A small ministry organisation asked us a fair question: what will all this AI cost us over the next year? They wanted a number they could take to the people who hold the purse.
The tempting thing is to give them one. Add up the obvious costs, write a figure at the bottom, send it over. Everyone feels good for about a week — until the world moves and the figure quietly stops being true.
So we didn't give them a number. We gave them something they could keep using.
Think of it like budgeting a kitchen, not buying a kettle
A kettle has a price. You pay it once, you take it home, done. A kitchen is not like that. The ingredients are priced in a currency that drifts. The big appliances get cheaper or vanish from the shelves depending on the month you walk in. And you don't fit the whole thing out in one afternoon — you buy the essentials now and add the rest as you actually need it.
An AI budget is a kitchen, not a kettle. If you treat it like a kettle, you'll be wrong on the day you print the page. There are three honest layers, and skipping any one of them turns the budget into a guess wearing a suit.
Layer one: price it in the money you actually earn
Almost everything in this kind of setup is priced in dollars. The organisation's income is in rand. That gap between the two currencies is a real cost, and the usual mistake is to hide it — quote everything in dollars and leave the reader to do the sums in their head.
We did the opposite. We built the whole budget in rand, and we wrote the exchange rate down in plain sight as something you can change. When the rand moves — and it always moves — the person looking at the budget changes that one figure and watches everything else update. The rate stops being a hidden assumption buried in every line and becomes a dial anyone can see and turn.
Layer two: leave room for prices that won't sit still
The machines this runs on don't hold their prices. While we were still building the budget, the ground kept shifting under us — a popular configuration got discontinued, a key component started getting more expensive almost weekly, and a price we'd written down was already on its way to being out of date.
So we built in a cushion. Not padding for its own sake — a deliberate line that says, out loud, "the hardware is going to cost more than today's sticker, and here is where we've allowed for that."
A budget that assumes today's prices hold for a year is a snapshot of one afternoon, sold as a plan for twelve months.
If your numbers quietly pretend the prices will freeze, you've already lied to the person reading them. The cushion is where you tell the truth in advance.
Layer three: a path you can walk, not a bill you must pay at once
The last layer is sequence. Instead of one big total, we laid out three steps — a modest starting setup, a heavier middle one, and a full version held back for a future decision when the rest is real.
These aren't three quotes for the same thing. They're a path. You commit to the first step now. You take the second when you've outgrown the first. The third stays on the shelf until someone with the authority decides the moment has come. Each step earns the next one, instead of asking anyone to swallow the whole cost on day one.
Why we handed over a spreadsheet instead of a figure
A single number is easy to produce and easy to be wrong with. It hides the currency risk, hides the hardware risk, and pretends a year of buying decisions is one moment of writing a cheque.
The thing we actually delivered does the opposite. The exchange rate is a dial. The hardware cushion is its own visible line. The three steps are a path the organisation walks at its own pace. When the rand shifts or a price jumps, the owner changes one input and reads the new total — without phoning us.
That's the whole method, and it isn't only for ministries. Any small organisation here needs the same three honest layers: the currency it's truly exposed to, the hardware risk it can't dodge, and a path it can afford to take one step at a time. Hand someone a single number and you've given them false confidence. Hand them the working model and you've given them control.
For the technical reader
The brief was a twelve-month total-cost-of-ownership budget for the AI capability of a seven-to-ten-person ministry organisation. The deliverable was a six-tab Excel workbook, sent on 2026-06-10 — not a slide deck, not a memo with a figure on the last page. The workbook was the product: the owner can open it, change one assumption, and watch the rest recalculate.
Layer one — currency. The whole budget was modelled in rand as the primary currency, with the USD/ZAR rate stated explicitly as an editable cell at R16.55 to the dollar, rather than baked invisibly into every conversion. Subscriptions, cloud compute and GPUs are all dollar-denominated; income is in rand, so the exchange gap is a first-class, stress-testable line item.
Layer two — hardware contingency. The budget split across six cost lanes: subscriptions, cloud compute, on-premise inference, fine-tuning, infrastructure, and contingency. The contingency lane was a direct response to live volatility during the build: Apple had discontinued the larger Mac Studio configurations, DRAM was repricing weekly off a memory shortage, and a unit priced at R89,999 was pre-hike stock unlikely to hold. A hardware-horizon section was added the same day, after that week's product announcements and fresh South African retail pricing, distinguishing what was shipping, what was announced but not shipping, and what was not worth waiting for.
Layer three — staged procurement. Three adoption tiers, costed as a gated path rather than three competing quotes:
- Tier one — R485k: the working baseline (subscriptions, cloud burst, first inference hardware).
- Tier two — R1.07M: a heavier on-premise build for sustained local inference and fine-tuning.
- Tier three — R2.16M: the full split-stack vision, deferred to a later financial-year decision once the earlier tiers are proven and the next hardware generation can be re-specced in.
Each tier is gated on the previous one actually working — approve tier one now, buy the first inference node when the right hardware launches, leave tier three as a future leadership call.