Cash handling in small NZ teams — the small AI that actually pays back

Cash handling is one of those parts of a small business that nobody really wants to talk about until it goes wrong. Tills come up short. A bank drop happens late. Someone reconciles a week behind because it was busy. None of these things look serious in isolation, and most of the time they aren't. They just quietly cost you money, and they hide other problems while they do it.

The reason I keep coming back to it is that cash handling is one of the few places where small, narrow AI assists actually pay back fast. Not the big "AI transformation" you read about. The small kind — categorisation, anomaly flagging, reminders, a simple weekly check that runs whether you remember it or not. The reason it pays back fast is the same reason cash issues hurt: the underlying process is high-volume and money-sensitive. Get it slightly more reliable and the savings compound week after week.

This is the playbook I keep recommending to small NZ teams that still process meaningful amounts of physical cash, mixed payment types, or rely on a single person to keep reconciliation on track.

Why cash handling breaks first when staff change

The pattern I see is consistent enough that I assume it before I even ask. When a small team has steady staff, cash handling is fine. The same person opens up, the same person banks, the same person reconciles, the rules live in their head. Then one of three things happens: that person leaves, that person goes on leave, or the business gets busy enough that the rules get bent on a Saturday and nobody writes the bend down.

After that, things drift. A coin float doesn't get topped up. A refund gets processed differently than usual. EFTPOS settlements stop matching daily takings cleanly. Two weeks later someone notices a $300 gap and there's no way to walk back through which day it came from.

This isn't really a cash problem. It's a process-resilience problem with cash as the medium. The reason I'm pointing it out before any AI talk is that the fix order matters: you fix the ownership and the cadence first, then you add tooling. AI doesn't fix a process that has no owner. It just makes a worse process run faster.

What "small AI" actually means in this context

When I say "small AI" I don't mean an agent that runs your accounts. I mean three or four narrow capabilities that quietly remove friction from the existing process. They're easier to set up than people expect, they fit inside tools you probably already use (MYOB, Xero, your POS), and they're easier to govern because the scope is small.

The four that pay back the fastest:

• Transaction categorisation. Most accounting tools now suggest a category for a bank line based on the description and history. Letting that suggestion auto-apply for high-confidence matches (and only those) saves a real chunk of weekly admin. The human still reviews exceptions.
• Anomaly flagging. A simple rule that says "tell me when today's takings are more than 30% off the same weekday last month" catches more real problems than a quarterly review ever will. You don't need anything fancy — a weekly summary with the outliers highlighted is enough.
• Reminders against a cadence. Not generic reminders. Specific ones tied to the cash process: "bank drop not logged by 4pm Friday", "EFTPOS settlement variance over $50", "till float not topped up before open". Things that fail silently otherwise.
• First-pass reconciliation drafts. AI is reasonably good at matching obvious lines (settlements, recurring suppliers, regular customer payments) and flagging the rest. You still approve. You just stop spending an hour clicking through the easy ones.

None of this is glamorous. That's the point. The teams that win with this stuff aren't running anything experimental — they're running boring, narrow assists in places where the process was already a bit fragile.

A simple weekly cadence that catches most problems

Here is the cadence I'd suggest if you're starting from scratch. It assumes you already have a basic accounting tool and some kind of POS. It doesn't assume any custom build.

1. Daily, end of trade. Cash takings counted, bank drop logged in a shared sheet (or your POS) with the person who counted. Five minutes. Owned by whoever closed.
2. Weekly, Monday morning. A 20-minute pass over the previous week. Bank lines reconciled with auto-categorisation handling the easy 70%. Variances over a set threshold flagged automatically and investigated. Anomaly summary reviewed.
3. Monthly, before the BAS/GST cycle. A clean review with the bookkeeper or accountant. By this point the noise should be gone and the conversation should be about the actual exceptions, not the data quality.

The weekly pass is the one most teams skip. They go from daily counting straight to monthly. That's where the timing problems hide — payroll funding gaps, supplier double-payments, missing settlements. A 20-minute weekly pass with one or two automated flags catches most of them while they're still small.

AI doesn't fix a process that has no owner. It just makes a worse process run faster.

What I'd not bother automating yet

I get asked about this stuff often enough that it's worth saying out loud what I'd leave alone for now in a small team.

Cash forecasting. Tempting, and there are tools that promise it, but for a 2-to-15-person business with reasonably steady patterns, a spreadsheet built once with your accountant beats anything an AI will generate this year. The accuracy improvements aren't worth the setup time and the false confidence is a real risk.

Fraud detection. Most small businesses don't have the transaction volume for the patterns to mean anything. The kind of fraud that hits small NZ teams is usually a person, not a pattern, and your weekly anomaly flag will catch it long before any model would.

Customer payment chasing via AI-generated emails. I'd send templated reminders, sure. But auto-generated personal-sounding emails from "you" that you didn't write are a trust risk and a customer relationship risk. The cost of getting one of these wrong in a small NZ town is real. Stick to clear, factual reminders that look like reminders.

The rule I use: if an AI assist is operating in the background and a person reviews exceptions, that's fine. If an AI assist is generating something that goes out the door under your name without review, that's where the wheels come off.

The piece most teams miss: the recovery plan

This part isn't about AI. It's the bit that makes any cash process actually work, and most small teams don't have it written down.

When something breaks — a missed bank drop, a system outage, the bookkeeper sick on reconciliation day, a power cut at end of trade — what's the fallback? Who steps in? What's the manual process you use until normal service resumes?

I'd write a single page. Three scenarios, two paragraphs each. Pin it next to wherever cash gets counted. Walk a new staff member through it on day one. The point isn't that the document is impressive. The point is that "what do we do when X breaks" has been answered once, by the owner, and not improvised at 4:55pm on a Friday by whoever is around.

This is the piece that makes the small AI assists worth doing. They reduce the day-to-day load enough that there's room for the team to actually follow the recovery plan when they need to, instead of being permanently behind on the basics.

Where to start this week

If your team handles cash and you've nodded at a few things in this post, here's the smallest useful step:

• Write down who reconciles and when. One sentence. Stick it somewhere visible.
• Turn on whatever auto-categorisation your accounting tool already has, set the confidence threshold conservatively, and review what it does for two weeks before expanding.
• Set one weekly anomaly flag. Pick the metric you'd most want to know about — a takings variance, a missed settlement, a recurring supplier amount that changes — and have your tool email it to you Monday morning.

That's it for the first month. No platforms, no agents, no transformation. Just the boring controls that quietly stop costing you money.

If you want help setting this up for your business — picking the right one or two automations and getting the cadence to actually stick — that's the kind of thing I help with. The parent framework I use for picking the first automation covers how I think about it more broadly, and you can see how this fits into a wider workflow setup if you want to go further.