Guide · How to Implement AI in Your Business
No hype, no jargon: what to automate first, what it costs in New Zealand, how long it takes, and the mistakes that sink most first AI projects.
If you feel behind, you're not. The data says most New Zealand small businesses are still at the starting line:
of NZ SME decision-makers say they currently use AI, up from 17% a year earlier (MYOB Business Monitor, 2025)
of NZ SMEs had no plans to even evaluate AI (Spark/NZIER survey, via MBIE)
of NZ businesses overall now use some form of AI, up from 48% in 2023 (Datacom, 2025)
The gap between those numbers is the story: big NZ companies have largely adopted AI, while most SMEs haven't started. That gap is a competitive opening. The SMEs that implement AI well over the next couple of years will be operating at a structurally lower cost per job than those that don't.
The right first AI project is a single, repetitive, high-volume process that eats staff hours every week, not a company-wide "AI strategy". Look for work where someone reads unstructured input (emails, documents, forms), makes a routine judgement, and produces structured output (a record, a reply, a report). That pattern is what today's AI does reliably.
Good first candidates I see in NZ businesses:
Rule of thumb: if a task is done more than 20 times a week, follows roughly the same steps each time, and nobody enjoys doing it, it's a candidate.
Before automating anything, record what the process costs you today: hours per week, error rate, and turnaround time. Without a baseline you can never prove the AI worked, and "it feels faster" is not something you can take to your accountant.
Keep it simple: have the person who does the task track it for one normal week. Time per item, items per week, mistakes caught later. Multiply hours by loaded staff cost and you have your number. If the process costs less than a few thousand dollars a year, automate something bigger first. My free ROI calculator does this arithmetic for you.
Use an off-the-shelf AI tool when your process is generic; build custom when the value comes from your data, your systems, or your rules. Both are legitimate. The mistake is not knowing which situation you're in.
A useful test: if you can describe the task without naming any of your own systems or rules, buy a tool. If the description is full of "then it checks my…" and "unless the customer is…", that's an integration job. See my AI workflow automation page for how those get built.
You don't need "big data" or perfect systems to implement AI. What you need is access. The practical questions are: can the AI reach the systems the process lives in (via API or export), is the input consistent enough to describe, and is there a written or tribal-knowledge rulebook for the judgement calls?
Cloud accounting and CRM platforms common in NZ (Xero, MYOB, HubSpot, Microsoft 365, Google Workspace, Shopify) all have solid APIs, so most SME processes are automatable without touching your core systems. Also decide early where data is allowed to live. For regulated industries I default to AWS Sydney for ANZ data residency, with NZ-only or on-premises options where compliance requires it.
Never let a new AI system touch live data or customers on day one. Run it in "test mode" first: it does the full job, logs what it would have done, and a person checks the results. Two weeks of test-mode output tells you the real accuracy rate before anything is at stake.
A pilot should be small enough to ship in 2–3 weeks and cheap enough that abandoning it is painless. This is also where you find the edge cases nobody mentioned in scoping, like the supplier who emails invoices as photos or the customer who replies in all caps with no punctuation. Every automation I build ships with test mode as the default, because switching one on blind is how businesses get burned.
Good AI implementations automate the reading, drafting and routing, and send the high-stakes decisions to a person. Set explicit thresholds: over a dollar value, below a confidence level, or touching a sensitive customer, the workflow stops and asks.
This is also the answer to the hallucination worry. A well-built system is allowed to say "I'm not sure" and escalate; it is never allowed to guess confidently in a customer-facing channel. Ask any AI vendor how their system behaves when it's uncertain. If the answer isn't "it refuses and escalates", keep shopping. (It's how I build customer-facing chatbots that don't invent refund policies.)
AI projects fail in the handover more often than in the technology. The person whose job changes needs to know three things before go-live: what the system now does, what they still own, and exactly how to override or escalate when it gets something wrong.
Frame it honestly: the goal is removing the repetitive 60% of a role so the human 40% gets more attention. In a small business that's usually true, because nobody was idle. Give the team the pilot's test-mode logs to review; watching the AI's actual work (including its mistakes) builds calibrated trust far faster than a training slideshow.
After 4–6 weeks live, compare against the baseline from Step 2: hours saved, error rate, turnaround time. If the numbers hold, you now have a proven pattern, and the second automation is always faster and cheaper than the first, because the plumbing and the trust already exist.
Typical results across SME automation projects:
Ranges reflect typical SMB process-automation outcomes; your project gets baselined against your own numbers, not these.
For a New Zealand SME, a custom AI automation project typically costs NZ$500 to $15,000 fixed-price, plus ongoing running costs that are usually tens of dollars a month, not thousands. Simple single-process automations sit at the bottom of that range; multi-system integrations with compliance requirements sit at the top.
Off-the-shelf tools are cheaper ($20–$100/user/month) and the right call for generic tasks (see Step 3). Beware of two pricing traps: open-ended hourly consulting engagements with no fixed scope, and paid "discovery phases". A competent AI consultancy can scope your first project in a free conversation. Discovery with me is always free; you pay only for a fixed-price build.
A well-scoped first automation should be working in 2–3 weeks and fully deployed within 4–6 weeks, not months. The timeline splits roughly into: scoping and process mapping (week 1), build with weekly check-ins (weeks 1–3), test-mode pilot on real data (weeks 3–5), then go-live with training and documentation. If a vendor quotes six months for your first SME automation, the scope is too big. Cut it down to one process and prove it first.
No. For off-the-shelf tools you need nothing technical. For custom automation, a consultancy handles the build, integration, training and documentation. Your job is knowing your own process. Most of my clients have no in-house IT beyond "the person who's good with computers".
If a repetitive task costs you more than a few hours a week, you're big enough. The economics work from about one saved hour a day. Sole traders usually do best with off-the-shelf tools; custom automation starts making sense once a process involves multiple systems or people.
The Privacy Act 2020 applies to AI systems the same as any other data processing. The practical controls: keep data in ANZ-region infrastructure where possible, redact personal information before it reaches a language model where feasible, log every automated decision, and keep a human accountable for outcomes. The Office of the Privacy Commissioner has published expectations for AI use, and a competent implementer designs to them from day one.
The models will keep improving, but waiting doesn't help: implementation cost is mostly in mapping your process and integrating your systems, and that work carries forward as models improve underneath it. Meanwhile the hours you're trying to save are being spent every week you wait.
Write down the three most repetitive tasks in your business and roughly how many hours a week each consumes. That 20-minute exercise turns "I should look at AI" into a scopeable project, and it's exactly what I work through in a free strategy call.
No. Most of what I build runs in the background inside the tools you already use, so there is nothing new to log into. Where a person does need to act, it is a button or a short form. You get training and written documentation at handover, and monitoring and support continue on a monthly plan from $99.
I find that out before you spend anything. Discovery is free, and if the numbers do not stack up I will say so rather than sell you a build. When it does stack up, the first piece is deliberately small, something that ships in two to three weeks so it proves itself early. The price is fixed before work starts, so a build that runs longer than expected is my problem, not your invoice.
Book a free 30-minute strategy call. Bring the process that's eating your week and I'll tell you honestly whether it's an off-the-shelf job, a custom build, or not worth automating yet. Discovery is always free.
Book a free strategy call