Most AI pilots are approved with a sentence about running costs rather than a number. "The model costs are negligible." "A few thousand rand a month at most." "We will know after the pilot." Then the pilot runs, the invoice arrives, and the conversation becomes about why the estimate was wrong rather than whether the pilot worked.
The underlying difficulty is real. AI run cost is variable: it depends on volume, on how the agent behaves on hard cases, on prompt design and on the exchange rate. Nobody can quote it as a fixed price before the work has been measured, and anyone who does is guessing.
But "we cannot know" is not the only alternative to a false fixed price. A run-cost quote can be a range built from stated assumptions, each of which the pilot is designed to measure, with budget ceilings and stop conditions agreed before the first call. That is how other variable costs are handled in a business, and it works for AI too.
Tugela Claims Services is fictional, and its systems and figures are illustrative. Tugela's original triage pilot was approved on the strength of "low running costs". It ran at roughly three times what the sponsor had in mind, mostly because nobody had counted the calls per claim. For the decision to roll the agent out across all 18,000 claims a month, the finance director asked for a proper quote.
A quote is a chain of assumptions
Run cost is the product of a handful of factors. Writing them down separately turns a guess into something that can be checked.
For Tugela's rollout the chain was:
| Factor | Assumption | Source | Range |
|---|---|---|---|
| Cases per month | 18,000 | Last 12 months of claim volumes | 16,000 to 21,000 seasonally |
| Model calls per case | 9 | Pilot traces | 6 to 14 by claim type |
| Tokens per call | About 11,500 | Pilot traces | Grows with document length |
| Price per million tokens | Published list price, mid tier, input and output separately | Provider pricing page, dated | Subject to provider change |
| Exchange rate | R18 | Budget rate | R17 to R20 |
The calculation gives a central figure of about R87,000 a month for model costs before any caching or batch savings, within a range of roughly R65,000 to R140,000 when calls per case and the exchange rate move together. That range is wide. It is also honest, and it is far more useful to a finance director than a single number that will be wrong.
Each row has a source. The volume comes from operations data, not from the business case. Calls and tokens per case come from pilot traces, which is why the original pilot should have been instrumented to produce them. Prices come from a dated page. The exchange-rate range is the finance team's own.
What the pilot is for
Framed this way, a pilot has a second job beyond proving that the AI works. It is the measurement exercise that narrows the quote.
That changes what a pilot must record. At minimum:
- Calls and tokens per case, by case type. Averages hide the expensive tail. Tugela's complex motor claims took 14 calls; windscreen claims took 5.
- Acceptance rate and review time, because, as Part 2 showed, these drive cost per useful outcome more than the model price.
- The distribution, not just the mean. If 10% of cases cost five times the median, the forecast must say so. The distribution matters more than the average makes the same point for coding agents.
- Which savings were tested. Caching and batch processing, from Parts 4 and 5, can reduce the run cost materially, but a forecast should only assume savings that the pilot actually demonstrated.
A pilot that cannot answer these questions has not finished, even if the demo impressed everyone. Your first AI project should include failure and cost argues the same from the engineering side.
Ceilings and stop conditions, agreed in advance
A range is not permission to spend without limit. The second half of a good quote is a set of controls that are agreed before the pilot starts, with named people who will act on them.
Tugela's eight-week rollout pilot had a model-cost budget of R60,000 and four thresholds:
- At 50% of the budget, the business owner is notified and reviews the spending trend against the plan.
- At 80%, cost per case is reviewed with finance, and the remaining weeks are re-forecast.
- At 100%, the pilot pauses until the owner explicitly approves more spend. It does not quietly continue.
- Any single case over R40, about eight times a typical run, stops and goes to an assessor. This one is enforced in code, because a runaway loop can spend that in minutes. Your agent spent the money. Can you prove why? shows the evidence trail that should accompany any such stop.
Alongside the budget controls were quality stop conditions, because a cheap pilot that is doing harm should also stop:
- any missed must-escalate claim pauses the pilot immediately for review;
- acceptance below 75% after week four triggers a redesign decision rather than more of the same.
The point of agreeing these in advance is that nobody has to win an argument in the middle of the pilot. The decision was made when everyone was calm.
The forecast after the pilot
When the pilot ends, the quote is re-issued with the measured values in place of the assumptions. Tugela's measured calls per case came in at 8.4 rather than 9, and caching on the agent loop reduced input cost by more than half. The post-pilot forecast for full rollout was R48,000 to R68,000 a month, including the volume and exchange-rate ranges, narrower and lower than the original quote.
That outcome is not guaranteed. Pilots also reveal costs nobody expected: a document type that triples token counts, or an acceptance rate that makes review the dominant expense. The value of the method is that either result arrives as a measured update to a known quote, rather than as a surprise on an invoice.
A one-page template
The quote that went to Tugela's executive committee fitted on a page. Its sections, which work for most AI pilots:
- What is being piloted, in one sentence, with the accepted-outcome definition.
- The assumption chain, as in the table above, with sources and ranges.
- The central estimate and range, in rand, with the exchange-rate assumption stated.
- Savings assumed, and whether each has been demonstrated or is still to be tested.
- The pilot budget, thresholds and who acts at each one.
- Quality stop conditions.
- What the pilot will measure to narrow the quote, and when the revised forecast will be issued.
- What is excluded, for example review time, platform costs and engineering effort, if these are budgeted elsewhere.
It is not a fixed price. It is something better: a cost that can be managed.
Next: Part 7, When AI tooling is a substitution and when it is just another cost. All parts are on the series page.
How CloudNala can help
We help sponsors and finance teams quote AI run costs before a pilot starts: building the assumption chain from real volumes, setting budget thresholds and stop conditions with named owners, instrumenting the pilot so it measures what the forecast needs, and re-issuing the forecast with measured values at the end.
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