Summary

Enterprise AI’s binding constraint is becoming measurement, not capability. Fradin’s analogy is digital advertising: Google and Facebook benefited from an ecosystem of measurement, planning, and governance that helped budgets and revenue scale. AI needs analogous infrastructure. If the industry bull case takes global IT spending from $1 trillion toward $10 trillion—and JPMorgan Chase’s roughly $18 billion–$19 billion IT budget materially higher—the enabling opportunity is infrastructure built “not with the goal of stopping anything, frankly, with the goal of accelerating it.”
Roughly $700 billion of enterprise AI spending is colliding with executives’ fear that they cannot prove what works. Among 350 heads of IT interviewed, about 70% believed money was being wasted, while 80%–85% of the companies thought they had only 18 months to become a leader or fall behind. One PE-backed executive could report progress on four board directives, but for AI, “all I have is the amount of stuff we bought”—a setup that creates pressure for tools to demonstrate usage or value.
Procurement numbers substantially overstate controlled adoption. More than 80% of Laridan customers discover employees using far more AI tools than the company knew about or licensed, even as formal enterprise usage remains lower than outsiders assume. Shadow adoption can indicate either risk or valuable bottom-up demand; the first requirement is visibility, because “they can’t all get retrained all at once with perfect knowledge and perfect security.”
No single metric can establish AI productivity; the defensible approach triangulates passive usage, work output or time, and survey evidence. The Harvey example divides six nominal users into two who never returned, two light users, and two heavy users, then compares their work rather than asking whether they liked the tool. Rampell’s warning is Goodhart’s law: “When a measure becomes a target, it is no longer accurate as a measure,” whether the target is emails, lines of code, or AI spend.
AI creates a principal-agent problem whenever an employee turns an eight-hour assignment into one minute but keeps the method secret. The worker captures leisure while the company receives no additional output; highly competitive, equity-driven organizations are likelier to reinvest the saved time, while larger employers may ultimately adjust workloads and staffing. The diffusion imperative is to “make this person a hero,” memorialize the workflow, and distribute it safely rather than merely buying more seats.
Governance can increase adoption when it gives employees a safe place to experiment without looking foolish or getting fired. Fradin calls a European bank’s response—having one skilled 28-year-old build a 30-slide deck and teach the entire investment bank—“an absurd way to hope people adopt world-changing technology.” Laridan’s Nexus wraps models with company-specific safeguards, while the upside remains highly uneven: “Cursor has taken mediocre engineers and made them good, but it’s taken amazing engineers and made them gods.”
Fradin rejects mass AI unemployment because competitors will reinvest productivity gains into growth rather than leave excess margin undefended. A $100 million company that fires 90% of its staff to earn $90 million could be attacked by a better-funded rival willing to keep hiring and accept 10% margins—“your margin is my opportunity.” He allows that billion-dollar solo businesses may emerge and that older white-collar workers face painful continued learning, but predicts the Fortune 500 will not employ fewer people in 30 years, while conceding, “There is a chance I will turn out to be wrong.”

Deep dive

1. AI needs the measurement stack that made digital advertising scalable

Rampell frames the problem through adtech’s old attribution disputes: did a sale belong to Yahoo’s banner, Google’s last click, or a coupon site that stuffed a cookie? AI has a similarly difficult measurement problem, but the board-level question is simpler: “Did it actually yield a benefit?”
Fradin recalls joining the first online ad network after moving to Silicon Valley in 1996. Digital advertising subsequently required DoubleClick, Flycast, Omniture, comScore, and other layers corresponding to the planning and measurement infrastructure already surrounding television, radio, and pharmaceuticals.
His founding thesis for Laridan is that every rapid budget migration forces infrastructure to be rebuilt. Measurement and governance are not meant to act as gatekeepers; they should let a 35,000-person company expand AI while answering mundane questions about retraining, security, D&O insurance, and whether a project was ultimately valuable.

2. Software is eating labor, and CFO scrutiny follows the spend

Rampell’s stylized model starts with a company spending $10 billion on labor and almost nothing on software. AI might move labor toward $8 billion while software rises to $1 billion or $2 billion, leaving the company more profitable—but making software productivity, not license cost optimization, the important question.
Fradin cites the industry bull case that global IT spending could rise from $1 trillion to $10 trillion because of AI and agents. JPMorgan Chase reportedly spends about $18 billion–$19 billion on IT and a couple hundred billion dollars on people; spending will not jump to $180 billion “in the next couple months,” but any material increase demands CFO-grade evidence.

3. Procurement is not adoption—and shadow AI is already widespread

Laridan begins with inventory: what tools exist inside the company, and are employees actually using them? More than 80% of its customers find substantially more employee-used AI software than IT knew about or licensed—a mix of genuine security problems and popular tools worth bringing “into the fold.”
Formal usage is also lower than many people imagine. Email and Workday have compulsory mechanics, but ordinary enterprise software often reaches only a fraction of its intended population; buying AI tools does not remove the longstanding rollout problem of getting employees to change their workflows.
Fradin emphasizes the psychology of a 42-year-old employee balancing work, travel, and home responsibilities while suddenly being expected to become an AI expert. Adoption rises when employees can experiment without “look[ing] dumb” and know they will not accidentally upload restricted data, violate EU rules, or get fired.

4. Productivity starts with passive behavior, then adds imperfect surveys

Laridan currently combines established productivity surveys with proprietary behavioral data, comparing departments’ heavy and light users rather than monitoring named individuals. The immediate question is whether legal users of an expensive tool, or marketers using Claude or ChatGPT, become more productive than otherwise similar nonusers after the software has definitely raised opex.
Fradin calls self-reported productivity the worst standalone measure: definitions vary, employees infer the answer management wants, and the researcher may not know whether respondents ever used the product. His comScore precedent paired survey answers with observed behavior; fully passive productivity measurement remains the destination, but enterprises are not yet sharing enough data.
The useful unit is an aggregate cohort, not the individual level. Vacations, flights, illness, and training make individual days noisy—“Was I productive yesterday? It’s unknowable”—whereas group-level usage, time, and work volume can tell a CFO whether a requested 50% increase in opex moved anything meaningful.

5. AI breaks the FTE baseline and turns every exposed metric into a game

Rampell’s principal-agent challenge uses a corporate lawyer who compresses eight hours of drafting into four, then plays golf. The employee is “lazier and richer,” but unless expectations or throughput change, the company has paid for AI without capturing the productivity gain.
Fradin’s counterexample is the equity-driven Silicon Valley worker who saves four hours and then works “four more hours and then another four.” That behavior exists at GE too among people aiming to become CEO, but the proportion varies with culture, incentives, and competitive intensity.
AI will disrupt a CFO’s instinctive “horse sense” for what 500, 1,000, or 2,000 FTEs produce. If a large workforce truly falls from eight-hour days to four, Fradin expects management eventually to retain fewer people who work perhaps six hours—not simply tolerate everyone working half as much over the next several years.
Rampell invokes Goodhart’s law: email volume or lines of code may describe activity until compensation makes them targets. The Harvey example instead compares actual nonusers, light users, and heavy users against the same productivity questions and observed output—the minimum combination needed to begin judging value.

6. Responsiveness is a better enterprise output than raw activity

Department-specific outcomes remain essential because sales, legal, engineering, and marketing do not produce interchangeable units. One promising unobtrusive measure is the internal service level: after AI adoption, are colleagues “comfortable sending more things to legal,” and do legal or engineering teams answer other departments faster?
Fradin argues that most CFOs do not dream of firing familiar colleagues: the CFO knows Tina, her husband, and her children, and would prefer that she perform well and never quit. Call centers may be treated differently, but ordinary cost centers can create value through greater productivity, happiness, retention, and responsiveness rather than immediate headcount cuts.
Today Laridan’s customer is the CIO; Fradin expects the buyer over time to become a partnership between the CIO and CFO as the budgets grow.
Independent measurement should eventually benefit effective AI vendors. Fradin says some providers may look askance at third-party scrutiny today, but argues that proof unlocks enterprise budgets—just as Google ultimately acquired Urchin and built Google Analytics because customers tracking value was helpful when the product genuinely delivered it.

7. $700 billion of spending is colliding with an 18-month panic

Fradin cites Gartner’s estimate that roughly $700 billion is being spent on enterprise AI, with rapid growth expected to continue.
About 70% of leaders said they were sure money was being wasted. Fradin does not treat that perception as proof that 70% of projects failed; the deeper problem is that companies lack systems capable of distinguishing successful projects from unsuccessful ones.
A profitable PE-owned company illustrates the gap. Its owners set five annual priorities, including organization-wide AI adoption; the executive had evidence for the other four, but at each board meeting his AI report amounted to “the amount of stuff we bought.”
Meanwhile, roughly 80%–85% of the companies interviewed believed they had only 18 months to become an AI leader or fall behind. That anxiety releases budget before measurement or training exists, producing a “perfect storm” of fast spending, uncertain returns, low employee usage, and confusion over what workers are permitted to do.

8. Safe, social diffusion matters more than another training course

Rampell says AI remains underhyped because spectacular individual experiences have not diffused through institutions. Someone in every large company has likely discovered how to turn eight hours into one minute; “the worst thing that can happen is that guy keeps it a secret.”
At a heavily regulated European bank, one 28-year-old who used ChatGPT well was asked to create a 30-slide presentation and teach the entire investment bank on a global call. Fradin says it may have been cool for the employee but calls the rollout absurd, much like buying an optional learning-management course that almost nobody completes.
Nexus is designed as a safe wrapper around models rather than a directive to choose Claude, Gemini, or ChatGPT. A customized Llama model can block prohibited requests and warn against prohibited data—such as Social Security information, a race-and-gender employee database, or AI-written performance reviews where the company interprets European rules as forbidding them.
The intended bargain aligns both sides: skilled employees receive recognition, nervous employees receive guidance, and the company accumulates reusable knowledge about what works. Fradin’s benchmark is the coding market, where “Cursor has taken mediocre engineers and made them good, but it’s taken amazing engineers and made them gods.”

9. Competition will recycle AI gains into growth, not mass unemployment

Fradin “doesn’t buy for a second” that AI will cause large-scale job loss. A company that preserves output while firing workers may enjoy short-term profit, but a competitor can keep the people, use AI to produce more, and take the market. He points to generally rising GDP and employment and says there is still no proof that the economy is zero-sum.
His VC thought experiment is a $100 million business firing 90% of employees to generate $90 million of profit. Rampell should dislike that plan because another firm could fund a direct competitor willing to keep hiring, accept 10% margins, and destroy that strategy: “Your margin is my opportunity.”
Rampell relays economist Ed Glaeser’s argument that this transition may unusually hit highly educated white-collar workers—but those workers are, almost tautologically, equipped to adapt. Fradin concedes that well-paid professionals who stopped learning in their 40s or 50s may now face an uncomfortable requirement to push themselves again.
More highly profitable solo businesses—and perhaps billion-dollar one-person firms—could emerge, while employment shifts toward data centers, plumbing, media, or jobs not yet imagined. Fradin still predicts the Fortune 500 will not employ fewer people in 30 years, but keeps the uncertainty explicit: if wrong, “maybe I’ll spend more time on vacation.”

10. Horizontal magic still needs a tip-calculator use case

Rampell diagnoses an enterprise product-marketing problem: telling users AI “can do anything” leaves them with no starting point. Adoption accelerates when the promise becomes concrete—“I will help you code better”—rather than requiring every employee to invent a use case.
Fradin learned the same lesson at comScore. “We know everything; what would you like to know?” was a poor sales pitch, while specific products—Visa versus Mastercard share in Japan, pharmaceutical research behavior, or whether employee searches for Ford got worse after the Firestone tire crisis—matched questions buyers already wanted answered.
Rampell’s Sharp Wizard analogy captures the paradox: Jerry Seinfeld’s father loved its tip calculator while ignoring everything else the device could do. Horizontal vendors may resent being reduced to one simple feature, but enterprises need more “tip calculator things”; ChatGPT-style self-evident magic is the rare exception, not a dependable go-to-market strategy.

How a16z Growth Invests Turing CEO Jonathan Siddharth: Who Wins in Data Labelling & Why 99% of Knowledge Work Will Disappear