The Ramp AI Index now covers 70,000 businesses. That figure arrived in a Crypto Briefing brief carrying no methodology, no model list, no sample breakdown, no update cadence. The two operative claims are "token spend" and "model-level usage." I have spent the last three months auditing an AI agent's decision logic against historical bear-market data, and the first rule any serious auditor internalizes is this: a coverage number is not a measurement. It is a marketing surface. Seventy thousand is impressive phrasing. It tells me nothing about attribution.
Understand what Ramp is before you trust its index. Ramp, by reasonable inference, runs corporate spend management and card infrastructure. Its raw material is card transactions, accounts payable lines, and reimbursement records. From those primitives, the company claims to derive model-level AI usage across a vast enterprise base. No published attribution method. No vendor list. No weighting. No confidence interval. Ledgers do not lie, only the auditors do — and here, nobody has signed the audit.
The timing matters. Enterprise AI measurement has, until now, run almost entirely on surveys. Menlo Ventures publishes adoption studies. a16z runs founder surveys. Model vendors disclose their own growth selectively, and only when the numbers flatter them. Every one of these sources shares a structural defect: respondents self-report. A procurement director who spent six figures on AI tooling in a flat quarter has every incentive to describe the spend optimistically, and a vendor has every incentive to count seats provisioned rather than seats used.
Against that backdrop, a spend-derived index is genuinely interesting. Payment rails capture intent that surveys miss. If you can see the actual transaction, you do not need to trust the respondent. That is the theory, and it is a good theory. It is also exactly the theory that collapses on contact with merchant descriptor data.
Here is why the mechanism deserves scrutiny rather than applause. An AI dollar does not arrive at the payment layer with a clean label attached. When a company buys OpenAI API access, that is a direct vendor charge. When it buys the same model through AWS Bedrock or Azure OpenAI, the charge reads as a cloud line item, indistinguishable at the descriptor level from S3 storage or EC2 compute. When it buys ChatGPT Team seats, that is a subscription. When it runs Llama or Mistral weights on its own servers, there is no external charge at all — only a GPU bill, which looks identical to any other compute spend. So the index claims model-level granularity from data that is, at the raw layer, hostile to that granularity.
Map the attribution problem precisely, because this is where the whole product either works or does not.
Layer one: direct API spend. Billed by OpenAI, Anthropic, Google, Mistral, Cohere. Cleanly attributable. This is the only layer a card descriptor can resolve to a specific model vendor with confidence.
Layer two: cloud marketplace resale. Bedrock, Vertex, Azure. The same inference dollar is now buried inside a hyperscaler bill. To recover model attribution, Ramp needs either a deep integration with each cloud's billing API — which exposes line-item detail most enterprises do not export to third parties — or a probabilistic model that infers the provider from spend magnitude and timing. Probability is not attribution. It is a guess wearing a decimal point.
Layer three: self-hosted inference. Open weights, on-premise, or reserved GPU capacity. Zero external token spend. The index cannot see this layer at all, yet it is the fastest-growing segment of production AI for privacy-sensitive and cost-sensitive firms. Any index built on external transaction data systematically undercounts open-weight adoption, and it undercounts it precisely among the most sophisticated buyers.
Layer four: pass-through and internal reallocation. Platform teams buy API credits centrally and re-bill internal units. The transaction is real, but the user attribution is fictional. One charge, fifty internal consumers, one recorded adopter.
I built a smaller version of this machine in January 2024, tracking the spread between spot Bitcoin ETF pricing and the Coinbase Premium Index with a Python script. That was a two-variable problem with clean inputs. Attributing a messy corporate card charge to a specific model family is a hundred-variable problem with dirty inputs. The engineering effort is not in the AI. It is in merchant normalization, entity resolution, and a mapping table that decays every quarter as vendors rename and re-bundle products. The hard part of model-level usage is data plumbing, not machine learning.
Now the sample itself. Ramp's 70,000 businesses are not a random draw from the economy. They are Ramp clients — firms that chose a modern spend platform, which skews toward venture-backed, software-native, mid-market companies. A Fortune 100 manufacturer running proprietary models on dedicated GPUs may be entirely invisible. A bootstrapped logistics firm still on spreadsheets is invisible. The index measures Ramp's customer base and then narrates it as enterprise AI adoption. Those are two different populations. Sample bias is the tax you pay for convenient data.
There is also a compliance surface nobody in the brief touched. Aggregated spend across 70,000 firms can reveal supplier choices, budget scale, technology stack, and strategic direction. That is sensitive commercial intelligence. Under GDPR and CCPA-adjacent regimes, the question of whether an enterprise can opt out of a third party monetizing its transaction metadata is not academic. If a single company can be re-identified from an industry-and-size slice, the index becomes a competitive-intelligence leak wearing a statistics label.
There is also a semantic trap worth naming. In crypto, token spend means something. It means on-chain value moving through wallets, visible on a public ledger, auditable by anyone with an RPC endpoint. In the Ramp index, token spend means LLM inference units metered by a private vendor. Same words, opposite epistemics. One is verifiable by design. The other is verifiable only by trusting the aggregator. Liquidity is the only truth in a fragmented chain. In AI spend data, the only truth is a disclosed methodology — and this brief disclosed none.
Here is the counter-intuitive part, and it cuts against my own skepticism. The index does not need to be accurate to be valuable. It needs to be first and cited. Once analysts and journalists anchor on a single spend-derived number, that number becomes the reference class, and every subsequent estimate calibrates against it. The product's real output is not measurement. It is narrative control over who is winning enterprise AI.
That is why the missing vendor list is not an oversight. A published model-share table invites dispute from every vendor it ranks low. A vague index invites citation without accountability. Efficiency demands the elimination of sentiment — but it also rewards the elimination of falsifiable claims when you are selling positioning. The vendors themselves will not volunteer validation either, because third-party share data can only ever confirm or embarrass, never flatter.
Watch for three disclosures, and grade the index on each. One: a methodology document naming the attribution rule for cloud-resold and self-hosted inference. Two: a vendor list and a stated aggregation threshold. Three: historical revision — does the index restate past months when vendors change billing structures? If all three appear, the product becomes a real benchmark. If only the headline number survives into next quarter, treat it as a press release with a chart. The algorithm executes, but the human decides what counts as evidence.