Strategy
Estimating the cost and ROI of an AI initiative honestly
A structured way to estimate AI value and total cost, including review effort, exceptions, running costs and the scenarios that make a case fail.
2 min readLast reviewed
Many AI business cases compare the full cost of today’s manual process with an idealised automated future. Real AI-assisted processes still include review, exceptions, maintenance and model usage costs. An honest estimate includes all of them and still often makes a strong case. When it does not, it saves you from an expensive pilot that was never going to pay back.
Start with a measured baseline
Measure the current process: volume per month, time per item, error rate and the cost of errors, cycle time, and seasonal peaks. Use system timestamps and timed samples rather than estimates from memory, which tend to be optimistic in both directions.
Model the AI-assisted process in full
- Share of items handled automatically, based on measured accuracy rather than hope.
- Review time for items that still need a person, which is usually shorter but not zero.
- Exception handling for items the AI cannot process.
- Time spent correcting AI errors that slipped through.
Count every cost
| Category | Examples |
|---|---|
| Build | Discovery, engineering, integration, evaluation, change management |
| Run | Model usage, hosting, vector storage, monitoring tools |
| Operate | Support, model upgrades, content maintenance, retesting |
| Risk | Controls, audits, incident handling, insurance considerations |
Model usage costs depend on tokens per item, which depend on prompt design and how much context is retrieved. Estimate them from a prototype on real samples, then add headroom for growth and retries.
Run sensitivity analysis
Identify the assumptions that matter most, usually automation rate, review time and volume, and show how the result changes if they are worse than expected. A case that only works under optimistic assumptions should be treated as an experiment, with a small budget and clear stop criteria.
Value beyond labour
Labour savings are the easiest benefit to quantify, but often not the largest. Faster cycle times, fewer errors, better compliance evidence and improved customer experience can matter more. Quantify them where you can, and describe them plainly where you cannot, without inventing numbers.