AI agents
Agents that hold a conversation, remember it and act on it. Built on the Claude API, with memory through Zep and Graphiti and guardrails written in code.
AI Development
Everyone is selling AI. Almost nobody has shipped it.
Plenty of agencies will bolt a chatbot onto your website and call it innovation. We are a New Zealand studio building a real AI product, Abbi, on the Claude API with long term memory and retrieval. Your system gets that standard: rules decide, AI delivers.
Rules decide. AI delivers.
An AI system is not a chat window with your logo on it. It is software: typed code that decides what happens, a model that handles language and judgement, retrieval that grounds answers in your real data, and memory that carries context between sessions. That is how we are building Abbi, and how we build for you. The model is brilliant at delivering. It should never be the thing deciding.
- A chatbot bolted in the corner
- Forgets every conversation
- Invents answers under pressure
- No tests, no evals, no logs
- Dies quietly after launch
The wrapper demos well. The system keeps working.
A lot of what is sold as AI right now is a wrapper: a prompt, a widget, a launch post. It does not survive the first month of real customers. A system is engineered to survive them.
The chatbot wrapper
- A prompt in a widget, bolted on
- No memory, no retrieval, no rules
- Guesses when it does not know
- Breaks quietly, nobody is watching
You, with Igloo
- Wired into your site and your data
- Memory, retrieval and rules in code
- Answers grounded in your own facts
- Evals and monitoring after launch
AI development, from your website to your back office.
Web facing AI first, because that is where your customers are. Business automation second, because that is where your hours go. Everything below runs on engineering we already use ourselves.
Retrieval & RAG
Assistants that answer from your documents, prices and policies, not the open internet. PostgreSQL and pgvector keep every response grounded in your facts.
AI features in websites
Smart search, instant quoting, lead triage and drafting tools inside a fast site. AI that helps visitors buy, without slowing the page down.
Workflow automation
Power Automate flows that route enquiries, send reminders and move data between systems on schedule, without a human in the loop. Boring, reliable and compounding.
Internal systems
Power Apps, SharePoint and Entra builds like SpruceOS, the system running our own cleaning company: jobs, checklists, documents and one login.
Evals & monitoring
Test suites, logging and drift checks, so you know the system still answers correctly in month six. The step every wrapper skips.
How is custom AI development different from just using ChatGPT?
ChatGPT is a chat window, and a genuinely useful one. If it covers what you need, use it and keep your money. But it does not know your prices, cannot see your bookings, forgets you between visits, and answers to the whole internet, not your business.
Custom AI development starts from the opposite end. We take one workflow, wire the model into your actual data through retrieval, add memory where the job needs it, and put deterministic rules around every action that matters. If the answer must be right, code decides it. That split, rules decide and AI delivers, is the whole difference between a demo and a system.
It is the same architecture we are building Abbi on. Everything else we have shipped is on our work page.
AI automation for NZ businesses.
Not every win is on the website. Most NZ businesses lose their hours in the back office: enquiries retyped into spreadsheets, reminders sent by hand, documents nobody can find. That is automation territory, and much of it does not need a model at all, just well built Power Automate flows and a tidy Power Apps front end on the Microsoft 365 you already pay for.
We know, because we run a company on it. SpruceOS and MySpruce are internal systems we built on SharePoint, Power Apps, Power Automate, Teams and Microsoft Entra, and they run Spruce, our own cleaning business, every day. When AI genuinely earns a seat in a flow, summarising an enquiry, drafting a reply, triaging a form, we add it there, with the same rules first discipline.
Can you build AI into the website you already have?
Usually, yes. Retrieval backed search, quoting tools and assistants can be added to most modern sites through an API layer, without a rebuild. The honest caveat is weight: AI features ride on top of your front end, and a slow site makes every response feel worse.
If your platform is already struggling, we will say so, and fix the foundation first with website speed optimisation or a rebuild by our web design team. And if the goal is publishing at scale, our AI assisted content systems pair the drafting speed of a model with a human editor who knows your business.
Carve. Build. Weather.
One method, three moves. Most AI projects fail at the first one: picking a workflow that was never worth automating.
01
Carve the workflow
We find the one job where AI genuinely pays: hours saved, leads answered, work won. One workflow, sharply scoped, with a measurable before and after. Not a moonshot roadmap.
02
Build the system
Rules in typed code, the model only where it earns its place, retrieval for grounding, memory where the job needs it. Shipped into your real stack, not a sandbox.
03
Weather the drift
Models update, data changes, edge cases arrive. Evals, logging and monitoring keep the system honest after launch, so it is still right in month six, not just on demo day.
The stack Abbi is being built on.
This is the stack we build on. Most of it is running today inside Abbi, the AI tutor we are building, and SpruceOS, the automation behind Spruce.
- 01 Claude API Session logic, structured content, summaries and custom AI features built on Anthropic models.
- 02 OpenAI Assistants, workflows, report generation and structured outputs where GPT models fit best.
- 03 Zep and Graphiti Agent memory and temporal knowledge graphs, so your AI remembers context over time.
- 04 PostgreSQL and pgvector Structured data plus semantic search in one reliable database.
- 05 TypeScript and Node.js The backend layer that keeps AI systems deterministic, typed and testable.
- 06 Next.js and React Dashboards, internal tools and product surfaces for AI-driven systems.
- 07 Microsoft Azure App services, storage, queues and hosting for production AI workloads.
- 08 Microsoft Entra Identity, permissions and secure access management from day one.
- 09 Power Automate Email flows, reminders, internal automation and workflow routing.
- 10 Power Apps and SharePoint Internal apps, checklists, documents and operational systems your team actually uses.
- 11 API integrations Connecting forms, CRMs, booking systems, email, data and reporting.
Abbian AI tutor in build: Claude API, memory, retrieval
SpruceOSin production, running our own company daily
7 figureswe grew our own brand, Spruce, from zero on $0 ad spend
No surprise invoices.
Nobody should have to book a discovery call just to learn what AI work costs. The ranges below are indicative, not a quote. Your figure comes down to scope, which we agree in writing before we start.
AI feature
from $2,500
2 to 4 weeks
One AI feature inside your existing site or workflow: smart search, lead triage, a drafting tool. Small, sharp and shipped.
Automation build
$5k to $12k
4 to 8 weeks
Power Platform automation or a retrieval backed assistant, wired into the systems your team already uses every day.
Custom AI system
from $15k
phased, 2 to 4 months
Agents, memory, retrieval and evals, engineered as a product. The full discipline we apply to Abbi, on your problem.
Indicative ranges, not a quote. We scope the exact number with you up front, in writing, before anything begins.
The things people ask first.
How much does custom AI development cost in NZ?
It depends on how much system sits around the model. A single AI feature inside an existing site is the low end. A full custom system with agents, retrieval, memory and evals is a bigger engineering project. The ranges on this page are honest indications. We agree the exact figure in writing before work starts.
Can you build AI into our existing website or systems?
Usually, yes. Most modern websites can take AI features through an API layer without a rebuild, and businesses on Microsoft 365 can automate a long way with the Power Platform tools they already have. We start by mapping what you have. If the foundation cannot carry the feature, we say so before you spend.
Which AI model do you use and is our data safe?
Mostly Claude, from Anthropic, the model we are building Abbi on, and we can build on OpenAI models where they fit a job better. We use the commercial APIs, which do not train on your data by default, and your information stays in databases you control. Access runs through proper identity tools like Microsoft Entra, not shared logins.
How is this different from just using ChatGPT?
ChatGPT is a general chat window. A custom system knows your business: it reads your real data through retrieval, remembers context where the job needs it, and follows rules written in code for every decision that matters. If a general tool already covers your need, we will say so.
How long does an AI automation project take?
We scope a single automation or AI feature at two to four weeks, and a retrieval backed assistant at four to eight. A full custom system runs in phases over a few months, with something working in your hands early, not a big reveal at the end. We scope one workflow first, prove it, then extend.
Strong alone. Stronger wired together.
Let’s build the one that ships.
Tell us the workflow that eats your week, or the AI idea nobody can scope for you. We will tell you honestly whether it needs AI at all, and what it takes if it does. No hype, just the plan.
Or start smaller: get a free AI teardown. Send us your website or your current setup, and we will map where AI would genuinely pay off, and where it would not.
Start a projector email hello@igloostudio.co.nz