Most AI business cases in procurement fail before anyone builds anything. Not because the idea is bad, but because the numbers are vague — "saves time", "improves visibility" — and a CFO cannot approve a feeling. The fix is boring: a small set of formulas, honest inputs, and a payback period that fits on one line.
This is the calculation I use for AI use cases in procurement, with three worked examples from builds that exist on this site. Every number is illustrative. Replace them with yours, and measure your current process for two weeks before you trust any of them.
Where the value of an AI use case comes from
AI in procurement creates value in four ways, and a business case should name which ones it relies on. Time: hours a buyer no longer spends reading, retyping or drafting. Price: money recovered or protected because something got checked that used to go unchecked — invoice prices, renewal dates, index clauses. Speed: a shorter sourcing cycle, which matters when a production line is waiting. Risk: an auto-renewal that did not happen, a supplier issue caught a month earlier. Time is the easiest to prove, price is the easiest to sell, and risk is the hardest to put a number on, so use it as a bonus rather than the core.
The formulas
Loaded hourly cost means salary plus employer contributions and overhead, not the hourly wage. For a mid-senior buyer in the EU, €40–60 is a sensible range; use your own HR number. Count build cost in hours too, even when the person building it is you — especially when it is you.
Three worked examples
A. Quotation analyzer. A buyer spends six hours a week comparing supplier quotes in different formats and languages. An AI quotation analyzer reads the PDFs and fills a comparison table; the buyer checks it in one hour. Five hours a week over 46 working weeks is 230 hours, or €10,350 at €45. Because every line item is now compared, assume it also catches €4,000 a year in quotes that hid freight or packaging in a separate line. Build: three days, €1,080. Running: AI credits and an hour a month of maintenance, about €1,200. Net year one €12,070, ROI about 1,100%, payback in roughly a month.
B. Contract renewal monitoring. The value here is not hours; it is one avoided auto-renewal. Contract Nemesis reads every new contract and posts renewal dates and price clauses to Teams. If it prevents a single unwanted €20,000 renewal with a 5% uplift, that is €1,000 avoided — plus the leverage of renegotiating on your timeline instead of the supplier's. Build: half a day. This is the use case where the business case is one sentence: it pays for itself the first time it fires.
C. Spend classification. Classifying 2,000 suppliers into categories took an analyst about a week each quarter: 160 hours a year, €7,200. A language model with your category tree in the prompt does a first pass in minutes; a human reviews the uncertain 10%. Net time saved is roughly 130 hours, €5,850. The bigger value is indirect — clean categories are the precondition for every other savings initiative — so present it as an enabler rather than a standalone return.
| Use case | Main value | Build | Payback |
|---|---|---|---|
| Quotation analyzer | Time + price caught | ~3 days | ~1 month |
| Contract renewal monitoring | Risk avoided | ~½ day | First avoided renewal |
| Spend classification | Time, enables savings work | ~2 days | ~6 weeks |
Estimating the AI running cost
People overestimate this. Language models are priced per token, roughly per word fragment, and procurement documents are short. The method: pages per document × tokens per page (about 500–700 for a dense contract page) × documents per month × your model's price per token, then double it for the prompt and the answer. For most single-category use cases this lands in the tens of euros per month, not hundreds. If your company uses AI Builder, the same logic applies in credits. Get the actual price from whatever your IT has approved — it changes too often to print here.
The procurement savings formulas, for completeness
When the AI use case recovers price rather than time, you also need the classic savings maths. Price savings are (baseline unit price − new unit price) × volume, with forecast volume for negotiated savings and invoiced volume for realized savings. Savings percentage divides by baseline spend, never by the new spend. Cost avoidance — an increase you prevented — is (requested price − agreed price) × volume and goes in its own column, because spend did not go down. And procurement ROI overall is realized savings divided by the cost of the procurement function, usually shown as a ratio such as 4:1. AI savings tracking shows how to get the realized number without a spreadsheet marathon.
Mistakes that kill AI business cases
Overstating hours saved is the classic one — measure, don't estimate. Counting the model's first draft as finished work is the second: always include the human review time. Forgetting maintenance is the third; flows break when a supplier changes their invoice layout. And presenting risk avoidance as a guaranteed return invites the first sceptical question in the room. Put the certain value in the headline and the risk value in the footnote.
FAQ
How do you calculate the ROI of AI in procurement?
Add time value (hours saved × loaded hourly cost) and price value (savings or leakage caught), subtract running cost and build cost, and divide by build cost. Payback in months is build cost divided by monthly net value.
How much does running an AI use case in procurement cost?
For one category, usually tens of euros a month in AI credits or tokens plus a few hours of maintenance. Estimate pages × tokens per page × documents per month × price per token, then double it.
What is the formula for procurement savings?
(Baseline unit price − new unit price) × volume. Use forecast volume for negotiated savings and invoiced volume for realized savings.
Which AI use case in procurement pays back fastest?
Contract renewal monitoring and quotation comparison usually pay back within weeks, because they are quick to build and either save hours every week or prevent a single expensive renewal.