More AI, Smaller Bills? Ramp's Economist Tests the Link Between Adoption and Revenue

1. Guest Background

Ara Karazian, introduced as Ramp's lead economist, studies AI adoption through business purchasing behavior. His relevant experience in this episode is the analysis of enterprise spending: which providers businesses pay, how intensively they spend, and how those patterns change across different customer groups. In this installment of ARK's The Brainstorm, he discusses that work with Brett and other participants who challenge the implications he draws from it.

Karazian says Ramp's research uses a dataset covering more than 80,000 businesses. Over roughly two years of tracking, he reports, AI spending has grown faster than any other expenditure category the company follows. That gives him a useful vantage point on commercial adoption, particularly where enthusiasm becomes an actual payment.

The vantage point has a defined boundary. Karazian acknowledges that Ramp skews toward technology-forward companies likely to adopt AI early and intensively. Their purchasing behavior can illuminate the leading edge without representing the entire economy. Nor do spending records, by themselves, measure the productivity of the work purchased. The supplied background establishes his role and research focus; it does not establish a broader biography or justify treating every interpretation he offers as a directly measured result.

2. What the Episode Covers

The episode examines a connection that can disappear inside an optimistic AI forecast: the passage from better capabilities to more usage, and from more usage to supplier revenue. Karazian's starting point is that AI spending continues growing in Ramp's dataset. His concern is the structure underneath that growth. He says approximately 1% of businesses account for 80% of spending, making the behavior of the most intensive customers unusually consequential. That is his description of the platform sample, not an independently established percentage for the whole enterprise market.

He therefore favors a distributional view: follow the median, top 10%, and top 1% of spenders. The denominator matters. The median discussed here is a business already spending on AI, rather than the median of all businesses. Karazian reports that spending at this level rose from about $12 to $15 per employee per month over a month. The top 10% also continued increasing expenditure. At the top 1%, however, he describes spending of approximately $7,000–$10,000 per employee per month and an emerging slowdown in growth.

The observation has a short time horizon. Karazian identifies September and early October relative to July and the earlier period of rapid expansion. Brett questions whether that interval can support a conclusion about lasting demand. The disagreement is therefore less about whether a recent deceleration appears in the data than about what mechanism it reveals and how far it can be extrapolated.

Competition adds another dimension. Karazian invokes Anthropic's rise in Ramp's business spending data as evidence that market positions can change quickly. But when pressed about Meta, he explicitly says Ramp does not yet see significant share. An observed shift toward one supplier and an expectation that another entrant will become important are different kinds of claims.

The conversation also asks where demand beyond technology companies might emerge. Brett proposes cybersecurity through a hypothetical attack scenario affecting ordinary firms. It is a possible demand mechanism, not a measured market opportunity. Meanwhile, Karazian notes that the category being forecast is itself unstable: as software acquires AI functionality, the boundary around AI spending changes. The episode consequently tests both the growth story and the measurements used to tell it.

3. Core Views: Reasoning, Examples, and Limits

Karazian's strongest argument concerns the conversion of usage into revenue. He says token volumes continued rising during the period in question while AI prices fell by almost half. Under those conditions, substantially more consumption can coexist with a flat bill. Enterprise discounts and improving standard or light-tier models introduce additional ways for customers to buy useful capability more cheaply. These are reasons to question a fixed relationship between usage and spending, rather than evidence that customers have abandoned AI. The reported price decline is also specific to his account of the period; it is not a guaranteed future rate of improvement.

Ramp itself supplies a concrete illustration of spending discipline. Karazian describes the company as both profitable and a top-1% AI spender. He says mature adopters can restrict frontier-model use on automated tasks, improve defaults, and route work to appropriate models. The mechanism is straightforward: deeper adoption can produce better knowledge of where expensive capability is necessary. A customer can become more dependent on AI while becoming less willing to pay frontier prices for every operation. However, the interview does not isolate the savings from each measure or demonstrate equivalent output quality through controlled comparisons.

Brett's objection concerns what follows from that mechanism. He sees aggressive adopters experimenting with business acceleration while more typical enterprises proceed cautiously. Even a superior product must overcome established practices, so commercial diffusion can lag technical progress. Another participant offers an optimistic reading of the same slowdown: capabilities may be advancing faster than leading customers can absorb them. That interpretation is plausible within the discussion, but it is not established as the cause. The interview cannot settle whether later demand will outweigh current efficiency gains.

Long-term contracts reveal a related tension between customer and supplier objectives. Discounts help customers control costs. Brett interprets suppliers' pursuit of longer commitments as an attempt to make revenue more predictable against large infrastructure obligations. That explanation connects pricing strategy with financing concerns, but it is not an examination of contract terms or cash-flow coverage. More predictable revenue does not, on the evidence supplied here, establish that the infrastructure commitments are affordable.

Revenue quality raises a separate concern. Karazian says the largest spenders on Ramp are predominantly venture-backed technology companies. If financing or their own customer growth weakens, several buyers could cut spending together. The risk is correlation: many customer accounts can still depend on the same capital environment. The discussion also repeats another commentator's suspicion that 70–80% of AI-company revenue comes from venture-funded AI companies. The episode does not independently verify that industry-wide proportion. Its narrower support is Karazian's characterization of Ramp's highest-spending cohort. His description of Ramp as profitable also prevents a blanket assumption that every intensive adopter depends on its next funding round.

The exchange over consumer visibility shows why competing explanations matter. Karazian links Anthropic's enterprise acceleration in time with product launches and consumer promotion, suggesting that familiarity can support business adoption. Brett emphasizes a different purchasing mechanism: enterprises want confidence, customization, and fast delivery. He invokes OpenAI and Google's consumer positions to question a simple consumer-to-enterprise progression. The episode offers hypotheses, not causal identification. Coinciding publicity and procurement cannot reveal which factor persuaded a buyer without additional evidence about the decision.

Finally, the participants disagree about the value available outside software work. Karazian argues that coding agents' productivity benefits for relevant technology roles cannot simply be assigned to the median business. He believes nontechnology firms may obtain substantial returns at modest spending levels. Brett counters from his own intensive use that useful business deliverables still require considerable supervision, with stronger models and greater spending affecting how much autonomy is practical. Neither account resolves the question across industries: the first supplies no measured customer ROI for those examples, and the second is personal experience rather than a representative comparison. Their disagreement makes output quality and supervision costs central to the economics.

4. Learning and Application

For a business buyer, the episode suggests separating four records that are often collapsed into one: consumption, accepted task output, money paid, and business benefit. This is an application of the discussion rather than a measurement system demonstrated in it. Karazian's account of rising token volumes and falling prices explains why the separation helps. A stable bill may purchase more capability, while a larger number of calls may still generate rework. Where feasible, compare similar tasks under similar acceptance standards before attributing an apparent improvement to the model or its price.

Cost controls can begin with the deployment choices Karazian describes at Ramp. A team could select repeatable work with inspectable results, compare cheaper models with frontier alternatives, and establish escalation rules when the lower-cost option fails. The relevant comparison is cost per acceptable deliverable, including review and correction. A low token price is insufficient if it transfers substantial work back to employees. Brett's experience also suggests a boundary: exploratory assignments or difficult-to-verify core outputs may need stronger models and closer supervision. The interview does not establish a routing policy that works for every task.

For someone assessing supplier revenue, customer concentration should be paired with shared exposure. Ask what sustains major customers' budgets, whether their businesses are growing independently, and how many rely on similar financing conditions. Karazian's account supports constructing a scenario in which funding pressure affects several customers simultaneously. It does not support predicting that those customers will necessarily cut spending together. Buyers considering discounted commitments can apply the same discipline to their own contracts by checking how the commitment fits realistic workloads and their ability to adjust.

Ordinary businesses should also resist using the largest spender's budget as an adoption target. Karazian argues that AI could become a competitive necessity while worthwhile returns remain compatible with modest expenditure. A practical response is to choose a workflow with a clear completion standard, establish its current cost, and test whether AI improves the result after supervision is included. Coding-agent success in a software role does not substitute for that local test, and Karazian's low-spend argument is not a promise of inexpensive success for every firm.

The episode's closing forecast provides a final lesson in source handling. Karazian initially discusses a five-year figure below $1,000 per employee per month, then recognizes confusion between monthly and annual units in his computer-purchase analogy. He mentions a possible annual figure and ultimately declines to give a definite level. A planning model should not preserve the first number while discarding the correction. Instead, construct explicit scenarios for workload, prices, acceptable quality, and adoption breadth, identifying each as an assumption. That retains the possibility of substantial diffusion without turning a withdrawn conversational estimate into a market-size finding or a procurement budget.

Source

• Original episode: Ramp's Economist Found Cracks In The AI Thesis | The Brainstorm 156