Skip to content
a16za16z

The $700 Billion AI Productivity Problem No One's Talking About

Russ Fradin sold his first company for $300M. He’s back in the arena with Larridin, helping companies measure just how successful their AI actually is. In this episode, Russ sits down with a16z General Partner Alex Rampell to reveal why the measurement infrastructure that unlocked internet advertising's trillion-dollar boom is exactly what's missing from AI, why your most productive employees are hiding their AI usage from management, and the uncomfortable truth that companies desperately buying AI tools have no idea whether anyone's actually using them. The same playbook that built comScore into a billion-dollar measurement empire now determines which AI companies survive the coming shakeout. Timecodes: 0:00 — Introduction 1:07 — Early Career, Ad Tech, and Web 1.0 2:09 — Attribution Problems in Ad Tech & AI 3:30 — Building Measurement Infrastructure 5:49 — Software Eating Labor: Productivity Shifts 7:51 — The Challenge of Measuring AI ROI 13:54 — The Productivity Baseline Problem 17:46 — Defining and Measuring Productivity 20:27 — Goodhart’s Law & the Pitfalls of Metrics 21:41 — The Harvey Example: Usage vs. Value 24:18 — Surveys vs. Behavioral Data 27:38 — Interdepartmental Responsiveness & Real-World Metrics 30:00 — Enterprise AI Adoption: What the Data Shows 32:59 — Employee Anxiety & Training Gaps 37:31 — The Nexus Product & Safe AI Usage 41:08 — The Future of Work: Job Loss or Job Creation? 43:40 — The Competitive Advantage of AI 52:45 — The Product Marketing Problem in AI 54:00 — The Importance of Specific Use Cases SOCIALS Follow Russ Fradin on X: https://x.com/rfradin Follow Alex Rampell on X: https://x.com/arampell Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://x.com/a16z Find a16z on LinkedIn:https://www.linkedin.com/company/a16z Listen to the a16z Podcast on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Podcast on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see http://a16z.com/disclosures.

Russ FradinguestAlex Rampellhost
Dec 1, 202557mWatch on YouTube ↗

CHAPTERS

  1. 0:00 – 1:11

    Enterprise AI urgency: 18 months to lead or fall behind

    The conversation opens with the intensity executives feel about AI adoption and the uneven diffusion of AI skills inside large organizations. Russ and Alex frame the core problem: everyone is buying AI, but few can explain whether it’s working beyond spend and licenses.

    • Executives feel a shrinking window to become an AI leader
    • AI impact is obvious anecdotally (minutes vs. hours) but hard to operationalize
    • Adoption is happening unevenly—isolated power users vs. the broader workforce
    • Boards ask for AI progress; companies often only report “what we bought”
  2. 1:11 – 2:49

    From Web 1.0 ad tech to AI: the return of the attribution problem

    Alex and Russ connect AI ROI uncertainty to early internet advertising, where the biggest challenge was proving what drove outcomes. Russ recounts his early career in ad networks and measurement companies and explains why AI needs similar measurement infrastructure.

    • Ad tech’s core challenge: attribution and proving ads worked
    • Russ’s background: early ad networks and comScore-era measurement building
    • Internet ad growth required a full tooling/measurement stack (DoubleClick, Omniture, comScore, etc.)
    • AI is at a similar stage: powerful tech, missing enterprise-grade measurement and governance
  3. 2:49 – 5:47

    Why Larridin exists: measurement and governance to accelerate AI spend

    Russ explains the thesis behind founding Larridin: big budget shifts require new tooling, especially for measurement and governance. The goal is not to slow AI down, but to make adoption safe, legible, and scalable across complex enterprises.

    • Large budget transitions create opportunities for “boring but critical” infrastructure
    • Governance enables faster adoption (security, compliance, risk management)
    • Enterprises must retrain thousands of knowledge workers quickly
    • CIO/CFO questions: what did we buy, is it safe, and was it valuable
  4. 5:47 – 9:02

    Software eating labor: the macro shift from labor budgets to AI software budgets

    Alex introduces the idea that AI is shifting spend from labor to software—massive productivity potential without necessarily eliminating jobs. This raises the CFO-level question: if software spend rises dramatically, how do we know it’s efficient and value-producing?

    • AI increases the share of spend moving from labor to software
    • Companies may trade some labor costs for higher software/AI OpEx
    • The bull case predicts global IT spend multiplying with agents and AI
    • CFOs need ROI clarity before expanding AI budgets materially
  5. 9:02 – 11:46

    The first baseline: discover what AI tools exist and whether anyone uses them

    Before measuring productivity gains, Russ argues companies must know which tools are in use and by whom. Larridin commonly finds significant “shadow AI” usage—tools employees use that IT hasn’t approved, licensed, or monitored.

    • Enterprises often don’t know all AI tools employees are using
    • Shadow usage can be risky—or a signal of valuable tools to formalize
    • Baseline step: inventory tools and measure adoption/engagement
    • Classic enterprise challenge: software rollout success depends on driving usage
  6. 11:46 – 13:49

    Measuring AI productivity: surveys, behavioral usage, and the privacy constraint

    Russ outlines Larridin’s current approach to productivity measurement: combining traditional survey instruments with proprietary behavioral usage data. They aim for passive measurement eventually, but customer data-sharing and employee privacy concerns set limits.

    • Pure surveys are weak: unclear definitions, biased answers, unknown actual usage
    • Behavioral usage data (who uses what, how often) improves measurement fidelity
    • Combining survey responses + usage mirrors comScore’s methodology
    • Long-term ambition: more passive productivity measurement, constrained by data access/privacy
  7. 13:49 – 20:29

    The productivity baseline problem: outputs vs. time saved (principal–agent tension)

    Alex explores the mismatch between individual incentives (work less) and company incentives (get more output). Russ responds by emphasizing group-level baselines, segmentation of heavy vs. light users, and the CFO’s need to understand whether higher OpEx drives real gains.

    • Time savings don’t automatically translate into company value unless output increases
    • Baseline is hard: what does “twice as productive” mean in practice?
    • Measurement should be aggregate, not individual, due to noise and fairness
    • Key comparison: heavy AI users vs. light/non-users within the same function
  8. 20:29 – 25:59

    Goodhart’s Law and why naive metrics break: from Harvey to Cursor spending

    They dig into the pitfalls of turning measures into targets and how that corrupts metrics. Examples include legal AI tools like Harvey (usage vs. value) and developer metrics like spending on Cursor, illustrating why passive usage plus outcome context matters.

    • Goodhart’s Law: targets distort the measurement itself
    • “Do you feel more productive?” surveys are unreliable and gamed (consciously or not)
    • Usage distribution is uneven (never log in vs. power users), so averages mislead
    • Developer-heavy orgs can sometimes use simpler proxies (e.g., tool spend + manager judgment)
  9. 25:59 – 29:43

    Defining real-world outcomes: responsiveness as a practical enterprise metric

    Russ proposes a pragmatic way to see productivity gains without reducing work to crude counts like emails or lines of code. Measuring interdepartmental responsiveness (informal SLAs) can reveal whether AI is reducing coordination friction and accelerating throughput across teams.

    • Outputs differ by function; many “activity metrics” aren’t true value metrics
    • Interdepartmental responsiveness can proxy productivity improvements
    • AI may reduce bureaucratic coordination costs that slow large enterprises
    • Metrics should inform leaders without being broadcast as gamable employee targets
  10. 29:43 – 32:41

    What 350 IT leaders say: $700B spend, perceived waste, and an 18-month clock

    Russ shares findings from interviews with hundreds of enterprise IT leaders: AI budgets are exploding, leaders believe large portions are wasted, and most feel intense time pressure to keep up. The absence of measurement systems amplifies uncertainty and anxiety.

    • Enterprise AI spend is large and accelerating (often cited around ~$700B)
    • ~70% of leaders believe much AI spending is waste, partly due to lack of measurement
    • ~80–85% believe they have ~18 months to avoid falling behind
    • Board-level mandates exist, but reporting often reduces to purchases rather than outcomes
  11. 32:41 – 37:18

    The human side of adoption: employee anxiety, training gaps, and unsafe experimentation

    The discussion turns to why employees underuse enterprise AI tools: fear of looking incompetent, fear of breaking rules, and lack of training. They argue adoption won’t scale through one-off hero presentations or generic LMS courses; it needs embedded support and safety.

    • Employees worry about what’s allowed and fear consequences (including termination)
    • Too many new tools at once overwhelms workers used to 1–2 annual rollouts
    • Traditional training (LMS) often fails except for mandatory compliance topics
    • Real diffusion requires making employees feel safe and supported in daily workflows
  12. 37:18 – 41:09

    Nexus and “safe AI”: wrappers, policy-aware guardrails, and compliance-by-design

    Russ explains Larridin’s Nexus product: a controlled environment around models that encourages usage while preventing prohibited prompts and data handling. The idea is to create a “safe space” where employees can use AI productively without regulatory or security missteps.

    • Wrappers around models reduce confusion (don’t force employees to choose tools)
    • Guardrails block disallowed prompts and risky data (e.g., SSNs, regulated HR uses)
    • EU/regulated industries need enforceable restrictions to avoid fines
    • Safety and clarity increase adoption and help companies capture internal AI best practices
  13. 41:09 – 52:09

    Future of work: augmentation, competition, and why mass unemployment is unlikely

    Alex asks whether AI will destroy jobs or create new ones; Russ argues competitive dynamics push firms to reinvest productivity into growth rather than shrink headcount. They discuss how white-collar disruption feels different, but skilled workers can adapt, and new roles emerge around AI-driven infrastructure and services.

    • Competition incentivizes doing more with teams rather than firing everyone
    • AI may enable highly profitable solo entrepreneurs, but not broad Fortune 500 shrinkage
    • White-collar disruption is psychologically scarier, but education aids adaptation
    • New jobs likely expand in areas like data centers, new services, and emerging industries
  14. 52:09 – 57:07

    AI’s product marketing problem: “does everything” vs. specific use cases that stick

    They close by arguing adoption often fails because AI is marketed too broadly. Like comScore’s early “we know everything” pitch, enterprises need concrete “tip calculator” use cases—clear, specific value propositions that map to daily work.

    • “AI can do anything” is hard for buyers/users to translate into action
    • Specific promises (e.g., “code better”) drive faster adoption than horizontals
    • Historical parallel: comScore sold better once it packaged clear, verticalized insights
    • AI diffusion will accelerate as companies operationalize narrow, repeatable use cases

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.