The Economics of AI Delivery
Tokens are how AI is metered, not how it is valued. This eight-part series treats AI run cost as something to manage rather than fear: why a token is a unit of compute and not of value, why cost per prompt misleads and cost per accepted outcome does not, how to pick the smallest model that passes your own evaluation, how prompt caching and batch pricing actually work, what the rand does to a dollar bill, how to quote a pilot before it starts, when AI is a substitution rather than another cost, and how to price an AI feature you sell. One fictional South African claims administrator carries the examples, every figure is illustrative, and provider prices were checked on 27 October 2026.
8 articles
- Part 1A token is a unit of compute, not a unit of valueTokens are the meter on every AI invoice, so they end up on every AI dashboard. They tell you how much computing the work took. They say nothing about whether the work was worth doing. Part 1 sets up the economics of AI delivery on that distinction.9 min read
- Part 2Cost per prompt is the wrong metric: measure cost per accepted outcomeOne request from a user can trigger nine model calls, and one in five results may be thrown away. Cost per prompt hides both. Part 2 builds the number that does not: everything it took, divided by what was actually used.6 min read
- Part 3Use the smallest model that reliably passes the evaluationModel choice is usually argued from benchmarks and price lists. It should be decided by your own evaluation: set the quality bar first, then pick the cheapest option that clears it, and route by segment where a smaller model earns its lane.6 min read
- Part 4Prompt caching is an architecture decisionPrompt caching can cut the input cost of repeated context by 90% or more, or it can add a 25% premium for nothing. Which one you get depends on prompt order, timing and tenancy, all of which are design choices. Part 4 explains the mechanics.7 min read
- Part 5Batch, realtime and the randMost AI workloads run at the realtime price because that is how the prototype was built, even when nobody is waiting for the answer. Batch pricing halves the bill for work that can wait. The exchange rate then moves the rand total without a line of code changing. Part 5 deals with both.6 min read
- Part 6Quote the AI run cost before the pilot startsAI pilots are often approved with a vague promise that running costs will be small, then surprise everyone. You cannot quote a fixed price for variable work, but you can quote a range built from stated assumptions, with ceilings and stop conditions agreed before the first call. Part 6 shows how.5 min read
- Part 7When AI tooling is a substitution, and when it is just another costMost AI business cases claim to replace effort. Most AI invoices arrive on top of the effort they were meant to replace. The test is whether a cost actually stopped. Part 7 also gives the honest counterweight: often the token bill is a rounding error, and sometimes it is the whole bill.6 min read
- Part 8Pricing an AI feature you sell: margin per case versus per seatWhen you sell software with AI inside it, your costs become variable and your price may not be. A flat seat price makes your heaviest users your least profitable. Part 8 works through the margin maths, the price models that fit, and the clauses a rand price needs when the cost is in dollars.6 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala