AI to ROI

Most companies treat AI as a layer they add on top of how work already gets done. Dan Zhang, Chief Business Officer and CFO at ClickUp, argues that it is exactly backward.
In this episode, Ray sits down with Dan to unpack the "100x Org," ClickUp's framework for rebuilding the business around AI rather than sprinkling tools and tokens on top of a human-driven workflow, and why that distinction determines whether an AI initiative shows up as activity or as income statement impact.
The conversation covers:
Why ClickUp expanded the CFO's charter to own AI transformation end to end, after both a top-down mandate and a bottoms-up experimentation push failed to produce results that made it into production

The "jobs to be done" framework Dan uses to separate primary work that actually moves the business from secondary work that just generates busy AI activity, and why most companies have a work redesign problem before they have an AI problem

Dan's psychological test for AI ROI (would you pay for it with your own money) and why ARR per headcount is the right metric, but a lagging one that plays out over years, not weeks

How ClickUp instrumented daily, not monthly, visibility into AI cost and token consumption, and why over half of enterprise companies in Ray's own research have blown their AI budget by 25% or more

Why the real anxiety CFOs have about AI ROI isn't the return, it's confidence in cost management, and how a governance layer turns that anxiety into a guardrail instead of a monthly surprise

Dan's rapid-fire advice on who should own AI ROI measurement, the two-to-three variables every CFO needs in place, and how early and mid-career professionals protect their relevance by owning the question, not just the answer

If your organization has AI activity but can't yet point to AI impact, or your CFO isn't sure whether AI spend is a competitive advantage or a leak, this episode gives you the operating model and the financial discipline to tell the difference.
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Token prices have fallen 98 percent since GPT-4, but enterprise AI bills are up 320 percent. In this week's Big Story episode, Ray Rike and Peter Buchanan trace where that money is actually coming from, and the answer is not the software budget.
Drawing on Gartner, Oxford Economics, Zylo, Challenger Gray and Christmas, and Goldman Sachs data, the two lay out why labor, not IT, is becoming the primary funding source for AI at scale, and why almost no company has the measurement infrastructure to manage it.
The price paradox. Per token costs have collapsed, but usage has grown faster than costs have fallen. Ray walks through the math behind average enterprise AI budgets rising from $1.2 million to $7 million in two years.

Three cautionary tales. Uber consumed its entire annual Claude Code budget in under four months, Microsoft revoked thousands of Claude Code licenses over cost, and one unnamed enterprise ran up a $500 million bill in a single month. Ray and Peter break down why each was a governance failure rather than a technology failure.

Only two budget pools are big enough. The IT and software budget represents just 3 to 4 percent of revenue, while labor represents 25 to 40 percent depending on industry. Ray makes the case that labor is the only pool large enough to absorb the AI spending trajectory Gartner and Oxford Economics are projecting.

The attrition lever. Ray and Peter unpack how not backfilling open roles has quietly become the primary way enterprises are funding AI investment, supported by data showing over 113,000 tech layoffs in 2026 with 48 percent explicitly attributed to AI.

Revenue per FTE as the tell. Ray shares benchmark data showing SaaS company revenue per employee up 25 to 35 percent over the last twelve quarters, and explains why this metric will be the clearest signal of whether the AI budget transfer is actually working.

Five metrics every CFO needs now. The episode closes with a practical starting list: AI spend as a percent of revenue, AI spend per employee, inference spend as a percent of opex, inference cost as a percent of COGS for AI enabled products, and revenue per FTE tracked against labor cost and agent cost as a percent of OPEX.

AI to ROI is always looking for guests with real-world examples of measuring AI budget impact. Reach out Ray on LinkedIn (@rayrike).
Subscribe to AI to ROI at ai2roi.substack.com for the full June 9th edition, and leave a review wherever you listen
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Professional services firms have billed by the hour for over 200 years. That model is now under direct attack from a new category of company: the AI Native Services firm. Ray Rike and Peter Buchanan break down exactly what makes these companies structurally different from traditional professional services firms and AI-augmented incumbents, profile four companies proving the model at scale, and lay out the six critical success factors that will separate the winners from the well-funded failures.
Episode Highlights:
Defining the category. Drawing on the Emergence Capital AI Native Services Playbook, Ray and Peter establish a clear three-part taxonomy: AI Native Services companies (AI does 80 to 90% of the work, a licensed human reviews and is accountable for the outcome, and the client pays for results), AI Augmented Services companies (humans still do most of the work with AI as a productivity layer), and traditional SaaS tools (the customer's team does the work, and the vendor takes no accountability for the outcome). The distinction matters enormously for enterprise buyers evaluating contracts and liability.

How the operating model actually works. The AI Native Services delivery model runs in four phases: client intake and data ingestion, AI-driven execution of the primary service work, licensed human review and approval, and outcome delivery back to the client. Critically, that fourth phase is where the model compounds, because every accepted output becomes training data that makes the system smarter and harder to displace over time.

Four companies are proving the model. Ray and Peter profile four AI Native Services companies at different stages of scale, each dominating a regulated vertical wedge. Top AI-Native Service companies covered include: 1) Field Guide is automating audit workflows for nearly half of the top 100 US accounting firms, including KPMG and RSM, and recently raised a $75 million Series C at a $700 million valuation; 2) Even Up has built a proprietary PI AI model trained on hundreds of thousands of personal injury cases, processing 10,000 cases per week for over 2,000 law firm clients, following a $385 million funding round; 3) A-Bridge converts physician-patient conversations into structured clinical notes integrated with Epic and other EMR platforms, serving over 150 enterprise health systems including Kaiser Permanente, Mayo Clinic, and Johns Hopkins, with $100 million in ARR and a $5.3 billion valuation; 4) Harper is a licensed commercial insurance broker, not a software tool, that processes applications across 160+ carriers simultaneously and delivers final coverage in 24 to 48 hours versus the industry standard of five to seven days.

Six critical success factors. The hosts lay out what separates durable AI Native Services companies from those that will stall: genuine domain expertise on day one, a proprietary data flywheel that compounds with every case resolved, a clear migration path from labor-based to outcome-based pricing, honest gross margin accounting that properly classifies LLM inference and human labor as cost of goods sold, narrow vertical focus on a specific wedge rather than broad horizontal expansion, and distribution through regulated industry incumbents who provide both credibility and enterprise access.

Gross margin as the early warning signal. If gross margins are declining as an AI Native Services company scales, that is a signal the human-in-the-loop is becoming the bottleneck rather than the leverage point. The financial goal is to continuously increase gross margins as AI does more of the work, moving from a 35 to 45% gross margin profile toward 50 to 60% over time.

What enterprise buyers should ask. Ray closes with a direct call to action for executive buyers: if an AI vendor is pricing by the seat, by the partial FTE, or by the hour, push hard on how much of that is human supervision of AI versus a truly AI-native delivery model. You are no longer contracting a resource to do the work. You are contracting an organization to deliver the outcome you need.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Token spend is exploding across the enterprise, but the value it creates remains largely invisible on corporate dashboards. In this week's Big Story episode, Ray Rike and Peter Buchanan unpack why AI investment and ROI visibility are moving in opposite directions, and what enterprises need to do about it.
Drawing on Ramp data, Exponential View, Semianalysis, and Ray's recent conversation with Russ Frayden, CEO of Lariden, the two dig into the measurement infrastructure gap that is turning individual AI productivity gains into an unmeasured expense line.
The productivity paradox. Individual output is up across engineering, sales, and research functions, but those gains are not translating into company-level financial impact. Ray connects this to Parkinson's Law and explains why more productive workers do not automatically produce more profitable companies.

AI dark output. Peter introduces the concept from Semianalysis: real economic value created by AI that never registers on a P&L, using the example of a legal document that drops from $400 to $5 to produce, where the savings disappear while the token expense shows up in plain sight.

The cost to compensation shift. Ray walks through why token spend approaching 50 to 100 percent of engineering compensation changes the entire calculus for measurement, contrasted against IT's historical 3.5 to 6 percent share of revenue.

Case studies in good and bad. The episode breaks down three real examples: Uber's Claude Code rollout that ran out of budget without measurable output gains, Lowe's cross-functional agent deployment that built proper context tracking, and Petrobras's narrow tax compliance pilot that identified $120 million in savings and is scaling toward $1 billion.

A four-stage framework for real measurement. Ray and Peter lay out the progression from cost visibility to utilization, proficiency, and business impact, and explain why almost every enterprise is stuck at stage one.

Six tactical takeaways. The episode closes with concrete actions: measure outcomes not activity, design for the middle 70 percent of users rather than power users, give CFOs real budget ownership, start narrow and expand from proof points, redesign decision rights as AI scales, and treat organizational role changes as a planned program rather than a side effect.

Subscribe to the AI to ROI newsletter at ai2roi.substack.com for the full breakdown, and leave a review wherever you listen
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

The April 2026 announcement that SpaceX may acquire Cursor for $60 billion, or alternatively pay $10 billion for a compute partnership, stopped the enterprise tech world in its tracks. A four-year-old company founded by four MIT students with $2.7 billion in annualized revenue but nearly $900 million in losses on $700 million in actual revenue. This deal is not primarily a valuation story. It is a signal and a cautionary tale about the economics of the AI coding tool market.
In this Big Story edition, Ray Rike and Peter Buchanan break down what is really happening in the AI coding wars, why Cursor ended up at SpaceX's door, and where this market goes from here.
Key topics covered in this episode:
From copilot to autonomous agent: how the AI coding market structurally shifted. Four years ago, AI coding tools suggested your next line of code. Today, they read entire codebases, plan multi-step tasks, edit files across a project, run tests, and submit pull requests with minimal human direction. Claude Code reached $1 billion in annualized revenue six months after launch, the fastest of any enterprise software product in history, and crossed $2.5 billion by February 2026. Meanwhile, 90% of enterprise developers now use at least one AI coding tool, and nearly half of all GitHub code is AI-generated or AI-assisted.

The productivity gains are real but uneven, and the risks are underappreciated. JPMorgan deployed AI coding agents to 40,000 engineers and reported 10 to 20% productivity gains in code creation and conversion, along with a 70% increase in code deployments. But CodeRabbit's research found 1.7 times as many defects in AI-authored pull requests as in human-authored code. Meta's brief "token maxing" leaderboard experiment, designed to spotlight power users, had to be taken down within two weeks after producing high token consumption and limited usable code. Senior developers are shifting toward architecture and review roles while junior developer pipelines are shrinking, even as total software developer job postings are up 5 to 10% year over year.

A tour of the seven major players and where the structural tension lives. Ray and Peter profile Anthropic Claude Code, GitHub Copilot, Cursor, OpenAI Codex, Google Gemini Code Assist, Replit Agent, Lovable, and Cognition's Devin across revenue, differentiation, and risk. The common thread: most point-solution coding agents run on Anthropic or OpenAI models, and those same model companies have now launched their own competing coding products. The Oracle database-to-applications parallel is not subtle.

Why the Cursor-SpaceX deal happened and what it actually reveals. Cursor had $2.7 billion in annualized revenue, negative 23% gross margins, and was losing money faster than it was growing. Even with a $2 billion funding round in process from Andreessen Horowitz, Thrive Capital, NVIDIA, and Battery Ventures, Cursor's leadership concluded they would need to raise billions more by year-end to fund compute costs. SpaceX's acquisition offer, or the $10 billion partnership payment that Ray reads as a very generous breakup fee, solved that problem while giving XAI a revenue base three times its current size ahead of a $1.75 trillion IPO valuation push.

The Chinese open-source threat and three scenarios for where this market goes. Kimi, DeepSeek, and Qwen models are improving rapidly and are significantly cheaper. They are, as Peter puts it, lurkers haunting every company on the list. Ray and Peter then lay out three scenarios: model makers consolidate, and IDE players get marginalized; a durable multi-tool ecosystem persists because different tools serve different workflow stages and buyer profiles; or a compute-native player builds a fully autonomous coding agent that eliminates the need for an IDE entirely. Claude Code already resolves 64.3% of real-world GitHub issues, and full autonomy for defined-scope tasks may be 18 to 36 months away.

The AI coding market is projected to reach $49-$50 billion by 2030, at a 38% CAGR. The speed gains are real. So are the defect rates, the governance gaps, the model-dependency risks, and the token-budget surprises landing on CFO desks.
If you are a CIO, CFO, engineering leader, or investor trying to make sense of who wins and what it costs, this is the episode to start with.

Read the full story at ai2roi.substack.com.
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.