Newsletter · · Ashutosh Agarwal
Meta Asks Its Managers to Come Back While Founders Keep Deleting Work - How They Build - Week of September 19, 2026
How They Build for the week of September 19, 2026: Meta asked staff in its 7,000-person applied AI division to volunteer to return to management and capped managers at around 20 direct reports, while Kiteworks ran acquisition due diligence for roughly 5,000 dollars in tokens instead of 500,000 to 700,000 in outside counsel, a founder runs 90% of his go-to-market inside Claude, and Retool's CEO said about 90% of token spend is probably negative ROI.
How They Build
Week of September 19, 2026: Meta Asks Its Managers to Come Back While Founders Keep Deleting Work
The flat-org dream hit a wall this week, while the operators quietly kept deleting the work anyway.
For a year, the story in this newsletter has been one long demolition: fire the managers, delete the middle, run the company with a skeleton crew and a swarm of agents. This week, the biggest AI builder on earth quietly hit the undo button, and a couple of podcast hosts noticed. But go one level down, to the founders actually running things, and the demolition never slowed. A 600-person company just did the legal diligence on an acquisition for the price of a used car. A founder now runs 90% of his sales operation inside a chat window. And a guy selling used iPhones in Norway staffed his entire back office with bots he named like real employees.
The Number: $5,000
That is roughly what it cost the legal team at Kiteworks, a ~$200 million-revenue cybersecurity company with 550-600 employees, to run the due diligence on its most recent acquisition. Entirely in-house. Entirely on AI.
General Counsel Camilo Artiga-Purcell walked through the math on The Legal Department. His team is 14 people total, with just four lawyers touching M&A. Historically, the legal due diligence alone on one of these small-to-mid-sized deals ran half a million to $700,000 in outside-counsel fees, junior associates at big firms "billing you 500 to 800 [dollars] plus an hour reading through" employment contracts and data-room documents.
Instead, his four lawyers now supervise what he calls thousands of AI "associates." They point Claude at the data room, iterate on a detailed prompt until it's airtight, then let it produce a red-flag report and issue-spotting checklist. Humans spot-check at random across every category (IP, employment, tax, customer contracts) until they're "95-plus percent confident" it's accurate.
"This whole process will cost in tokens, when we're all done, probably somewhere in the order of $3,000 to $7,000... but historically, the spend on outside counsel, even for a small to mid-sized deal, easily the legal due diligence alone is somewhere in the order of a half a million to $700,000."
That's a roughly 100x cost collapse on one of the most expensive, most defensively lawyered tasks a company ever does. And the pressure to do it came from the top: "It's from the CEO and the board saying, you have to leverage these technologies. You have to identify inefficiencies. Get out of your comfort zone and push forward."
The Legal Department, "Small But Mighty: Using AI to Scale Your Legal Department" (Sep 17, 2026)
What Founders Changed
A founder runs 90% of his go-to-market inside Claude, and learned to fire his own agents. The founder of BackEngine demoed his actual setup on the GTM AI Podcast: "90% of what we do on the go-to-market side is actually happening inside of Claude." He's wired it into Fireflies, Gmail, HubSpot, Notion, Slack, Zoom and his own home-built connectors, so he never leaves the chat window. Scheduled agents run the operation: a "Founder Sales Daily Pulse" fires every day at 4pm and emails him everything that happened across outreach; another reviews every batch of 20 prospect calls and tells him what's landing and what isn't ("probably the highest value email I get"). The interesting confession is what he stopped doing. He once had 200 agents firing at once, "and then the second that happens, no one's reading them." He cut it to about 10. His lesson sounds exactly like managing people: "Everyone likes to hire as many people as they can, but they'll only do a great job if you have the time to actually devote to each of them... less is more." Each surviving agent has an owner and a real workflow attached, or it gets killed. GTM AI Podcast, "200 AI Agents to 10: How a Founder Runs 90% of GTM in Claude" (Sep 16, 2026)
A used-phone seller built an org chart out of bots, and gave them names. Phil Atkinson runs Greenphones, a seven-figure used-iPhone brand in Norway that buys $500,000-$600,000 of phones a month from private sellers and wholesalers. His physical team is tiny, a handful of technicians and salespeople who "do the physical selling." Everything else is staffed by what he openly calls "AI employees." He wrote job descriptions the way you would for a real hire, fed them into GrokBot, and now has a "chief of staff" named Max with four direct reports: Eric, Simon, Matt and Sally (Sally runs SEO and site research). They introduce themselves and even "discuss the business with each other." The work they've absorbed (a chief marketing manager pulling from a "market intelligence officer," building the ad creative, optimizing Google and Meta campaigns) "would normally be done by probably two or three people from the creative side, the analytic side." The tooling bill for the bot layer: about $30 a month for GrokBot. Winning With Shopify Podcast, "How I Built A 7-Figure Shopify Brand With AI, Automation & Almost No Team" (Sep 14, 2026)
Hinge fused its two most senior tech jobs into one, but refused to gut the layer below. Ben Celebicic was promoted from CTO to Chief Product and Technology Officer at Hinge, absorbing a role a separate executive used to hold. He now runs a team of nearly 300 of Hinge's ~360 people, and the company is chasing $1 billion in revenue in 2027. It's a real example of the C-suite collapsing: one person now owns the product-versus-engineering fight that used to happen between two executives ("I had to have the tension with myself"). But he was pointed about where he draws the line: he deliberately keeps strong, separate leaders for engineering, AI, data, product, design and research so they still argue prioritization in the room. When those debates go quiet, he takes it as a warning: "That kind of lights a bulb in my head that something is wrong." Collapse the title at the top; keep the friction in the middle. The Product Podcast, "Hinge CPTO on Building an App Designed to Be Deleted, Merging the CPO and CTO Roles" (Sep 16, 2026)
EY is teaching all 400,000 of its people to build agents, using an agent. On HR Leaders, EY Chief Learning Officer Simon Brown described "Agents of Change," a program where every one of the firm's 400,000 employees learns to build AI agents through hands-on practice with an agent named Maestro, which teaches you how to build agents. In each 60-90 minute module you define a real use case (his example: podcast research), build an actual working agent with Maestro's help, and deploy it into Copilot. 130,000 people have gone through it since May. It's the "everyone builds" idea most founders talk about, executed at a scale that would have been unthinkable a year ago, turning agent-building from a specialist skill into a company-wide baseline competency. HR Leaders, "How EY Is Upskilling 400,000 People for the Age of AI" (Sep 16, 2026)
A law firm is inventing the "legal engineer." A Chief AI Officer on Legal Innovation Spotlight described deliberately collapsing the traditional middle layers of a law firm and standing up roles that simply didn't exist before: a "legal engineer," plus teams organized around data, applied science and applied AI. Her framing captures where a lot of knowledge-work orgs are heading: does "knowledge management" quietly become "knowledge engineering"? (Her wry aside: law firms didn't even have C-suites until about five years ago, and now they're redrawing the whole chart around AI.) Legal Innovation Spotlight, "Your Org Structure Is Your AI Strategy" (Sep 16, 2026)
The Other Side
This was the week the mood turned. Not because the tools got worse, but because the biggest builders started admitting what the flattening actually costs.
Meta is quietly begging its managers to come back. This is the story of the week. As broken down on the Elon Musk Podcast, Meta is asking staff in its 7,000-person applied AI division to volunteer to return to management, directly reversing the exact flattening it spent enormous energy on, when it stripped managers of their titles and shoved them into individual-contributor roles. The tell is that it's opt-in: they're asking people to go back. Because you can delete a manager's box on the org chart, but you can't delete the work. The coordination, the cross-team dependency tracking, the coaching a junior engineer through a hard call: all of it "lands somewhere. Usually it lands on your most expensive, most experienced technical specialists," who end up doing "two full-time jobs": their real work by day, "shadow management work... [that] spills into their evenings and weekends."
The remedies Meta promised read like a confession. It agreed to cap managers at around 20 direct reports, which means, as the hosts put it, "you can safely assume some managers were carrying 30, 40, or even 50 direct reports." A 50-to-1 ratio is arithmetic that doesn't close: 30 minutes a week with each of 50 people is 25 of a manager's 40 hours gone before any actual work. It promised to stop reshuffling people between managers so often, to let trapped engineers apply for other roles, and, tellingly, "to restock the micro-kitchens." Their verdict: "The entire industry essentially copied the flat organizational model" like a toggle switch, and treated a complex human structure as a setting on a dashboard. Elon Musk Podcast, "Why Meta is bringing back middle managers" (Sep 12, 2026)
Oracle's giant layoff isn't the AI story it looks like. Oracle cut roughly 21,000 jobs (about 13%) and booked a ~$2.1 billion restructuring, headlines that got filed under "AI replaces workers." But on Future Ready Leadership, Jacob Morgan made the un-sexy point: this is "AI CapEx displacement," not automation. Oracle is trading payroll for concrete and infrastructure to fund a roughly $100 billion data-center buildout, cutting people to pay for the buildout, not because software ate their jobs. Two different stories that keep getting collapsed into one. Future Ready Leadership With Jacob Morgan, "Time to Kill AI? Anthropic Wants AI Slowed, Oracle Cuts Jobs" (Sep 14, 2026)
The AI-native darlings are hiring like crazy. For all the "lean team" evangelism, Go Nimbly CEO Jen Igartua, who works with AI-first companies like Fireworks, Perplexity and Exa, said on Topline that these companies are hiring huge numbers of people annually and growing "despite their operations," not because of some elegant agent stack. Typically there's one overwhelmed ops person "who gets rapidly replaced by a small specialized" team. The fastest-growing AI companies are, in practice, adding humans fast. Topline, "RevOps for Hypergrowth (Fireworks AI, Perplexity, Exa & More)" (Sep 13, 2026)
The Cost Corner
The spending reckoning got a blunt new headline number this week, and a clear playbook for what to actually do about it.
"About 90% of the tokens you're spending are probably negative ROI." That's Retool CEO David Hsu, quoted on the AI Proving Ground Podcast from a Wall Street Journal executive summit, the line that everyone will screenshot. "It's just the last 10% of tokens that are really driving a lot of that ROI." The host noted his own firm, World Wide Technology, has watched its token bill climb from a few hundred thousand dollars a month earlier this year to well over $2 million now, "and that number may be even growing as I speak." WWT CTO Mike Taylor offered the fix, and it's not about picking a cheaper model: "If you're asking me to save $100 on tokens, I'm going to get 90 cents of that by building out a data-ready architecture." The remaining dime comes from publishing a shared "skills library" so hundreds of employees aren't all re-solving the same problem from a blank prompt and burning tokens while the model "just churns." AI Proving Ground Podcast, "AI Slowdown Calls Mount as Enterprises Push Ahead With Agents" (Sep 19, 2026)
The cheap-model math is now impossible to ignore. On Everyday AI, the specifics were laid out plainly. DeepSeek's V4 Pro model lists at $0.43 per million input tokens and $0.87 per million output, "more than 25 times cheaper than the premium closed-source models." Translated to a real budget: a task costing $40,000 a year on a frontier model could cost about $1,000 a year on an open one. A 100-agent swarm that runs "$1,200-plus on Opus" drops to "maybe like $60 on DeepSeek." And capability has gone local: Google's Gemma 4 delivers roughly the quality of a top frontier model from 14 months ago, now running "on a consumer laptop" for free. The move is to stop treating AI as one-model-fits-all and route high-volume grunt work (summarization, extraction, parsing PDFs, classification) to the cheap stuff. The catch worth flagging: using open or foreign models via API "strips away all of that legal protection" you get from the big US labs, a cost that's real even when it doesn't show up on the invoice. Everyday AI Podcast, "Open Source AI 101: Why Local Models, Cheap APIs, and AI Agents Change Everything" (Sep 18, 2026)
The floor keeps dropping. For context on just how fast the price of intelligence is falling: on The Twenty Minute VC, Positron co-founder Thomas Sohmers noted token prices have gone from roughly $60 per million to about $1 per million over five years. The DHUnplugged hosts tallied the current spread the same week: a premium model around $12 per million, DeepSeek's cheapest tier at 66 cents, and, on the raw hardware side, GPU time at "$8 an hour" versus 20 cents an hour for cheaper alternatives. The through-line of the whole Cost Corner this week: the sticker price of a token is collapsing, but the bills are still exploding, so the discipline has moved from "which model" to "how much of this were we ever going to read." The Twenty Minute VC, "How Many Planned Data Centers Will Actually Get Built?... Thomas Sohmers, Positron" (Sep 19, 2026); DHUnplugged Podcast, "DHunplugged #818: Cyberdyne IRL" (Sep 16, 2026)
The bottom line this week: the org-chart demolition finally met the person who has to clean up after it. Meta learned that deleting managers doesn't delete management; it just hides it in your best engineers' evenings. But the founders one rung down aren't slowing down: they're doing $600,000 of legal work for $5,000, running sales out of a chat window, and staffing the back office with bots that have names. The lesson isn't "flatten everything." It's "delete the work, not just the boxes on the chart."