The Token Tape: methodology

What it measures

The Token Tape records what 60 major US-listed financial institutions say about the cost and economics of AI on their quarterly earnings calls. It measures disclosed discussion, not spending. There is no survey, no estimate, and no score.

Universe

60 US-listed companies in four segments: banks (money center, custody, super-regional, brokerage), asset managers and alternatives, insurers, and exchanges and market data. Companies exit on delisting; history is retained and replacements are documented. Non-calendar fiscal years (FactSet, Franklin Resources) are mapped to the calendar quarter in which the covered period ends.

Source

Public earnings-call transcripts, speaker-attributed. Quotes published on the page are verified verbatim against the transcript text before publication; load-bearing quotes are additionally checked against the call audio.

Measurement

• Candidate sentences are found by a keyword scan (token, inference, LLM, generative AI, agentic, copilot, named model providers, machine learning, AI).
• A model-based classification pass assigns each candidate to one of: AI-cost discussion; AI-capability discussion; crypto and tokenization language (excluded); boilerplate (excluded); not AI (excluded). Speaker role (management, analyst, operator) is recorded.
• Reported series: share of calls mentioning AI, share of calls with AI-cost discussion, AI-cost sentences per call, management vs analyst counts, and calls using “token” in the AI-cost sense.

Reading the series

Year-over-year comparisons are primary: fourth-quarter calls cover full-year results and run longer and more strategic than interim quarters. Current-season figures are labeled “to date” until all institutions have reported, and early reporters skew toward the largest banks.

Corrections

Factual errors are corrected on the page with a dated note.

← The Token Tape