Nearly every AI business case contains a sentence about replacement. The agent will replace hours of manual work. The coding assistant will replace a contractor. The document tool will replace an outsourced service. The cost of the AI is then set against the cost of what it replaces, and the case looks strong.
A year later, the AI invoice is real and recurring, and the cost it was meant to replace is often still there too. Nobody lied. Time was saved. But saved time is not saved money unless something that cost money actually stopped.
This part sets out a simple test for telling the two apart, and then argues the counterweight that most AI cost conversations skip: for a great deal of AI-assisted work, the token bill is too small to matter, and effort spent squeezing it is effort wasted. For some work, the opposite is true. Knowing which situation you are in is the whole skill.
Tugela Claims Services is fictional, and its systems and figures are illustrative. Tugela's triage agent business case, approved before the work in this series, said the agent would "free up the equivalent of six assessor posts". Tugela has 26 assessors.
Substitution means a cost stopped
The test has three questions, applied in order.
Did a cost stop? Fewer hires than would otherwise have been needed, overtime that no longer happens, an outsourced service cancelled, a temporary team not renewed. If yes, the AI is a substitution, and the comparison is like for like: the full cost of the AI path against the cost that stopped.
If not, did measured output or quality rise? More claims handled by the same team, faster turnaround that customers notice, fewer errors that previously cost money. If yes, the AI is an addition with a return. It may well be worth funding, but the case rests on that evidence, not on replacement.
If neither, is a measurement planned, with a review date? If yes, it is an experiment with a deadline. If no, it is another cost, and it should be treated as one: budgeted, reviewed and cancelled if nobody can show what it does.
None of these outcomes is a failure. An addition with a measurable return is often the better investment. The failure is presenting an addition as a substitution, because then the budget assumes a saving that never arrives.
What happened at Tugela
Nobody was let go, and six posts were not freed in any sense a finance system would recognise. But when the agent's first year was assessed honestly, two things had stopped costing money.
Tugela had taken on a new insurer client during the year, increasing claims volume by about 12%. The pre-AI staffing model would have needed three more assessors. None were hired. That is substitution: three posts avoided, at a loaded cost the finance director could state precisely.
Month-end overtime, previously a recurring line, had fallen by about two thirds because the backlog no longer peaked. That is also substitution.
The rest of the assessor time saved, about nine minutes per accepted claim according to Part 2, went into more thorough handling of complex claims. That may be producing value in fewer leakage payments, but it has not yet been measured. Under the test, it is an experiment, and it now has a review date.
The revised business case is smaller than the original and much easier to defend.
The whole cost line, not only the invoice
The comparison only works if the AI side includes everything. In practice that means:
- model and tool costs, from the invoice;
- licences for any AI products involved;
- review time by the people who check the output;
- rework when the output is rejected;
- engineering and assurance: the people who maintain the prompts, run the evaluations, watch the monitoring and handle incidents;
- platform costs for hosting, logging and storage.
A coding assistant that "replaces a contractor" but requires a senior engineer to review and fix its output for several hours a week has a different substitution case once that time is counted. You bought Copilot. So why is there another AI bill? and Why AI is an operating cost cover the licence side of this; the point here is that the people around the tool are part of its cost.
The honest counterweight: often the tokens do not matter
Most of this series is about measuring and controlling AI run costs. It would be misleading not to say plainly that in many organisations, for many workloads, the token bill is too small to be worth optimising.
Tugela's whole AI invoice, after the batch changes in Part 5, is about R90,000 a month. The triage function it supports costs nearly R1 million a month in assessor time alone. Model tokens are under 10% of the cost of the work. A heroic engineering effort that halved the token bill would save around R45,000 a month. A change to the review screen that saved one minute per claim would be worth considerably more, as would one percentage point of acceptance rate.
This pattern holds widely. Where AI assists skilled, salaried people with high-value work, the people are the cost. Token optimisation should be proportionate: get the obvious things right, such as prompt order for caching, batch for work that can wait and a sensible model tier, then spend effort on acceptance, review time and the quality of the work.
When the tokens are the bill
The second picture in the diagram is a different kind of service: a free public question service answering several hundred thousand questions a month with two support staff behind it. Here the model cost is more than half of the total, and it scales directly with use.
Token cost tends to dominate when several of these are true:
- high volume, low value per interaction, such as public information, consumer chat or free tiers;
- few people in the loop, so there is no large salary line to dwarf it;
- agentic loops or long context, where each interaction carries many calls or very large documents;
- the service is free or flat-priced to the user, so cost grows with popularity and revenue does not.
In those situations everything in Parts 3 to 6 matters a great deal, and so does the pricing question in Part 8. A public-sector service in particular can find its AI budget consumed by success. Public-sector AI workflows shows the shape of such services; the economics need to be designed with them, not discovered afterwards.
Questions for the next business case
Before approving the next AI investment justified by replacement, ask:
- Which specific cost will stop, and when? Name the line in the budget.
- If no cost stops, what measured output or quality will rise, and how will it be measured?
- Does the AI cost include review, rework, engineering, assurance and platform, or only the invoice?
- Is the token bill likely to be a rounding error or the main cost? Which of the conditions above apply?
- What is the review date, and who decides whether to continue?
A business case that answers these is less exciting than "frees up six posts". It is also one that will still be true a year later.
Next: Part 8, Pricing an AI feature you sell. All parts are on the series page.
How CloudNala can help
We help leadership teams test AI business cases before approval and review them honestly afterwards: separating substitution from addition, counting the whole cost line, and deciding where token optimisation is worth the effort and where attention belongs on review time, acceptance and the work itself.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an AI Cost Review or write to us at consult@cloudnala.co.za