What it costs to put AI on your company data
Three cost lines, one of them almost always forgotten, and the return calculation worth doing before starting. With the numbers that surprise both sides.
Conversations about AI cost stall because each side looks at a different line. Sellers talk about the assistant subscription. Buyers think about consultancy. The real bill has three lines, and the one that varies most is the one almost nobody mentions.
The three lines
| Line | Order of magnitude | What makes it vary |
|---|---|---|
| Assistant subscription | Tens per person, per month | How many people actually use it |
| Data platform | By volume ingested and queried | Number of sources and events, not of users |
| Connection work | Concentrated in the first month | How many sources, and how many only export spreadsheets |
The third line decides the project. Two sources with APIs is days of work. Seven sources, two of them PDF-only, is months — and the return arrives too late to sustain the decision.
Start with two sources. Not because the rest does not matter, but because the project has to prove value before consuming six months of attention.
The cost that shows up later: querying
Cheap questions have a predictable side effect: people ask a lot of them. In a real account the connected assistant read 143.6 GB in a period when the product’s interface read 102.7 GB — with 16 times more executions. That is not waste, it is usage; but a buyer should know that line exists and grows with adoption.
There is a simple technical fix: materialise what repeats daily, and keep live reads only for what is new. A fixed page costs once; an exploratory question costs each time.
The return calculation that makes sense
Do not try to estimate "revenue increase caused by AI" — that is guesswork. The three savings below are verifiable and usually enough.
- 01Report assembly hours that stop existing. Multiply by weeks in the year and the real cost of the hour.
- 02Days of budget stuck in a mistake. Finding it on day 3 instead of day 30 is worth the proportional difference of the month.
- 03The data team queue. Repeated questions that disappeared, measured in requests per week.
If those three savings together do not cover the bill, the problem is probably not price: the company does not yet have a question expensive enough to justify the effort. That is a legitimate answer.
Mistakes that add cost and bring nothing
- Ingesting everything from day one. A source nobody queries costs storage and maintenance forever.
- Buying by seat count before knowing who uses it. Real adoption is always smaller than the access list.
- Reprocessing full history on day one. History matters, rarely in week one.
- Automating what has not been checked by hand. Automating an error only multiplies it faster.