Money

What it really costs, and where the money goes

Updated November 2026 · About 5 minutes to read · Watch the 3-minute video

The short answer

Each answer costs about a cent or two in AI fees. That's rarely where the money goes.

Most of it goes on people: getting your documents ready, connecting the system, checking the answers and keeping the documents current. Budget for those first.

Think of it as a company car

The price on the windshield is only the start. The real cost is the car plus insurance, servicing and the driver's time.

The AI is the fuel. The servicing is your documents, and servicing costs more.

Where the picture stops working: fuel is usually a big part of running a car. Here the fuel, the AI bill, is often the smallest line.

Most of the cost is people and documents, not the AI

A made-up company of 800 people builds a help desk that answers HR policy questions. Here is a rough first year. The numbers are illustrations, not data.

$40k

Getting the documents ready
Finding the current versions, retiring old ones, and naming an owner for each.
$30k

Building and connecting it
Setting it up, linking it to where the documents live, and respecting who may see what.
$25k

Checking and upkeep
An expert reviewing answers in the first months, and documents kept current after that.
<$1k

The AI bill
About 24,000 answers in the year, at a cent or two each. Too small to draw to scale.

Gartner, the research firm, predicts that through 2026 companies will abandon 60% of AI projects whose data isn't ready (Gartner, February 2025). The top bar is the one people leave out.

What one answer costs

AI use is billed in tokens. A token is roughly three-quarters of a word. Each answer reads the passages it found, then writes a reply.

Reads about 5,000 tokensAround 3,700 words of passages: about 1 cent.

Writes about 400 tokensAround 300 words of answer: under half a cent.

About 1½ centsPer answer, at a mid-priced model's list price.

Our arithmetic, using Anthropic's list price for Claude Sonnet 5.5: $2 per million tokens read and $10 per million written (Anthropic pricing, November 2026). Other vendors' mid-priced models are in the same range, and prices keep falling.

Three ways to pay

A price per person, per month

A subscription such as Microsoft 365 Copilot, listed at $30 per person a month on top of Microsoft 365 (Microsoft).

Best for: getting started fast, with documents already in that vendor's tools.

Gives up: you pay for everyone, asked or not. For 800 people that's about $288,000 a year.

A price per answer

Built for you on an AI vendor's model, paying for what's used.

Best for: one job done well, where you choose the documents and the checks.

Gives up: someone has to build and run it, and that costs more than the answers.

A project fee for outside help

A firm sets it up with your team, then hands it over or keeps running it.

Best for: a first project, when nobody inside has done one before.

Gives up: a bigger bill up front, and you need to keep the know-how afterwards.

What goes wrong

Budgeting only the AI bill. The cents per answer get approved; the document work comes as a surprise.
No money for upkeep. The project ends, nobody owns the documents, and the answers go stale.
Seats for everyone. A license for all 800 people, and 80 who use it.
No cap on usage. AI that takes steps on its own can retry and loop. Set a monthly limit.

Questions to ask your team or vendor

What does the first year cost in total, not just the license or the AI fees?
Who gets our documents ready, and how long will it take?
Who keeps the documents current after launch, and what does that cost?
If we pay per person, how many people will really use it?
If we pay per answer, what's the monthly limit, and who is told when we near it?

Other names you may hear

Tokens: The units AI use is billed by, roughly three-quarters of a word each.
Per-seat license: A price per person, per month, whether they use it or not.
Usage-based pricing: Paying for what's used, usually per million tokens.
TCO: Total cost of ownership: everything it costs over its life, not just the purchase price.
API: Application programming interface: the connection your software uses to send questions to an AI vendor's model.
Inference: The AI producing an answer. "Inference cost" is the AI bill.