How Many Tokens Will an Internal AI Assistant Use? A 5,000-Employee Estimate

Start from questions, not headcount

The request usually arrives as "we have 5,000 employees, how many tokens will the assistant use?" Headcount does not turn into tokens directly. The bill is the number of questions asked, times what each question sends to the model and gets back. The same 5,000 people can ask about 60,000 questions a month or more than 750,000, depending on how many of them use it and how often.

questions per month = employees × share active per workday × questions per active person × workdays
input tokens per question = instructions + retrieved passages + chat history + the question
monthly cost = questions × (input tokens × input price + output tokens × output price)

Every number below is an assumption for a typical knowledge assistant: one that searches company documents and answers with what it found. Swap in your own. We say where each number comes from and which ones move the bill most.

What one question sends to the model

The question an employee types is the smallest part of the request. Before the model sees it, the assistant adds its instructions, the passages its search found and the earlier turns of the conversation.

Two settings decide most of the input: how many passages the search adds to each question, and how much of the conversation you send again. Each extra 500-token passage is 9% more input. Token counts also differ between providers, because each one splits text differently; we measured how many characters a token is on real text. To get your own numbers, paste your instructions and a few typical retrieved passages into the Token Calculator.

Output is smaller but priced higher: five times the input rate on every model in the table below. If the model writes hidden reasoning before it answers, those reasoning tokens are billed as output too, and they can be larger than the visible answer. Check the usage numbers in the API response, not the length of the text you see.

Three adoption cases for 5,000 employees

The share of employees who use the assistant on a given workday is the number that is easiest to get wrong. We have no survey behind these three cases. They are planning ranges, chosen to show how far the answer moves. All three assume 21 workdays a month.

The middle case step by step: 5,000 × 40% × 6 × 21 = 252,000 questions. 252,000 × 5,500 = 1,386,000,000 input tokens and 252,000 × 400 = 100,800,000 output tokens. The high case asks 12 times as many questions as the low case and sends almost 31 times as many input tokens, because heavy users also run longer conversations and ask the search for more.

What that costs on five models

Standard prices, verified Oct 9, 2026 for Claude Sonnet 5.5; each model's date is in the table above the article. Long-context surcharges do not apply here, since every request stays far below 100,000 tokens.

Read the table both ways. On any one model, the high case costs about 30 times the low case. Between the cheapest and the most expensive model in the same case, the gap is 40 times. Both gaps are large, but only one of them is your choice on day one: you can pick a model in an afternoon, while adoption is something you find out. Settle the adoption range first, then test whether the cheaper models answer your questions well enough.

Want the same math for a customer-facing bot with retrieval? The RAG chatbot monthly cost guide works it out by monthly question volume instead of headcount.

What adds to the token bill

• Retries and regenerate clicks. Every retry is a full second request with the same input.
• Tools and agents. If the assistant searches, reads a file and then answers, one question can be three or four model calls. Count calls per question, not questions.
• Hidden reasoning. Models that think before answering bill that thinking as output tokens.
• Pasted documents. "Summarize this report" can send more tokens in one request than a full day of normal questions from the same person.
• Indexing the documents. Every page has to be turned into embeddings once, and again when it changes.

The last one is usually small. Indexing 50,000 documents of about 3,000 tokens each is 150 million tokens, which costs $3.00 once on OpenAI text-embedding-3-small at $0.02 per million tokens. Turning 252,000 questions of 100 tokens into search queries is $0.50 a month. Compare other embedding models on the embedding API pricing page.

Three settings that cut the bill

• Cache the fixed part of the prompt. The 1,200 tokens of instructions are the same on every question. With prompt caching they are billed at the cache read rate after the first request. In the middle case on Claude Sonnet 5.5 the bill goes from $3,780.00 to $3,205.44 a month, before the small charge for writing the cache when it expires.
• Retrieve fewer, better passages. Four passages instead of six saves 1,000 input tokens per question: 252 million tokens or $504.00 a month on Claude Sonnet 5.5. Only do it if answers stay as good on your real questions.
• Trim the conversation you send back. Resending the last two turns instead of the whole thread caps how much history a long chat carries.

The prompt caching savings calculator shows the caching effect for your own prompt size and cache hit rate.

Replace the guesses with a two-week pilot

• Give the assistant to 50 to 100 people from different departments, not only the team that built it.
• Log the input and output token counts the API returns with every response. OpenAI, Anthropic and Google all report them.
• Record how many pilot users ask something on a given day and how many questions each one asks.
• Work out the average and the 90th percentile of tokens per question. Budget on the 90th percentile.
• Scale to 5,000 people with the formula above, using a lower active share than the pilot showed. Pilot groups are often volunteers who were keen to try it.

Two weeks of real usage logs replace three of the four guesses in this article. The fourth, how many people will use it once it is everyone's tool, you only learn after launch, so set a monthly spend limit and review it after the first month.

What this estimate leaves out

• Fixed bills that are not priced per token: the search index or vector database, the app server, single sign-on and access control, logging and answer-quality checks.
• Per-seat plans, which price by user instead of by token. Compare them with this estimate at the usage you expect, not at full adoption.
• Enterprise discounts, committed-spend deals, taxes and Batch pricing. All prices here are standard list prices.
• Image, audio and file inputs. The token counts here are text only.
• Real usage data. Every token count in this article is a planning assumption, not a measurement from a live company deployment.

FAQ