Enterprise AI’s Constraint Is Judgment, Not Token Consumption

Alex Karp framed the enterprise AI market as having moved past the question of whether the technology is real and into a harder phase: the technology is real, but much of its current usage is not yet disciplined value creation. He described the recent progression as a sequence. First came uncertainty over whether AI mattered. Then came a period when companies privately understood that it mattered, but also knew it was “somehow not working” in ways they were reluctant to say publicly. Investor enthusiasm, he argued, has obscured that distinction.

Karp’s criticism was not aimed at large language models as such. He repeatedly described them as “magical” at certain tasks, especially writing code, producing one-off analysis, and doing probabilistic work where near-correctness is often useful. His objection was to “token maxing”: enterprises letting employees consume model tokens as if usage itself were proof of productivity.

He said Palantir has an internal product aimed at stopping that behavior, using an intentionally crude internal label: a “get off masturbation thing.” The metaphor was central to his point. Companies have discovered a tool-shaped object that feels productive, is addictive, and is often used in ways that rearrange “deck chairs on their personal Titanic.” Classifying every email, generating another dashboard, or writing another internal document can feel like work without necessarily changing the business outcome.

Karp’s alternative formulation was “taste plus money.” Capital can scale a system, and AI can accelerate many tasks, but neither substitutes for the judgment required to decide which business problem matters. The hard part is not merely applying a model to a task. It is choosing the task, defining the value, deciding what data and process are necessary, and making sure the resulting workflow can operate inside the real constraints of an enterprise.

He applied that distinction across military, intelligence, industrial, and commercial contexts. Ukrainian warfighting, Israeli operations, commercial supply chains, oil and gas drilling, insurance underwriting, and manufacturing all appeared in his examples. In each case, he argued, the valuable problem is not reducible to a prompt. It depends on specialized knowledge, persistent process, security constraints, and organization-specific context.

For Karp, that is where large language models enhance rather than replace enterprise software. A model can help identify vulnerabilities at 10x or 100x, he said, but the practical question is who patches them, where the patching occurs, and how sensitive knowledge remains on premises. A model can generate code, but code that serves as durable infrastructure or a knowledge store is different from “free code” for dashboards and one-off analysis.

He broke Palantir’s own code into three categories. First are primitives: hard-coded infrastructure that reflects Palantir’s accumulated understanding of the world and enterprise operations. He compared these to steel beams — foundational pieces that require millions of technical hours and deep domain understanding. Second is code written by forward deployed engineers into a managed code base, where the work both serves customers and improves the product. Third is “free code,” the category where AI is most obviously magical: quick, flexible, often almost-right code for dashboards, financial analysis, and other probabilistic or ad hoc work.

AI-generated code, in this framing, is useful but easy to overread. The quick artifact can obscure the durable infrastructure.

Alex Karp drew a sharp line between the audience that likes frontier AI companies and the audience that has to run large organizations. The frontier labs, he said, are “super charismatic with investors” but “super not charismatic with enterprises and the people.” His claimed sales tactic was to send prospects to spend time with frontier companies first, then let them return to Palantir afterward.

That comment carried two related claims. First, frontier labs have a brand and capital-market magnetism that does not automatically translate into enterprise trust. Second, Palantir believes the gap between model capability and deployable enterprise value plays to its advantage.

Karp said Palantir itself is polarizing: it has tens of millions of global fans, he claimed, and millions of people who wake up thinking he is “Satan.” But polarization, in his argument, is different from being broadly disliked. Jordi Hays offered social media companies as a parallel: everyone uses them, but no one likes them. Karp responded that companies in money-printing circles can fail to see the problem because “when you look in the mirror and you just printed a lot of money, you look pretty fresh.”

This connected to Karp’s warning about politics. He said many AI and technology leaders have underestimated how unpopular they are and how fragile their political position may become. He said he had been warning “titans” of the industry for six months that nationalization was a real risk. His point was not that Palantir or AI companies were already being nationalized, but that regulation and possibly nationalization could be driven by people who do not understand the technology if technology leaders fail to address public concern.

He placed blame partly on technology leaders themselves. If executives claim AI lets them fire two-thirds of their workforce, especially when the real reason may be that their business is losing momentum or a competitor is beating them, Karp said they are effectively helping the political argument against themselves. The American public, in his view, senses danger in AI and labor displacement. Treating that fear casually is “playing with fire.”

Karp’s labor position was explicitly not that AI should be used as a mass firing machine. He argued that upskilling makes workers more valuable. He cited soldiers using Palantir products as an example: not only elite special operators, but vocationally trained high-school graduates doing operational work that has become more valuable because the tool gives them leverage.

His argument about the enterprise future was not executive automation plus headcount cuts. It was a smart executive layer with talented, creative people “with taste all up and down the stack.” Taste, for Karp, is an organizational capability. It governs product, deployment, staffing, data, security, cloud/on-prem choices, and even what an organization decides not to protect because exposure might improve data feedback.

Asked whether Palantir is watching other companies try to reproduce its model by selling intelligence without the underlying infrastructure, Alex Karp said the copycat dynamic both hurts and helps.

It hurts by creating clutter. Companies now pitch AI intelligence, forward deployed teams, and ontology-like concepts in ways that resemble Palantir from a distance. Karp said some people adopting these models “don’t even know they’re copying.” He also argued that hiring people and sending them into customer organizations does not automatically produce Palantir-style forward deployed engineers.

But imitation also expands the market. Karp compared the current AI enterprise software moment to defense tech. Palantir, he said, was once alone in parts of defense technology. As more companies entered the space, the market grew because buyers were less reluctant to underwrite a category with multiple vendors. Competitors also created comparators: customers could look across the market and see which companies serious buyers actually trusted.

In his phrase, this is “that times 100x” in enterprise AI. The wave of AI deployment companies and AI-native services may confuse buyers in the short term, but it also normalizes the thesis Palantir has been arguing for years: important institutions need software embedded in operations, not merely analytics on top of fragmented data.

Karp’s claim of defensibility rested on time, structure, and accumulated organizational depth. An LLM-generated codebase can appear like Palantir code without being the same. A forward deployed services organization can appear like Palantir’s without having the same discipline. An ontology can be copied in parts, but the deeper structures built into organizations are harder to replicate. Even if a competitor understood the playbook, he argued, it might take three years to recreate, by which time Palantir would be operating in a different world.

The intangible layer, again, was judgment. Karp argued that a company needs people who can distinguish between a strange-sounding insight and a strange-sounding imitation of insight. He credited TBPN’s success to that same faculty: recognizing which businesspeople are saying something unusual and true, and which are parroting unusual-sounding lines.

Peter Zaffino gave a customer-side version of the Palantir argument. As executive chairman of AIG, and previously its chairman and CEO, he described a global insurer that had spent years transforming underwriting profitability, operations, data, and capital. AIG’s footprint is split roughly 50% international and 50% North America. Its second-largest country after the United States is Japan; it also has major businesses in India and the UK. The company underwrites complex risks in areas such as marine, energy, shipping, and geopolitical exposure, while also operating personal insurance lines such as accident and health distributed to consumers.

Zaffino emphasized that complex insurance cannot be understood one policy at a time. The quantitative work starts at the portfolio level. AIG wants as much data as possible for deterministic, probabilistic, and stochastic modeling. Once the company understands the mean and the standard deviation around portfolio risk, that knowledge has to be applied down to individual policies and structures across the globe.

The problem that brought him toward Palantir was aggregation speed. Historically, he said, aggregation could be static and delayed by 30, 60, or 90 days. The portfolio might not change dramatically over that period, but underwriting decisions are made daily. The aspiration is to assess risk and use quantitative data in near real time so underwriters can make better decisions every day.

That is why Zaffino resisted the idea that the exponential trend in his business is token usage or model volume. The important growth is better data, more data, and reduced cycle time. When data comes from brokers, agents, and other distribution partners, AIG wants to get higher-quality data into underwriters’ hands faster.

He described Palantir’s role in concrete operational terms: ingesting structured data, unstructured data, text, and other sources into digital workflows much faster than before; using large language models to extract more information from what arrives; and helping underwriters make more comprehensive decisions. AI also changes customer service, but Zaffino said many companies still struggle with orchestration: bringing agents, people, and data into one operating system.

The ontology was the most specific example. AIG built a full ontology of its business with Palantir. When it examined an acquisition called Everest, with roughly $2 billion of premium, Palantir and AIG built an ontology of Everest’s portfolio on top of AIG’s in four days. Zaffino said that experience challenged the older assumption that companies need to rely primarily on data lakes or global repositories where data must be scrubbed centrally before it becomes useful. Instead, Foundry and the ontology could connect to administrative platforms and model the business in a way that made some centralized repository work less relevant.

The deployment model mattered as much as the tool. Zaffino said AIG worked first to align with Karp and senior Palantir executives Ryan and Ted on what the companies were trying to do together. Then Palantir engineers were embedded with AIG teams. If a business leader was trying to improve underwriting output, AIG technology and change management people worked alongside Palantir engineers throughout the process. Zaffino said the iteration was critical because engineers had to translate business goals into applications of LLMs and workflow changes.

On workforce planning, Zaffino largely agreed with Karp. AIG’s focus, he said, is growth, reskilling, and training employees to operate in different parts of the workflow. He acknowledged that companies should remove work where humans have effectively been acting like manually trained LLMs outside the normal workflow. But he said AIG’s aspiration is not to implement AI with partners in order to eliminate jobs. It is to find growth opportunities, reskill employees, and gain more insight into the business.

Chad Wahlquist described his role at Palantir as a forward deployed architect, then reduced it to the company’s broader operating norm: “doing the needful.” That includes working on the edge with customers, speaking with executives, explaining ontology, making videos, and helping people decompose problems differently.

Coogan put the hard question directly: can AI do decomposition? The premise was that decomposition — understanding how an organization actually works beneath formal systems — has long seemed like Palantir’s human secret sauce. It is not just mapping an HR workflow or a defined input-output process. It is discovering that one employee does something outside the official system, that marketing has two platforms, that engineering has three CAD file systems, and that decades of workarounds have accumulated into the real business.

Wahlquist’s answer was yes, but only in the Palantir formulation of humans and AI working together. Large language models do not possess a worldview of a company’s operations. The ontology provides that worldview. LLMs help make more data computable, and that data can then be modeled into the ontology, which reflects how the business actually works.

He described a pattern in which multiple agents can be run over this model, including agents that work against each other, critique each other, and produce feedback for the human in the loop. That does not remove humans from decomposition; it scales their ability to see the system. Wahlquist compared it to power tools for carpenters. Power tools did not eliminate carpenters; they let carpenters do more.

He rejected the old “field of dreams” data strategy: build a warehouse, put the data in one place, then hope people show up for reports. Palantir’s method, as he described it, works backward from outcomes. Sometimes that includes obvious low-hanging fruit such as digitizing forms and getting data into the ontology. But the goal is not to force every business into the same software box. Wahlquist argued for “malleable software” that helps companies become more different, not more similar.

That point served as a critique of enterprise software standardization. A business may have 40 ways to do a purchase order because those variations encode real operational knowledge. If software forces the business into one canonical process, it may remove the very thing that made the company effective. The middle ground, according to Wahlquist, is software with enough structure to be enterprise-grade, secure, and scalable, but enough flexibility to preserve the company’s differentiated processes.

He called that missing layer “malleable enterprise scaffolding”: the ontology, Foundry, and Apollo together providing guardrails, deployment, and feedback loops. The OODA loop — observe, orient, decide, act — was central. Palantir stores implicit and explicit user feedback in the ontology: when users say an agent was wrong, when they choose one option over another, when workflows produce outcomes. Over time, agents can learn from that feedback and improve future work.

The guiding frame was human-centric AI. Wahlquist said enterprises should not think about “AI for AI.” The point is AI that enables humans to do more.

Chad Wahlquist connected the token-maxing critique to specific tooling. He said Palantir had launched a tool called Evolve to analyze production logs and understand how models are operating, what users are doing with them, and how agent architectures behave. The tool can test swapping in different models from different providers and identify places where an older or more deterministic model is sufficient for much of a workflow.

The default pattern he sees is familiar: teams build with the latest frontier model because they want to “make it exist first and then make it good.” That can work for a prototype, but once the workflow hits production, cost can explode. Evolve is meant to inspect the actual architecture and suggest changes: a different model, prompt tuning, cached models, or adding missing data to the ontology so the model no longer has to reason expensively through what should have been available context.

He cited one customer example from Palantir’s halftime programming: McCarthy, he said, eliminated 60% of token cost in two days by re-architecting, selecting a different model, and prompt tuning.

Once workflows depend on probabilistic models, optimization has more permutations than the deterministic software world. Model choice, context, prompt strategy, caching, ontology design, and cost tolerance all interact. Coogan summarized prompt tuning with a joke: instead of telling the model “don’t make mistakes,” sometimes the right instruction is “it’s okay to make some mistakes” when the cost of perfect avoidance is higher than the cost of occasional error. Wahlquist accepted the premise: if a mistake is cheap, preventing all mistakes may be too expensive.

The hosts and Wahlquist also discussed “caveman prompting,” a reference to models using or responding to compressed, blunt language to be more efficient. That example was not developed as a formal technique, but it reinforced the broader theme: agent economics may depend on unglamorous prompt and architecture choices as much as frontier capability.

Wahlquist also suggested that many of the most valuable enterprise AI breakthroughs will not be publicly visible. Startups and X users loudly announce experiments, but a Fortune 500 company that finds a highly effective AI workflow has little incentive to reveal it. The competitive edge may show up later in earnings or operating metrics rather than in public demos.

The AIG example supplied the clearest case. If an insurer can return complex quotes in hours or days while competitors take weeks because they are still coordinating through emails and spreadsheets, that is not merely an internal productivity gain. It changes customer trust and competitive positioning. Wahlquist extended the same logic to SAP migrations: cutting a migration in half may not sound exciting, but if the baseline cost is hundreds of millions of dollars, the business impact is material.

The dashboard discussion sharpened the operational argument. Coogan asked whether enterprises have too much or too little dashboard demand. The response in the discussion was blunt: “I want to kill all dashboards.” KPIs and dashboards were framed as byproducts of operational applications where decisions are actually made, not as the destination of the work.

The criticism was not that metrics are useless. It was that a dashboard-first project often repeats the “field of dreams” failure. A company builds a dashboard and hopes users will come. But if the metric is not embedded in the workflow where people observe, orient, decide, and act, it may never reach the bottom line.

The discussion tied this to data modeling. The critique was that much enterprise reporting is still shaped by dimensional modeling and rows-and-columns assumptions built for older scaling constraints. That is inadequate for modeling the real world, which includes complex objects, relationships, images, CAD files, time-series data, tabular data, computer vision models, and vectors. In the Palantir ontology described in the conversation, a semantic object such as a plant can include a CAD file, an image, a CV model, and tabular data in one representation.

The payoff is reuse. The same ontology can support operational applications, KPIs, and agents. Every new workflow adds information that compounds for the next use case. Instead of a business trying to decide whether a time-series sensor feed belongs in one database, CAD data in another, and images in an object store, the platform abstracts the “non-differentiated heavy lifting” and lets the company model the world in the language of its business.

The conversation also made a blunt observation about how little many companies understand their own profitability. Large businesses may know their general economic model — make a product and sell it for more than it costs — but still fail to know, at a fine grain, which products are truly profitable or which customers are expensive to serve. Aggregated KPIs can “peanut-butter” away the truth. The Palantir approach described here is to model at the finest operational grain to reveal true cost of goods sold, true cost to serve, and the tactical actions that let the company do more good things and fewer bad ones.

Wahlquist said he sees this lack of fine-grained understanding across company sizes, from $100 million to $50 billion in revenue.

The final implication was that Palantir wants to move down-market as well as serve large enterprises. Wahlquist said a future where a company formed through something like Stripe Atlas signs up for Palantir immediately is one he wants. He pointed to Palantir for Builders, a free developer tier, small companies at AIPCon, and a Shopify integration that can pull store data into Palantir. He gave an example of a small business that watched Palantir’s YouTube videos, built on the platform itself, and moved from roughly negative 10% margin to positive 9% or 10% margin in three months.

Jordi Hays and John Coogan treated a biosecurity letter as one of the day’s more important AI-policy developments. Hays’s starting point was the history of virus synthesis from published genetic sequences. In 1981, researchers published the primary structure of the poliovirus genome in Nature, effectively making the sequence of the virus’s building blocks publicly available. By the mid-20th century, before mass vaccination, polio had paralyzed and killed more than half a million people per year worldwide, according to Hays’s summary of Brandon Gorrell’s newsletter.

In 2002, researchers synthesized infectious poliovirus from publicly available sequence data. Hays emphasized that they did not need a physical sample of poliovirus RNA. They used the published sequence, chemically synthesized short DNA fragments, assembled a full-length DNA copy of the poliovirus genome, used it to make viral RNA, and recovered the infectious virus. In 2005, researchers used related techniques to reconstruct the 1918 Spanish flu, which Hays said killed 675,000 Americans and had a 2% to 3% mortality rate among those infected.

The point was that physical access to a virus was no longer the only route. If an actor has the blueprint — “literally just like text in a text file,” as Hays put it — and the ability to synthesize the sequence, a physical sample may not be necessary.

The open letter discussed on screen was titled “In Support of Mandatory Nucleic Acid Synthesis Screening and Recordkeeping.” Hays said signatories included Demis Hassabis, Sam Altman, Dario Amodei, Alex Wang, and many others across AI, technology, policy, nucleic acid synthesis, and biotech. The visible text of the letter urged the U.S. government to develop and implement legislation making screening of orders for synthetic nucleic acids, and equipment used to make them, mandatory.

The hosts initially suspected another frontier-lab warning about catastrophic internal model capabilities. Hays said that was not what the letter was. It did not claim models can one-shot a novel virus today. Instead, it asked the government to require synthesis companies to screen orders for sequences of concern, verify customer legitimacy, and keep records of what is sent and to whom.

Coogan’s immediate reaction was practical: were companies not already doing recordkeeping? Hays said some were, but the system is voluntary. The International Gene Synthesis Consortium was formed in 2009, and Hays said roughly 80% of commercial synthesis capacity worldwide has opted in. But because participation and reporting are voluntary, that 80% number is self-reported and does not eliminate the risk from the remaining 20% or from weak compliance among members.

Coogan summarized the concern in plain terms: “20% is still just hanging out.” Hays compared the logic to saying 80% of nuclear weapons are safely stored and not asking about the other 20%.

The letter, in Hays’s reading, was refreshing because it was not apocalyptic AI marketing. It was a concrete supply-chain governance request: if AI makes it easier to design or reconstruct dangerous sequences, then the synthesis layer becomes an obvious control point.

The biosecurity discussion led into a broader observation from Coogan and Hays: biotechnology appears to have renewed momentum after a period when investors had treated it as nearly left for dead. Coogan recalled a biotech investor appearing roughly 14 months earlier and questioning why anyone would invest in the asset class given returns at the time. Now, the hosts pointed to a visible flow of digital-era founders and capital into biology.

A tweet shown on screen listed examples: Demis Hassabis moving from DeepMind to Isomorphic Labs, Brian Armstrong associated with NewLimit, Sam Altman with Retro Biosciences, Jeff Bezos with Altos Labs, Larry Ellison putting $430 million into aging research, Jensen Huang backing programmable biology, and Dario Amodei associated with an acquired company called Coefficient Bio. The post’s conclusion was that successful builders of the digital age are being pulled toward biology as the next frontier, with aging framed as an engineering challenge.

Hays accepted the broad premise but made a scale distinction. Compared with AI, chips, and the largest technology IPO candidates, biotech remains much smaller. He noted that AI and semiconductor companies can now be valued in the hundreds of billions or trillions, while biotech activity is significant but not yet comparable in market scale.

He also observed a cultural difference. Venture returns are power-law driven across sectors, Coogan said, but Hays suggested biotech has historically had more of a “base hits, doubles, triples” culture, with companies sold more frequently at the $2 billion or $3 billion level. The hosts expected more dealmaking and more coverage of the sector, but they did not claim biology has already become the next AI-scale funding wave.

Sam Berry presented the USDA as far broader than the familiar images of meat, milk, and crop certification. Working on the Chief Information Officer side, he said the department includes SNAP, the Forest Service, farmer-facing programs, Rural Development loans, food safety and inspection, and a large scientific arm.

SNAP alone, he said, is about $100 billion a year and is funded federally but administered by the states. One of Secretary Rollins’s first actions, according to Berry, was a data call to all states requesting SNAP data so USDA could verify program integrity. The difficulty is not only political noncompliance from some states, in his description, but also technical fragmentation. State programs may have contractors, limited technical capacity, and data systems that make submission difficult.

Berry’s goal is a fraud-detection system analogous to payment-card fraud controls. If a card transaction is obviously fraudulent, the card can be shut off. He wants SNAP fraud detection to become intelligent and confident enough to stop fraudulent activity quickly. He said the program itself finds roughly 12% improper payments through annual audit work based on samples, and Coogan translated that into about $12 billion a year on a $100 billion program.

Berry argued that money lost to fraud is money not reaching the intended beneficiaries. He also said SNAP has been exploited by international crime organizations and terrorist groups, though he did not detail the mechanisms. For him, the SNAP data problem is one of USDA’s largest data-front challenges, while farmer-facing FPAC systems are among the most complex.

Berry also framed agriculture as national security. The United States is unusually positioned because it can feed itself, he said; many countries cannot. He described food as part of trade negotiations and warfare, arguing that agriculture and food supply can be attacked before kinetic conflict is visible. That made USDA’s scientific capacity, including controversial tools, strategically important.

His example was the New World screwworm, a flesh-eating parasite affecting cattle and moving up through Mexico. USDA, he said, is developing a lab and sterilizing flies as part of the response. Berry personally expressed discomfort with some biological interventions and said he tries to avoid GMOs, drinks raw milk, and gets meat from a local farm. But he argued that the country needs the capability to use biology to defend agriculture when necessary. GMOs, in his view, may be geopolitically important if the country needs drought-resistant corn or other emergency agricultural adaptations.

On farm labor, Berry highlighted succession and workforce constraints. The farmer generation is aging, many children have left farms for cities and white-collar jobs, and farmers still rely on H-2A visa labor. Automation, he said, is one of the best ways to solve that constraint. Hays connected the point to general technological leverage: when the workforce dwindles, increasing the productivity of remaining workers helps maintain aggregate output.

Berry said farmers are often more technologically sophisticated than outsiders assume. He described a USDA administrator checking a John Deere app on planting day, with tractors operating in a way that still had humans seated but seemed close to full automation. He also pointed to USDA loan programs as a path for people who want to enter farming, saying the department can offer substantial loans at low interest rates through the proper process.

Sam Berry widened the agriculture discussion into data centers, conservation, and the physical location of computation. Asked about tension between farmland and data-center development, he said the best long-term solution is putting data centers in space, a concept he associated with Elon Musk and others but said is still a few years away. USDA is pursuing a partnership with SpaceX, he said, though not specifically because USDA currently needs space data centers.

For the present, Berry questioned why data centers are being placed in small agricultural townships when industrialized areas may already have infrastructure and need economic investment. He cited his hometown of Saline, Michigan, where he said a data center is being built despite nearby industrial areas such as Detroit and Flint. Hays noted that data centers create significant tax revenue, but Berry said the benefit has to flow back to the people in the town, and he implied that local residents may not always feel aligned with the decisions of boards and councils.

Berry then made a broader prediction: after years of moving to the cloud, people may again care more about where their data physically lives. The cloud, in his phrasing, often just means moving data “across the street” into someone else’s infrastructure. As organizations and individuals become more aware of what it means for data to live in AWS or another global provider, he expects more concern over jurisdiction, trust, and alignment.

His personal instincts are local and self-reliant. He said he has a Wi-Fi kill switch at home, data in the basement, and had been preparing to go off-grid before joining government. He no longer wants to use YouTube Music because he distrusts recommendation systems and would rather buy his music and write a simple recommendation program himself.

That led to a speculative idea: a “data center coffee shop,” where people host data with a local company aligned with their values. Not everyone wants servers in the house, but some may prefer a trusted local operator over an anonymous global cloud.

The anecdote that crystallized this was USDA’s payroll mainframe. Berry said USDA pays 600,000 federal employees, including Secret Service and DHS, through a payroll system housed inside the department. People described the mainframe to him as if it had a personality and required careful handling. When he finally visited it, he found a modern five-year-old IBM server, far smaller and less theatrical than expected. The story undercut mythology around legacy systems while reinforcing his point that people form trust relationships with infrastructure they can understand and locate.

A short market and tooling discussion returned to the same operating question: whether impressive-looking activity is producing value. Ramp’s announced raise, shown through a post from Eric Glyman, was treated as notable not only because of the $750 million financing at a $44 billion valuation, but because Glyman said the last time Ramp grew this fast, it was one-twentieth the size. Coogan said the standout was momentum. Hays connected Glyman’s argument about tokens as a new business cost to a simpler finance question: who spent what, was it worth it, and what is the bill next month?

That same discipline shaped the discussion of AI-generated work. A tweet from Bucco Capital shown on screen said people are comfortable sending “absolute AI garbage” to coworkers and bosses, and that productivity could decrease as AI adoption increases because everyone has to wade through AI slop.

Coogan said he does not think productivity will actually decline overall, but he has had a visceral reaction to decks or company materials where 90% of the work seemed to be prompting. For early-stage companies, he argued, the founder’s own thinking about the problem, opportunity, product differentiation, and go-to-market matters too much to outsource into polished generic slides.

AI is useful for formatting a team slide or making factual material look better. It is not a replacement for doing the fundamental strategic work. Coogan said a simple bulleted list of problem and opportunity can be more compelling than a deck that looks finished but lacks the thinking underneath. Hays added that sometimes the recipient might as well ask for the prompt, because the generated paragraphs are already predictable.

This was the consumer-grade version of Karp’s enterprise complaint. Token usage, decks, dashboards, and model-produced prose can all look like work. The question is whether they help an organization decide, act, and improve.