AI Adoption Costs · Token-Based Pricing
AI Cost and Token Pricing
What are AI adoption costs?
- Two axes: per-user subscription and usage-based pricing
- Both input and output are billed, and output costs more
- Training, administration, and verification effort are easily left out of the budget
The question about AI cost shifts from 'How much per month?' to 'Who uses what, and how?' Even within one team, a person who logs in to polish reports and a workflow wired into a system that processes hundreds of items a day follow completely different cost curves. The first stays fixed unless headcount grows; with the second, the invoice grows with the workload. A budget that doesn't separate usage types first will always come back from finance review. The first table in your approval request isn't a price list but a table of users and items processed, by usage type.
Tokens are the pieces a model splits text into for processing. Korean tends to break into smaller pieces than English, so the same amount of text produces more tokens. Rather than studying the price list first, it is faster to run ten real work documents through and measure input and output tokens per case. Input and output are priced differently, and output is usually more expensive. What drives cost isn't the choice of model but the attachments you push in whole every time. Attaching a 40-page manual to every request gets you the same answer at several times the cost, and trimming it to the clauses you need cuts cost in proportion. Keep in mind, too, that token fees are only part of total cost.
The items most often missing from approval requests lie outside the tool price. To make the numbers add up, include initial training, writing internal guidelines, the monthly effort of an operations owner who manages accounts and permissions, the time people spend verifying results, and the cost of attempts that were tried and dropped. Estimate benefits in the opposite direction, by discounting them. Convert time saved into labor cost, but conservatively, and write down what you will measure, when, and how. Prices and plans change quarterly, so leave the formula rather than specific amounts in the document.
The most common mistake in practice is the claim that 'token prices are less than a cup of coffee, so there's nothing to worry about.' That holds when one person is testing. But once nightly batches run, error retries kick in, and long search results get stuffed into prompts, costs jump by an order of magnitude within a month. Projects have actually been halted by sticker shock at the first invoice after starting without usage caps or alerts. The opposite misconception exists too. Buying accounts for every employee at once looks fair, but open the login records two months later and half are unused, with the money already locked into an annual contract. It is safer to open accounts only for people who will use them, and set a cap on usage.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayAs language model APIs became commercial, per-token billing became the standard. Enterprise subscription plans followed, and the billing structure settled into two tracks: one for on-screen use and one for system integration.
Consider a hypothetical case. Company A, an auto parts maker, has 520 employees and annual revenue of KRW 130 billion (≈ US$97,261,700). Its quality assurance team writes replies to overseas client claims at 40 minutes each, 1,400 a month. The corporate planning team had submitted AI adoption proposals twice and been rejected both times for 'no cost justification.'
- When Deputy Manager Park on the Business Planning team dug into usage across departments over two weeks, she found 38 people actually working in the tool. The other 12 said they had 'received an account but never use it.'
- Manager Lee on the IT team ran 100 claim replies through the actual API and measured 6,200 input tokens and 900 output tokens per case. Multiplying by 1,400 cases a month gave the basis for the usage budget.
- Measuring also revealed that every reply was being sent with the entire 40-page quality manual attached. Once only the relevant clauses were included, input tokens per case dropped from 6,200 to 1,800.
- Because Kim, the head of Finance, said, 'I can't submit an approval request that lists only tool costs,' we added separate lines for two days of initial training, 16 hours a month for the operations owner, and staff time spent verifying replies.
- They lowered the projected benefit to 25 minutes saved out of 40 per case, added a line promising to re-measure against actual processing logs after three months, and got approval on the third proposal.
- Three months after rollout, subscription accounts held steady at 38, actual billing came in at about 78% of the budget, and the time to write a reply fell from 40 minutes to 14 minutes per case.