Datacom’s managing director offered an analogy this year that I keep repeating to people: you wouldn’t hire a 50-seater bus to move two people from Hamilton to Wellington.
That’s roughly what happens every time someone uses an expensive frontier AI model to tidy up an email.
Now here’s why that stopped being a funny observation and started being a line item.
The pricing shift most businesses haven’t priced in
The major AI vendors are moving customers onto consumption billing. You pay for what gets used. Uber has capped its AI usage. Microsoft and Amazon — companies whose entire commercial interest lies in more AI usage, not less — have reportedly stepped back from internally encouraging maximum experimentation.
Meanwhile only a third of New Zealand organisations say their AI returns exceed their costs. Seventy-nine per cent increased AI investment over the past year anyway.
So the bill is becoming variable at precisely the moment most businesses can’t say which of their AI use is productive.
That’s a board conversation waiting to happen, and in a lot of New Zealand businesses it’ll happen in the next two quarters.
The instinct, and why it’s expensive
The instinct is to cap it. Set a limit, contain the exposure, revisit in six months. Entirely understandable. I’d also argue it’s the most expensive option on the table.
A cap is indiscriminate. It cannot distinguish between the person running a 50-seater bus to polish an email and the person who has quietly found something that saves eleven hours a week. It takes both.
And it takes the second one permanently. Because that person now has their answer about what this organisation does when someone finds something that works: it introduces a spending limit. They won’t bring you the next one.
Look at the shape of the Datacom data again. Seventy-nine per cent of organisations increasing investment. Core-operations use down from 8% to 4%. Spend rising, depth falling. That’s the signature of an organisation buying more of something it can’t steer — and the reflex response to that signature is always to cut, which fixes the spend and guarantees the depth keeps falling.
The lever that actually works
The alternative to capping spend is knowing which spend is working.
That sounds so obvious it’s almost insulting to write down, and it’s almost never in place. There’s a reason, and I’ve written about it separately: a meaningful share of your staff are concealing their AI use — 28% of New Zealand workers use tools their employer doesn’t know about. You cannot cost-optimise a system you can’t see. Every efficiency exercise run on a partial picture ends up cutting the visible use, which is not the same as the wasteful use.
So the sequencing matters. Visibility, then value, then spend. Reverse it and you’ll cut confidently and wrongly.
Three things that are cheaper than a cap
Match the task to the tool. Not everything needs a frontier model. Most routine drafting doesn’t. Most summarising doesn’t. A business that has never had this conversation is almost certainly overspending on capability it isn’t using — and, at the same time, underusing the expensive stuff where it would genuinely pay.
Make value visible before you make spend visible. If the first AI conversation your team hears is about cost, you’ve just told everyone that AI is a budget risk rather than a capability. Good luck getting honest information after that.
Fund the method rather than the consumption. A predictable, fixed cost for building capability is easier to govern than a variable usage bill — and, more to the point, it’s the thing that brings the variable bill down, because people who know what they’re doing use less of the expensive stuff.
I should declare an interest: the third point is close to what I work on, so read it with appropriate scepticism. But the underlying maths isn’t mine and isn’t controversial. The cheapest AI spend in any business is the spend that’s aimed.
The thing I’d actually worry about
Not the bill. Bills get managed.
What I’d worry about is that the cost conversation arrives before the capability conversation in most New Zealand businesses, and once it does, AI gets filed permanently under “expense to control” rather than “capability to build.” That framing is very hard to reverse, and I think a lot of businesses are about six months from locking it in.
If you’re going to have the cost conversation — and you are — try to have the other one first.