a classification or summarisation job is direct arithmetic from there.

Take a support ticket classification job, the kind of task described on our AI automation service page: a short model reads an incoming ticket and assigns a category. At Anthropic's published rate for Claude Haiku 4.5 (US$1 per million input tokens, US$5 per million output tokens as of September 2026), a 300-token ticket producing a 20-token category label costs roughly US$0.0004, or about US$0.40 per thousand tickets (Anthropic API pricing). At current exchange rates that's under NZ$0.70 per thousand (xe.com mid-market rate, late August 2026).

Summarisation costs more, because it reads more and writes more. A ticket plus its reply history at 1,500 tokens, summarised into 150 tokens using the larger Claude Sonnet 5 model (US$2 per million input tokens, US$10 per million output), costs roughly US$0.0045 per ticket, about ten times the classification cost. Roll 500 of those weekly summaries into a single digest and the digest itself, at maybe 75,000 tokens of input and 800 tokens of output, costs about US$0.16 a week to generate. OpenAI's equivalent tier, GPT-5-mini, prices at US$0.25 input and US$2.00 output per million tokens, in the same order of magnitude (OpenAI API pricing).

These are illustrative, not a quote. Token counts depend on how verbose your tickets are and how the prompt is written, and rates move as vendors release new models, so check current pricing before you build a budget around them. But the pattern holds regardless of the exact numbers: a well-scoped classification job runs for cents per thousand items, a summarisation job runs for several times that, and both are cheap enough that running cost is rarely what kills one of these projects. The next section covers what actually does, and it isn't the running cost.

The cost of the build itself

This is where the wide, unsourced ranges you'll find elsewhere come from, and where "it depends" is an honest answer rather than a dodge. A prompt-engineered classifier against a well-defined queue is days of work. A system that needs a fine-tuned model, a human-review path for edge cases, and integration into an existing ticketing platform is weeks. Nobody can quote either without seeing your data and your current process first, and a vendor who quotes a fixed price before that conversation is guessing.

The honest version of a build-cost estimate looks like a scoping phase, not a number pulled from a website. Ours runs as a paid discovery week: we look at your process as it runs today, and come back with a written assessment of what a machine could take over, an expected accuracy range, a fixed quote for the build, and the expected per-item running cost, before you commit to anything. That structure exists because guessing a number upfront and adjusting the scope to fit it is how AI projects end up half-built, which is the risk the next section covers.

Scoping mistakes cost more than the invoice

RAND Corporation's 2024 study of AI implementation found that more than 80% of AI projects fail, roughly twice the failure rate of non-AI IT projects (RAND Corporation, "Why AI Projects Fail and How They Can Succeed"). That statistic gets quoted constantly and rarely explained. The usual cause is scope, not the model: a project built against a demo instead of real data and a spec nobody checked with the people doing the job today, then shipped with nobody watching whether accuracy holds up as the input data drifts.

Budget overruns follow the same pattern at the enterprise end: McKinsey's 2026 global AI survey, fielded between May and June 2026 across 1,719 respondents, found 93% of organisations report exceeding their AI budgets, with one in five constraining AI use specifically because of cost (McKinsey & Company, "Burning through the AI budget"). That's enterprise-scale spending, and the dollar figures don't transfer to a project costing a few thousand dollars. The cause does: a project scoped before anyone checks it against real data tends to cost more than expected, because the gap between the demo and the live queue is exactly where the extra weeks go.

The fix is unglamorous and cheap relative to the alternative: see the model's output against cases you already know the answer to before you pay for production. A vendor whose plan skips straight from pitch to build with no prototype step against your actual data is asking you to accept the failure odds above rather than improve on them.

Pricing your own project before you call anyone

Four numbers get you most of the way to a sane budget conversation with any vendor:

Volume. How many items a week actually go through the process you want automated. Count it for two weeks rather than guessing; most people are wrong in one direction or the other.

Error cost. What a wrong answer costs if the system gets one wrong. A misrouted enquiry costs a few minutes to fix. A wrong price quoted to a customer costs the job. This decides how much human review the system needs, which is most of what drives build time.

Task shape. Whether the job is picking one of a known set of categories (cheap to run, faster to build) or generating open-ended text (costs more per item, harder to verify automatically).

Running budget. Multiply your weekly volume by the per-item costs above, add a margin for token counts running higher than expected, and you have a monthly running number to hold any vendor's estimate against.

A real discovery conversation still matters more than any of this, but walking in with these four numbers means you can tell a properly scoped quote from a guess.