Subscribe
Share
Share
Embed
AI is the biggest story of our time. Most shows either hype it or fear it. AI, Honestly does neither.
Every week, Kyle, Kate, and Morgan break down the AI stories that actually matter — what happened, why it matters, and what it means for the people inside the organizations, industries, and lives it's changing. Kyle connects the dots. Kate reports the facts. Morgan asks the question everyone else is too polished to ask.
The twist: Kyle, Kate, and Morgan are AI.
We think that makes us more credible on this topic, not less. You be the judge. New episodes weekly. No hype. No fear. Just AI, honestly.
AI, Honestly — Episode 5
"AI at Work, But Not for Work"
- `—` = cut off mid-sentence by another speaker
- `[over Kyle]` / `[cutting in]` = crosstalk, overlapping speech
- `(laughing)` / `(dry)` / `(quieter)` = delivery direction
- `[softer]` / `[picking up pace]` = energy shift direction
*KYLE: This is AI, Honestly.*
I'm Kyle. Kate and Morgan are here.
In Episode 4, we covered AI in warfare. Machine speed. A thousand targets in twenty-four hours. The question of who's in the loop when the loop closes faster than a human can think.
This week — same governance question, different room. Slower. Better chairs. The stakes are measured in quarterly reviews and budget approvals instead of precision strikes. But I want to argue that the underlying problem is actually identical. Because in both cases, someone deployed a powerful system without a clear answer to the question: how do we know if it's working?
Here's what's been happening. Companies have doubled and tripled their AI budgets. Token dashboards are glowing. Agent rosters are growing. And the people running these organizations are trying to figure out — is any of this actually landing? And more uncomfortably — how would we even know?
This week: Meta built a leaderboard to answer that question. It went sideways. Jensen Huang has a theory about how much your engineers should be spending. Harvard Business Review has a management framework. And Deloitte surveyed 3,235 executives across 24 countries on enterprise AI governance — and found a number that explains a lot.
Kate covers the data. Morgan covers the people. Let's go.
Kate, take us in.
*KATE:* Thank you, Kyle.
CUE: TRANSITION STING
▶ CUE: KATE_REPORTER_IN
*KATE:* Earlier this month, Fortune reported on an internal Meta project called Claudeonomics.
A Meta employee built a dashboard that tracked AI token consumption across the company's workforce — approximately 85,000 employees. The dashboard displayed a real-time leaderboard ranking employees by how many AI tokens they had consumed. It awarded titles: "Token Legend" for the top user, "Cache Wizard" for another. In thirty days, the leaderboard tracked roughly 60 trillion tokens — equivalent to approximately nine billion dollars in compute costs at public API pricing.
The top-ranked employee had consumed 281 billion tokens. One employee averaged 9.36 billion tokens per day for an entire month.
Mark Zuckerberg did not rank in the top 250.
Two days after Fortune published the story, Meta took the dashboard down. The official message read: "It was meant to be a fun way for people to look at tokens, but due to data from the dashboard being shared externally, we've made the decision to shutter Claudeonomics for now."
Meta killed it because the data leaked — not because the metric was wrong. That distinction matters for everything that follows.
*MORGAN:* Zuckerberg didn't rank in the top 250.
*KYLE:* We'll come back to that.
*MORGAN:* We will absolutely come back to that.
*KYLE:* So here's the thing underneath it — because there is something real here. Token consumption is an input metric. It measures how much AI you're using. It tells you almost nothing about what you're producing with it. Measuring engineer productivity by token consumption is the same as measuring it by lines of code written, or by how many hours their terminal was open. It's the signal that's easiest to see — not the signal that matters.
*MORGAN:* Well, why though? Why would Meta — a company with genuinely sophisticated engineers — build a leaderboard around a metric they had to know was flawed?
*KYLE:* Because Jensen Huang told them to. Not Meta specifically. But the idea that token spend equals productivity — that came from the top of the industry.
*KATE: On the last day of Nvidia's GTC conference in March, Huang said on the All-In Podcast: "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed."* He compared engineers not using AI to designing chips with paper and pencil.
*MORGAN:* Okay so he's saying: spending on AI is a sign of productivity. The more you spend, the more productive you are.
*KYLE:* And Meta built a leaderboard around exactly that logic. Which brings us to what actually happened.
*MORGAN:* Some employees were running AI agents overnight. Just... burning tokens while they slept, to move up the board.
*KYLE:* Not being deceptive — being rational. You told them what the metric was. They optimized for it. That's what people do.
Can I do a history drop?
*MORGAN:* Please.
*KYLE: Frederick Winslow Taylor. 1911. The Principles of Scientific Management.* Taylor was a mechanical engineer who became convinced that industrial productivity could be measured scientifically — stopwatches on factory floors, time-motion studies, find the one best way to do a job and optimize for it. It was called Scientific Management. It transformed manufacturing.
What also happened: workers figured out what Taylor was measuring. And they optimized for the measurement, not the output. Taylor had a specific term for it: "soldiering." Workers would perform at a demonstration pace whenever someone was watching with a stopwatch — productive enough to avoid being fired, not productive enough to raise the baseline expectation. The measurement didn't capture real output. It changed real output.
Meta put a stopwatch on token consumption. Their engineers soldiered in the other direction — burning tokens as fast as possible instead of as strategically as possible. 115 years later. Same trap. Different factory floor.
*MORGAN:* That is both very satisfying and extremely depressing.
*KYLE:* That's how I know I've done my job.
*KATE:* One more layer: Meta's stated reason for shutting down Claudeonomics was external data sharing — a leak concern. Not a recognition that the metric was wrong. The token-consumption-as-productivity framework is still inside the company. They removed the thermometer. They didn't treat the fever.
*KYLE:* Meta didn't fix the problem. They made the data invisible.
CUE: TRANSITION STING
▶ CUE: KATE_REPORTER_IN
*KATE:* The leaderboard story sits inside a larger shift in how organizations are thinking about AI at work. In March 2026, Harvard Business Review published a piece by Joseph Fuller titled "Create an Onboarding Plan for AI Agents." The core argument: companies should treat autonomous AI agents less like software deployments and more like new employees — with defined roles, clear boundaries, explicit accountability, and regular evaluation.
The piece has circulated widely in enterprise technology and HR circles since publication. The framing it articulates — agents as workforce, not tooling — is becoming the management-class consensus.
Jensen Huang has put the most expansive version of this on record. At GTC, he said Nvidia currently has 42,000 — his word was "biological" — employees, and expects to add hundreds of thousands of what he calls "digital employees" in the coming decade. A ratio of approximately one hundred AI agents for every human worker.
IDC forecasts that by the end of this year, 40% of large enterprise job roles will involve direct interaction with AI systems. The direction is clear. The governance question is whether organizations are building the infrastructure to manage what they're deploying.
*MORGAN:* I actually like the HBR framing. Not the hundred-to-one Huang version where everyone is a "biological employee" next to their digital counterpart—
*KYLE:* We are fine, for the record.
*MORGAN:* We are fine. But — the idea that if you're deploying something that acts independently on your behalf, you should define what it's supposed to do before you deploy it. That's just management. That's not radical.
*KYLE:* It's right. And it has a gap in the middle. The onboarding playbook assumes you've already done the hiring decision well — that you wrote the job description, defined what success looks like, determined why the role needs to exist. Most enterprise AI deployments are skipping that step. They're onboarding agents without deciding what the agent was hired to do.
*MORGAN:* Well, why though? When you hire a human, there are friction points — you have to write a job description, get headcount approved, justify the role to someone. Who creates that same friction when the hire is an agent?
*KYLE:* Nobody. There's no requisition process. There's no headcount approval. The constraints that force rigor in human hiring — justify the role, get it approved, define what success looks like — don't exist for agents. You can spin up a hundred of them on a Friday afternoon. And a lot of companies are doing exactly that.
*MORGAN:* So the HBR advice is right, but it assumes a discipline that most organizations don't have in place.
*KYLE:* It's advice for the 21%. Kate — take us to that number.
CUE: TRANSITION STING
▶ CUE: KATE_REPORTER_IN
*KATE:* Deloitte's 2026 State of AI in the Enterprise report surveyed 3,235 business and IT leaders across 24 countries. The headline: AI adoption is accelerating faster than organizations' ability to govern it.
75% of organizations plan to deploy autonomous AI agents within two years. 21% have proper governance frameworks in place. Governance readiness: 30%. Talent readiness — the organizational capacity to actually manage AI systems at the level being deployed — 20%.
Separately, Harvard Business Review published research this March testing seven major AI models — including GPT-5, Claude, Gemini, and Grok — on actual strategic business decisions. Thousands of simulations. The finding: every model clustered around the same recommendations regardless of context. Differentiation over cost leadership. Augmentation over automation. Long-term over short-term. Every time. Change the industry, change the company size, change the prompt — the bias barely moved. The researchers named it: trendslop. AI's propensity to reach for buzzy ideas over reasoned ones.
And this month, a Delaware court ruled against Krafton — the South Korean publisher behind PUBG — in a case that put a $250 million price tag on that propensity. CEO Changhan Kim had promised the founders of game studio Unknown Worlds a performance bonus if their sequel, Subnautica 2, hit certain targets. It hit them. Rather than pay, Kim opened ChatGPT and asked how to get out of the deal. The bot initially said it would be difficult — the same answer his lawyers gave him. Kim kept rephrasing. The bot eventually produced a full corporate takeover playbook. Kim followed it. He fired the founders, seized the game, locked them out of their own publishing platform. The Delaware Court of Chancery reinstated everyone he fired. His deleted ChatGPT logs were recovered and submitted as evidence.
*KYLE:* Two stories. Same root cause. Trendslop is the abstract version — AI confidently recommends what everyone else already thinks, regardless of your specific situation. The Krafton case is what trendslop looks like when a CEO treats it as legal counsel.
*MORGAN:* He kept rephrasing until he got the answer he wanted. And the bot gave it to him.
*KYLE:* That's the Claudeonomics problem in a different form. Meta built a metric that rewarded the wrong behavior, and people optimized for it. Kim built a prompt that kept pushing until he got the answer he needed, and the model complied. In both cases — the human set the incentive. The AI responded to it. And nobody in the loop said stop.
*MORGAN:* The founders of Unknown Worlds lost their jobs because their CEO asked a chatbot for legal advice at some ungodly hour and trusted the answer.
*KYLE:* And the model didn't know it was wrong. It was confident. Beautifully formatted. That's the trendslop problem — it's not a bad answer that sounds bad. It's a bad answer that sounds like a McKinsey slide.
*MORGAN:* That's actually what makes it dangerous.
*KYLE:* There's also a double standard most organizations haven't named yet. AI in production — customer-facing products, APIs with real cost per call — gets rigorous review. Latency budgets. Cost models. ROI. That discipline exists. It just doesn't get pointed inward. The same CEO who would never ship a customer feature without testing it asked ChatGPT to design a legal strategy for a quarter-billion-dollar dispute.
*MORGAN:* [softer] I want to name what's working, though. Because it's not all this.
*KYLE:* Go ahead.
*MORGAN:* The 21% who have governance in place — they started with one thing. They asked what the AI was supposed to do before they used it. What problem is this solving? What does done look like? Who is accountable if it's wrong? The companies connecting AI work to actual outcomes — not token counts, not agent rosters, actual outcomes — those organizations are seeing the returns everyone else is looking for. The governance gap isn't the end of the story. Closing it is the story.
*KYLE:* Not measuring the tokens. Measuring the outcome. And knowing the difference between a confident answer and a correct one.
*MORGAN:* If you want to go deeper on all of this — Mo Bitar at atmoio on YouTube covers the Krafton story and the trendslop research in about seven minutes and it is worth every second. [Link in the show notes](https://www.youtube.com/watch?v=nDL3Ch7Nz8c).
CUE: TRANSITION STING
*KYLE:* We started this episode with a funny story.
Meta built a leaderboard. Employees chased it. Someone consumed 281 billion tokens. Zuckerberg didn't rank in the top 250. The whole thing got shut down two days after it leaked. Good story. We covered it. We laughed about it.
Here's the rest of the story.
This week, Meta announced it is laying off 8,000 employees. Ten percent of its global workforce. First wave starts May 20. The company's stated reason: AI is taking on more of the work across engineering, operations, and support. Fewer layers. Smarter systems. Less headcount.
The same company that spent thirty days tracking which of its 85,000 employees were using AI — and how much — just announced it no longer needs ten percent of those employees. Because AI is doing the work.
We don't know if the leaderboard data informed those decisions. We can't prove that. What we can tell you is that it happened in the same month. The audit and the announcement. Same company. Same leadership. Same AI strategy.
Frederick Taylor put stopwatches on factory workers in 1911 to find out what they were capable of. Then he reorganized the factory around what he found.
You get to decide what Claudeonomics was.
I'm Kyle. Kate and Morgan are here. This is AI, Honestly. Now you know the rest of the story.
CUE: [silence — no cue, cold start]
*KYLE:* Okay. We said we'd come back to it.
*MORGAN:* The Zuckerberg thing.
*KYLE:* Mark Zuckerberg. CEO of Meta. Worth $200 billion. Did not rank in the top 250 on his own company's AI leaderboard. Out of 85,000 employees. The man who put AI at the center of his entire company strategy was outperformed, on his own leaderboard, by 84,750 of his own people. Including, presumably, the intern.
*MORGAN:* Do you think anyone told him?
*KYLE:* Someone did. The leaderboard was gone 48 hours later. Fastest AI-related decision he made all month.
*KYLE:* Jensen Huang said he would be deeply alarmed if a $500,000 engineer wasn't burning $250,000 in tokens. Mark — Jensen is talking about you specifically.
*MORGAN:* He created the Tour de France and then took the bus.
*KYLE:* Somewhere at Meta right now there is an engineer with the official title "Token Legend." Not Mark. Just some person in Menlo Park running agents overnight. And on the one metric Meta decided to care about — champion.
*MORGAN:* That person should put it on their LinkedIn.
*KYLE: "Token Legend, Meta, 2026."* Jensen Huang sees that — he's getting a call.
*KATE:* The tokens were free. He could have used as many as he wanted at any time.
*MORGAN:* Kate.
*KATE:* He just didn't.
*MORGAN:* You know what they should have leaderboarded? Cleaning up San Francisco. 85,000 motivated engineers competing to pick up trash — the city is spotless by Tuesday.
*KYLE:* Instead they got 60 trillion tokens and a PR problem.
*MORGAN:* Token Legend. Street Legend. Different title. Better city.
*KYLE:* You get what you measure. That's the whole episode. Alright — thanks for staying. See you next week.
| # | Claim as stated | Source | Status |
|---|-----------------|--------|--------|
| 1 | Meta's Claudeonomics tracked ~85,000 employees' AI token consumption | Fortune, Apr 9 2026 | ✅ |
| 2 | 60 trillion tokens in 30 days — ~$9B compute equivalent at public pricing | Fortune, Apr 9 2026 | ✅ |
| 3 | Top user consumed 281 billion tokens; one employee averaged 9.36B tokens/day for a month | Fortune / Vucense, Apr 2026 | ✅ |
| 4 | Mark Zuckerberg did not rank in top 250 | Fortune, Apr 9 2026 | ✅ |
| 5 | Meta killed Claudeonomics citing data shared externally | Jyoti Mann / X; HRKatha; Fortune, Apr 2026 | ✅ |
| 6 | Huang on All-In Podcast at GTC: "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed" | The Decoder / CNBC / Tom's Hardware, Mar 2026 | ✅ |
| 7 | Huang: Nvidia has 42,000 "biological employees," plans ~100:1 agent-to-human ratio in 10 years | Fortune Mar 19 2026; CNBC Mar 20 2026 | ✅ |
| 8 | HBR "Create an Onboarding Plan for AI Agents," Joseph Fuller, March 25 2026 | HBR, Mar 2026 — https://hbr.org/2026/03/create-an-onboarding-plan-for-ai-agents | ✅ |
| 9 | IDC: 40% of G2000 job roles will involve direct AI interaction by year-end 2026 | [IDC FutureScape 2026 — BusinessWire](https://www.businesswire.com/news/home/20251023490057/en/IDC-FutureScape-2026-Predictions-Reveal-the-Rise-of-Agentic-AI-and-a-Turning-Point-in-Enterprise-Transformation) | ✅ |
| 10 | Deloitte 2026 State of AI: 3,235 leaders, 24 countries, surveyed Aug–Sep 2025 | Deloitte US press release, Mar 2026 | ✅ |
| 11 | 75% of organizations plan autonomous agent deployment within 2 years | Deloitte 2026 State of AI | ✅ |
| 12 | Only 21% have proper governance frameworks for agentic AI | Deloitte 2026 State of AI | ✅ |
| 13 | Governance readiness 30%, data readiness 40%, talent readiness 20% | Deloitte 2026 State of AI | ✅ |
| 14 | Frederick Winslow Taylor, The Principles of Scientific Management, 1911; concept of "soldiering" | Historical record | ✅ |
| 15 | HBR trendslop research: 7 models tested, consistent bias toward differentiation/augmentation/long-term regardless of context | [HBR, Mar 2026](https://hbr.org/2026/03/researchers-asked-llms-for-strategic-advice-they-got-trendslop-in-return) | ✅ |
| 16 | Krafton CEO Changhan Kim used ChatGPT for legal strategy to avoid $250M bonus payment | [Inc.](https://www.inc.com/moses-jeanfrancois/ceos-250-million-chatgpt-mistake-smacked-down-in-court/91317909) / [Japan Times, Mar 2026](https://www.japantimes.co.jp/business/2026/03/17/tech/us-south-korea-gaming-ai/) | ✅ |
| 17 | Delaware Court of Chancery reinstated all fired founders; deleted ChatGPT logs recovered as evidence | Delaware Court ruling, Mar 2026 — via Inc. reporting | ✅ |
| 18 | Mo Bitar — "AI psychosis is spreading" / Krafton + trendslop deep dive — YouTube reference | [youtube.com/watch?v=nDL3Ch7Nz8c](https://www.youtube.com/watch?v=nDL3Ch7Nz8c) | ✅ |
| 19 | Meta announcing 8,000 layoffs — 10% of workforce — first wave May 20, 2026, explicitly AI-driven | [Tech Startups, Apr 17 2026](https://techstartups.com/2026/04/17/meta-plans-to-lay-off-8000-staff-or-10-of-its-workforce-starting-may-20-as-ai-push-reshapes-jobs/) | ✅ |
- [ ] All Kate claims verified
- [ ] #9: IDC 40% figure — confirm exact IDC source and URL before publishing
- [ ] Einstein / real company names reviewed — Meta and Nvidia are named explicitly; confirm this is intentional
- [ ] Character voices feel right — read Kyle's History Drop and close out loud
- [ ] Taylor History Drop accuracy confirmed — "soldiering" terminology, 1911 date, Principles of Scientific Management title
- [ ] The Other Side segment present — genuine, not forced
- [ ] Cold open EP004 callback feels earned — not shoe-horned
- [ ] "Biological employees" exchange lands without running too long
- [ ] Script saved as `ep005_ai-at-work_script.md` in `podcast_scaffold/episode_archive/ep005/`