AI workflow automation for New Zealand businesses
We take a repetitive task your team does by hand, work out whether a machine can do it reliably, and put the result into production. Most web studios do not build this. We have run classification models against live queues and shipped the tooling around them.
What you get
- A written assessment of which tasks in your process are worth automating
- A working prototype on your real data before you commit to the build
- The automation running in production against live inputs
- Accuracy measured against human decisions, with the number reported honestly
- A path for the cases the system should refuse to handle alone
- Monitoring, so you find out when quality drifts rather than a customer telling you
- Documentation your team can hand to another engineer
What we have actually built
A New Zealand support desk was receiving more tickets than the team could triage. We trained a classification model on thousands of their historical tickets, learning the categories, sub-categories and priorities their own staff had assigned, then put it in front of the live queue.
Every incoming ticket now gets classified and routed the moment it arrives. Each one also carries an LLM-written summary, so an agent has context before opening it. A second automation runs weekly, rolling the individual summaries into a digest that surfaces recurring problems as engineering action items. That job used to take hours of reading.
The work is under NDA, so we cannot name the client. The pattern transfers to any business with a queue of inbound text: enquiries, claims, applications, maintenance requests.
Where automation pays, and where it does not
Automation earns its cost on tasks that are high volume, rule-shaped, and tolerant of a small error rate with a human check behind it. Sorting inbound email. Pulling line items off supplier invoices. Drafting a first-pass reply. Flagging which of 400 records need a person to look at them.
It fails on tasks where every case is different, where the volume is low enough that a person handles it in ten minutes a week, or where being wrong once carries a cost you cannot accept. We turn down that third category rather than build something that will hurt you.
Our note on AI workflow automation for NZ small business works through which tasks in a typical operation fall on each side.
How a project runs
We start with a paid discovery week. You show us the process, we watch how it runs now, and we come back with a written view of what a machine could take over, what accuracy to expect, and what it would cost to run each month.
If the numbers work, we build a prototype against your real data before committing to production. You see the model's output on cases you already know the answer to, which is the only honest way to judge whether it is good enough.
Production comes last, with monitoring attached. Model quality drifts as your inputs change, and a system nobody is watching quietly gets worse for months.
Running costs and data handling
Ongoing cost depends on volume and the model. A classification job running on a small fine-tuned model costs cents per thousand items. An LLM summarising long documents costs more, and we give you the per-item number during discovery rather than after the first invoice.
On data, we work out what leaves your systems and tell you where it goes. Where the material is sensitive, we look at models that run in a region you are comfortable with or on infrastructure you control. New Zealand privacy obligations apply to your customers' information regardless of which vendor processes it, and that shapes the design rather than getting checked at the end.
Common questions
What does AI automation cost for a small business?
It depends on the task and the volume running through it. A paid discovery week comes first, and it ends with a written scope, a fixed quote for the build, and the expected per-item running cost, so you see the real numbers before committing to anything.
Do I need a lot of data to start?
For classification work, a few thousand historical examples with the decisions your team already made is enough. For tasks handled by a general model with good instructions, you need no training data at all. We work out which applies during discovery.
How accurate are these systems?
It depends on the task, and we measure it rather than estimate it. A well-scoped classification job usually matches human decisions on 85% to 95% of cases. We design the workflow so the remaining cases route to a person instead of guessing.
Will my customer data be sent to a third party?
That depends on the design, and it is a decision you make with the facts in front of you. We tell you which vendor processes what, where it is hosted, and what their retention terms say. Where the data is sensitive we look at models running on infrastructure you control.
Is this the same as adding a chatbot to my website?
No. A website chatbot answers visitors. This is automation of the work your team does internally, which is where the hours actually go. We build chatbots too, and our article on what they cost covers when one pays for itself.
Tell us what you are trying to build.
Send a couple of paragraphs about the business and the problem. You get a reply within one business day, from the person who would do the work.
Start a project