Using AI in business
By Custom Software Solutions Team ·
The Companies Winning With AI Aren't the Ones You'd Expect
Everyone assumes the businesses getting the most value from AI are the ones with the biggest budgets, the flashiest custom models, or a dedicated "AI transformation team." That's mostly wrong. In our experience building software for small and mid-sized businesses here in New Zealand, the organisations getting real, compounding value from AI are the ones that treated it as a workflow problem, not a technology problem.
The companies burning the most money on AI right now are the ones who bought the tool first and asked "what do we do with this?" second. The ones quietly getting ahead skipped that step entirely.
The Misconception: More AI Adoption Equals More Value
There's a pervasive belief that AI value scales with how much AI you're using — more tools, more integrations, more automated decisions. This leads businesses to chase breadth: a chatbot here, an AI writing assistant there, some predictive analytics bolted onto a dashboard nobody reads.
Why This Fails
Each of these tools, in isolation, produces a small, disconnected improvement. The chatbot saves a support agent a few minutes. The writing assistant speeds up a first draft. But none of it compounds, because the underlying workflow hasn't changed — you've just inserted AI into the gaps of an old process instead of redesigning the process itself.
The result is a business with a dozen AI subscriptions and no measurable change in output, margin, or customer experience. Worse, it creates tool fatigue — staff start ignoring the AI features because none of them meaningfully reduced their workload, they just shuffled it around.
What Actually Works: Narrow, Deep, Owned
The businesses seeing real returns picked one or two high-friction processes — the ones everyone complains about internally — and went deep. Not broad.
Depth Beats Coverage
A good example is quoting and estimating. It's tedious, it's repetitive, and it's usually done by your most experienced (and most expensive) staff, because it requires judgment. A business that applies AI narrowly here — training a model or workflow on past quotes, jobs, and outcomes — can compress a two-hour task into fifteen minutes and reduce variance between staff who are good at estimating and staff who aren't.
Compare that to a business that adds a generic AI assistant to every department "to see what sticks." The quoting business gets a defensible, compounding advantage. The scattergun business gets a bigger software bill.
Ownership Matters More Than Sophistication
The businesses doing this well also tend to own the workflow rather than renting someone else's black box. That doesn't mean building your own large language model — that's rarely necessary or sensible for a small business. It means understanding exactly what data goes in, what logic sits in the middle, and what comes out, so the system can be adjusted as the business changes.
If you can't explain why your AI tool made a particular recommendation, you don't have a system — you have a slot machine. That's fine for low-stakes tasks like drafting a marketing email. It's not fine for anything touching pricing, hiring, or customer commitments.
The Uncomfortable Part: It Exposes Bad Processes
Here's the angle most AI commentary avoids: introducing AI into a business is a forcing function that reveals how broken your existing processes already were.
AI Doesn't Fix Chaos, It Amplifies It
If your sales pipeline data is inconsistent, an AI layer built on top of it won't magically produce good forecasts — it'll produce confident-sounding bad forecasts, which is arguably worse than no forecast at all. If your customer support tickets are inconsistently tagged, an AI summarisation tool will summarise nonsense fluently.
This is why so many AI pilots quietly die after a few months. It's rarely the model's fault. It's that the business had no clean, consistent process underneath it to begin with, and nobody wants to admit that the real project was process cleanup, not AI adoption.
The Businesses That Benefit Most Do the Boring Work First
The companies we've seen succeed spent more time fixing data entry habits, standardising terminology, and clarifying decision rules before introducing AI than they spent on the AI implementation itself. It's unglamorous. Nobody writes case studies about "we spent six weeks getting our spreadsheet naming conventions consistent." But that's the actual unlock.
A Smaller, More Useful Question
Instead of asking "how can we use AI in our business," a more useful question is: "which specific, recurring decision takes up disproportionate time relative to its actual complexity?"
That's usually where AI pays off first — not in the glamorous, strategic decisions, but in the repetitive, moderately complex ones that never quite justified hiring another person, yet quietly drain hours every week. Rostering. Invoice reconciliation. First-pass customer enquiries. Draft contracts. Inventory reordering thresholds.
These are rarely the tasks featured in AI marketing material, because they're not exciting. But they're where the time actually goes.
Rethinking the Timeline
The other quiet misconception is that AI adoption is a one-off project with a start and end date. In practice, it behaves more like hiring a junior employee who happens to work at software speed: it needs onboarding, correction, feedback, and ongoing supervision. Businesses that treat their first AI implementation as "done" after launch tend to see performance quietly decay as the business changes around it and the tool doesn't.
The businesses that treat it as a standing responsibility — someone owns it, reviews its output monthly, adjusts its inputs as things change — get compounding value over time instead of a brief spike followed by drift.
One Thing to Do This Week
Pick the single task in your business that people complain about most often in passing — not in formal meetings, but in the "ugh, I have to do the roster again" sense. Spend thirty minutes this week writing down exactly how that task is currently done, step by step, including all the messy exceptions. Don't touch AI yet. That document alone will tell you whether the problem is a process problem, a tooling problem, or both — and that distinction will save you from buying the wrong solution.