What an answer costs: how K-Lake reduces the tokens an assistant spends

The K-Lake engineering team measured how many tokens an AI assistant consumed answering the same three questions over the same nine annual and quarterly filings, using four retrieval approaches. K-Lake put 21,504 tokens into the conversation. A carefully tuned command-line approach put in 50,666, a vector-chunk store 45,000, and naive file searching just over one million. The test ran on 31 August 2026.

Why this matters to a regulated buyer

AI assistants are priced by the token, and most send the full conversation history again with each new question. An assistant that reads whole documents to find one paragraph pays for that reading again and again. For a CFO, that makes AI spend hard to forecast. For a CISO, every extra token is more of your content sitting in a conversation with an external model. Reducing what enters the conversation addresses both concerns at once.

What we tested

The corpus was nine public filings from one listed company, about 6.3 million characters of extracted text, drawn from a wider set of 40 filings indexed in K-Lake. The three questions covered a broad concept (the company's risk factors), an exact fact (who held a named executive role) and a section-level question (the legal and regulatory risks). One agent ran every approach. Tokens were counted as characters entering the conversation, divided by four.

Where the saving comes from

K-Lake returned about 13.6 million characters across the session, far more than any other approach. 99.3% of it stayed on disk. The assistant saw short previews and read only the passages it chose, which came to under 1% of what K-Lake found. It searched widely and read narrowly.

What it means for a budget

At list prices of $2.50 per million input tokens and $10 per million output tokens, a three-question session cost about $0.10 with K-Lake, $0.16 with the vector store and $0.18 with careful command-line search. At 10,000 ten-question sessions a month, that is roughly $4,300 a month with K-Lake against $9,100 and $9,800 for the alternatives. The ratios hold whatever the price per token.

Limits we want you to know

This is a small test: nine filings and three questions, with one model in one harness. Larger estates have not yet been measured. The vector store used keyword-based similarity rather than dense embeddings, which probably flattered it, so its 2.1x gap is a floor rather than an estimate.

The on-disk saving depends on an assistant harness that writes large results to disk, and others will show a smaller benefit. Command-line tools remain faster for pure pattern matching, such as counting the files that mention a term.

The honest headline

Against a well-built alternative, expect K-Lake to put less than half as much into the conversation. The 47x figure applies to assistants left to search files on their own, which is how many rollouts start. The figure that matters for your organisation should come from a proof of concept on your own documents, and we will measure it with you.