The most valuable thing a fencing supplier in Auckland taught his AI was how long it takes to drive between jobs.
Not a chatbot. Not a content engine. Drive time. He is a one-man operation plus a contractor, and the thing quietly wrecking his week was getting booked into two jobs across town with no realistic way of reaching the second one. So the instructions he wrote for his AI included how to account for travel between sites when it builds his calendar.
It is a small detail, and it is the whole game. Because if you had asked that owner six months earlier what implementing AI looked like, he would have described something with a screen and a subscription. What actually moved the needle was somebody sitting down with him and working out, in unglamorous detail, where his week was leaking.
That gap between what people expect AI implementation to look like and what it actually looks like is where most of the wasted money goes.
The short version
Implementing AI into a business is a five-step sequence, and choosing a tool is step five, not step one:
- Review last week and list the repetitive tasks you did not want to be doing, in mechanical detail
- Describe the worst one to an AI tool and ask how it would map the process out
- Save the solution as reusable instructions rather than a prompt you retype
- Test it on your most awkward cases, not your cleanest one
- Share what works internally, then choose and pay for the tool that fits
Everything in steps one and two happens in a notebook. That is not a delay before the real work. It is the real work.
Why the tool question is the last question
Ask a business owner where they are with AI and you almost always get a product name back. ChatGPT. Copilot. Someone in accounts has started using Gemini for something.
It is a confident answer to a question that did not need asking yet. The major tools are broadly capable and getting better every few months. What varies wildly between two businesses in the same industry is not which software they bought. It is whether anyone has done the tedious work of identifying which specific, repetitive, unbilled tasks are eating the week.
Skip that, and you are shopping for a solution while still hunting for the problem. Which is precisely how a New Zealand business owner recently came within a signature of paying $100,000 to have an AI-integrated CRM built for him. The capability he was describing sells for something nearer $100 a month. He was not naive about technology. He just started at the product end, and once you start there, someone will always be delighted to sell you the biggest version of it.
The barriers holding New Zealand businesses back are not cost, and they are not scepticism about whether AI works. MBIE's research into AI adoption among SMEs points at lack of knowledge and uncertainty about where to start.
So start somewhere deliberate.
Step one: review last week, not your industry
Do not go and research what businesses like yours are doing with AI. Look at what you personally did in the last seven days, and write down every task you did not want to be doing.
Then get uncomfortably specific about the mechanics. Not "admin". Something closer to: I pulled the job details out of an email, retyped them into Xero, copied the address into my calendar, and guessed how long the drive would take.
That level of detail matters because the handovers are where the time actually disappears. Three systems, one process, all of it manual, and a context switch either side of every step. A task that feels like five minutes is rarely five minutes.
Most people skip this because it feels like procrastination. It is the only step that determines whether anything after it is worth doing.
Step two: ask it how, do not tell it what
Take the worst thing on that list, describe the process, and ask the AI how it would map it out.
The word doing the work there is ask. There is a strong habit of treating these tools as a pair of hands, issuing instructions and grading the output. The bigger opportunity is using one as a consultant on your own operation. Describe the mess. Ask how it would approach it, what it would need from you, and where it expects the process to break.
Two adjustments make this dramatically better.
Tell it to ask you questions first
Left to itself, it fills gaps with assumptions and then quietly builds on top of them. Told to interrogate you first, it puts those assumptions somewhere you can correct them.
Tell it to ask one question at a time
A block of ten questions is how people quietly abandon the exercise. One at a time reads like a conversation, and you finish it.
Underneath both of those sits the principle that explains most disappointing AI results. These tools have read an enormous amount of the internet and know absolutely nothing about your business. Every genuinely useful output comes from closing that gap on purpose, with your context, your templates, your way of doing things. What you get out is mostly a function of how much of your business you were willing to put in.
Step three: write it down once so it repeats
A prompt you retype every time is not a system. It is a chore wearing a costume.
Most serious AI tools now let you save standing instructions the tool follows whenever you call on it. Write out how you invoice, what your proposal template contains, who gets copied into calendar invites, how to handle travel between jobs. Then "draft my proposal" runs the whole thing, the same way, every time.
This is where the value compounds. It is also where dependence gets designed out, because those instructions are written in plain English. Change the way you invoice and you tell it, in a sentence, and it updates itself. Nobody needs to book a consultant to make that edit.
Step four: test the ugly cases, not the clean one
This is the step people get wrong, and it decides whether any of it survives contact with a real week.
Almost everyone tests a new automation on a tidy example. Of course it works. The tidy example was never in question.
Feed it the awkward one instead. If you have set up email drafting, do not check it against the routine enquiry that arrives forty times a month. Give it the annoyed customer, the half-relevant request, the question that sits just outside what you actually offer. Then make a decision on purpose: either the process handles that properly, or you build in a flag so it never touches those and a human picks them up.
Deciding where AI does not go is as much a part of implementation as deciding where it does. A narrow system you trust completely beats a broad one you have to check constantly, every time.
The same discipline scales to bulk work. One neat example: sorting several hundred photos and video clips from a year of travel and events, with AI tagging the contents and filing them. The move that made it safe was not turning it loose on the files. It wrote its proposed classifications into a document first, and that document got reviewed before anything moved. Nobody was going to inspect 600 files individually, but a two-minute scan of the logic was enough to see whether it was thinking sensibly.
Step five: share it, then go tool shopping
Two habits keep this alive after the first win.
Share what works
The standard failure in any business past a handful of people is two teams doing the same job, one of them at twice the speed, neither aware of the other. A channel where people post what they have worked out costs nothing and spreads sideways fast. Externally, talking about it with other owners reliably returns more than it gives away, and the best ideas often arrive from an industry nothing like yours, because the process turns out to be identical even when the application is not.
Then choose the tool
By now you know what you are shopping for, which changes the conversation entirely. Claude handles a lot of general business work well at the moment, largely because it connects natively into systems like Xero, Gmail and Google Calendar. If your team is already fluent in ChatGPT or Gemini, that familiarity has real value. Specialist tools exist for content production and technical workflows, and no single tool covers everybody.
Whatever you land on, pay for it. Free tiers generally train on what you feed them. Business plans switch that off.
Three rules that keep this out of trouble
Pay for the tools
See above. If anything confidential goes near it, the subscription is not optional.
Start read-only
The common mistake is connecting AI to every system at once with full write access and then finding out what it does. Give it read-only permissions, watch it for a fortnight, and add capabilities one at a time. Some connectors help by default: linked to Xero, an assistant can draft an invoice but cannot send anything to a customer.
Keep someone who can verify the output
AI is only ever as good as the person checking it. Owning an AI tool does not make you a chartered accountant, which means you cannot properly verify accounting output on your own. Run that test over every process you plan to automate. If nobody in the business could catch the mistake, it is not ready.
What the time is actually for
There is a comfortable story about AI where the hours it hands back turn into afternoons off.
Ask the people implementing this in New Zealand small businesses what owners actually do with recovered time and the answer is almost unanimous. Nearly all of it goes straight back into growth. More quotes, more site visits, more conversations. The fencing supplier's next project was not a holiday, it was a research routine so that when he drives past a business he wants to work with, he can drop the name in and get back who they are and who to call.
Which reframes the whole exercise. Back-office automation is not a cost-cutting measure. It is a way of buying more selling time. The admin reduction is a sales capacity increase in disguise, and that is a far better reason to bother.
There is funding for the planning part
Worth knowing before you spend anything of your own. MBIE runs an AI Advisory Pilot through the Regional Business Partner Network, co-funding up to 50 percent of AI advisory costs, capped at $15,000, to develop and implement a plan tailored to your business. After strong demand it was expanded to support up to 150 businesses and extended to run until 31 January 2027, with wider eligibility. You need to be an existing Regional Business Partner Network customer, so your local RBP is the place to start.
Note what it funds. Not software. Planning. Which is the correct end to begin at, and rather the point of all of the above.
Want the full conversation? Listen to "How to Actually Use AI in Your Business, with an AI Engineer" with Blake Harkness on the Marketing 4 Business podcast.
