The AI Series: Adoption Ramps Up

Episode Summary

Executive Summary: The episode features Ramp economist Ara Karazian discussing Ramp’s AI Index and what its spend data says about enterprise AI adoption. The big picture: AI use is rising fast, but the gains are highly concentrated among a small set of intense adopters, mostly tech-forward firms. Adoption alone is misleading; productivity effects remain limited at the macro level, while model competition and falling prices may pressure frontier AI revenues and valuations.

Main Topics: Ramp’s AI data advantage (Priority: 5/5): Karazian explains that Ramp’s finance-platform transaction data provides a more granular view of enterprise AI spending than surveys, capturing real paid usage, vendors, receipts, and usage intensity. AI adoption is rising, but unevenly (Priority: 5/5): The AI Index shows business adoption accelerating sharply in 2025, yet the market is highly heterogeneous: many firms use only simple tools, while a small minority spend heavily and use advanced agents or multiple models. OpenAI vs. Anthropic competition (Priority: 5/5): Ramp’s data suggests Anthropic overtook OpenAI in business adoption, but OpenAI recently regained growth momentum after launching a stronger model and cutting prices, highlighting a very contestable market. Why AI is not yet showing up in productivity data (Priority: 5/5): Karazian argues that aggregate productivity gains are muted because most adopters are casual users, while actual ROI is concentrated among a small number of intense, often fast-growing firms. Jobs and hiring effects (Priority: 4/5): A joint study with Revelio Labs finds intense AI adopters grow headcount faster, while light adopters do not. Surprisingly, intense adopters also increased entry-level hiring, suggesting firms are changing hiring, not just cutting labor. Pricing pressure and investor implications (Priority: 4/5): Model prices are falling quickly, which could compress future revenues even if usage volumes keep rising. That creates uncertainty for investors betting on durable high margins in frontier AI. Policy and diffusion challenges (Priority: 4/5): Karazian argues the main policy issue is not just data centers or white-collar job loss, but how to help more firms use AI productively. He believes diffusion is uneven and may eventually affect blue-collar work through automation.

Key Arguments: Ramp’s spend data is stronger than surveys because it records actual paid adoption, vendor choice, receipts, and token usage rather than self-reported behavior. Enterprise AI adoption is much broader than the most conservative estimates but still below saturation; however, breadth of adoption should not be confused with intensity of use. The biggest economic effects come from a small group of firms that use AI intensively; most adopters are still using simple chat-style subscriptions with limited ROI. AI’s diffusion is not following classic software patterns; advanced users often expand vendor lists and use multiple models, rather than converging on a single vendor. OpenAI and Anthropic are engaged in aggressive price and feature competition; falling token prices may weaken the revenue case for frontier AI companies. Jobs effects appear positive for intense adopters, with faster headcount growth and more entry-level hiring, implying AI is changing organizational demand for labor rather than eliminating it outright. Macro productivity statistics lag because AI is still concentrated in a subset of firms and use cases, especially coding agents and tech-heavy workflows. Government policy should focus less on speculative narratives and more on enabling broader productive adoption across non-tech sectors.

Data Points: Enterprise AI adoption in Ramp sample: 53% - Share of businesses in Ramp’s data set using AI; crossed 50% a couple months before the interview. Anthropic business adoption today: 43% - Ramp’s estimate of the share of businesses using Anthropic. U.S. government estimate of AI-using firms: low 20s% - Referenced as a lower official estimate of firm-level AI use. Other survey-based estimates: 70% to 80% - Alternative U.S. business survey estimates cited by Karazian. Total AI spend estimate: $200 billion to $300 billion - Public estimates of enterprise AI spend discussed in the episode. Top 1% spend per employee per month: $7.4 thousand - Ramp data showing extreme concentration of AI spending among the heaviest users. Top 10% spend per employee per month: $650,000 - As stated in the transcript, this reflects highly concentrated spend among top adopters. Median spend per employee per month: $12 - Illustrates that the typical AI-using firm is still spending very little. Open-source model adoption: about 6% - Share of AI-buying businesses using open-source models, usually via router platforms. Headcount growth at intense adopters: 10% over two years - Firms that adopt AI intensely grew headcount by this amount after adoption in the Ramp-Revelio study. Headcount growth at light adopters: no change - Light AI adopters showed no change relative to the control group. Overall hiring growth in study: 12% over two years - Aggregate growth observed in the study among adopters, driven by intense users. Entry-level worker share change: 1% higher share after two years - Firms that adopted intensely had a slightly higher entry-level worker share after adoption. Anthropic Fable 5 adoption/share: 5% to 10% - Karazian said Anthropic’s most expensive frontier model captured only a small slice of adoption/spend. Model price decline: 30% to 40% in a month - He said effective price per million tokens at the frontier fell sharply due to competition. OpenAI price cut on 5.6 Sol: 20% - Recent price reduction mentioned as further evidence of price competition.

Pivotal Quotes: "“The privilege of my data set is that it's actual business spend data.”" — Ara Karazian: Describing why Ramp’s AI Index may be more reliable than surveys for measuring enterprise AI adoption. "“Why doesn't AI show up in productivity metrics in the US? It's because the vast majority of firms that are using AI are not getting ROI from it.”" — Ara Karazian: Explaining why broad adoption has not yet translated into measurable macro productivity gains. "“That doesn't happen in bubbles where there's just spend without question.”" — Ara Karazian: Arguing that firms’ expanding vendor lists and rising contract sizes suggest real experimentation and conscious decision-making, not pure euphoria.

Implications: AI adoption is broad but not yet transformative at the macro level. The biggest near-term winners may be intense users and highly competitive model vendors, while investors and policymakers should watch falling prices, uneven diffusion, and the need for practical adoption support.