Summary

AI makes the legacy firm’s coordination machinery a liability: execution can now be cheaper than the meeting, approval chain, or IT review needed to authorize it. Salim Ismail argues that Ronald Coase’s transaction-cost logic has broken, while companies still survive as a “fiduciary wedge” holding purpose, IP, liability, and human accountability. His defining shift is to organize “around intelligence, not around hierarchy.”
The investable threat is a two- or three-person AI-native team attacking a high-margin incumbent workflow within 60 to 90 days. Ismail’s CEO test is blunt: could “two guys with OpenClaw” reproduce a lucrative business line while the incumbent remains trapped in human-to-human workflows? Diamandis adds the asymmetry: in a large company, one of 20 people can kill an idea; a startup needs only one of 20 investors to say yes.
AI-native does not mean uncontrolled autonomy; it means recursive workflow improvement inside a tightly governed agent architecture. Ismail’s six-layer loop spans purpose, sensing, interpretation, decision, orchestration, and learning, wrapped by trusted evaluations, searchable logs, granular rollback, human review queues, and an “agent passport” defining permissions and liability. Humans move upward into judgment, monitoring, exception handling, and approval.
Ismail expects the average company eventually to operate with roughly 20% to 25% of today’s workforce, with middle management absorbing most of the compression. He says its coordination function could fall about 90%; marketing businesses might reach 10% human staffing, while physical operations remain nearer 25%, and a power plant modeled for Fermi America fell from 800 people to an estimated 80. Diamandis and Ismail frame the alternative to mass unemployment as “five times more companies” and a blossoming of entrepreneurship—but concede that aggressive apprenticeship programs will be needed to rebuild the management pipeline.
Incumbents should not inject AI throughout the cash-generating core; they should construct an AI-native digital twin at the edge and migrate workflows only after parallel validation. Ismail recommends three to five internal outliers plus builders, copied—not moved—data and workflows, recursive improvement, quality checks, and gradual deprecation of the legacy process. “You cannot change and fix and transform the existing company,” he says; Ismail also says that even his influence could not force this through a 100-person organization.
The durable moats become proprietary data, temporary regulatory protection, customer relationships, brand, purpose, and above all a faster learning loop. Ismail estimates a properly operating digital twin could improve workflow performance by “100X or higher per year,” while citing Cognition Labs’ 73-fold ARR growth as an early signal; however, Diamandis notes that visible margins will summon competing agents and drive demonetization. The surviving majority may complete the transition over five to seven years, inside a broader “turbulent transition” they place at two to eight years.

Deep dive

1. AI has reversed the economics that created the modern firm

Ismail starts with Ronald Coase’s 1937 theory: companies grew because coordinating employees internally was cheaper than transacting externally. Later thinkers extended that model, while Exponential Organizations used community, crowds, and AI to stretch the firm’s boundary—Uber’s mission-critical driver-passenger match, for example, happens “out in the wild,” not inside the company.
AI breaks that bargain. An internal website can require meetings, brand approval, privacy review, and an IT veto; outside the company, one person can use Vercel for five minutes, incorporate the brand guidelines, generate a dozen versions, and test them. Salim quotes a tweet that captures the shift: “Building the feature is cheaper than having the meeting about the feature.”
Diamandis asks whether organizations disappear altogether. Ismail’s answer is no: they retain a “fiduciary wedge,” the gap between what AI can execute and where humans must hold judgment and liability. The company increasingly becomes a purpose, legal, fiduciary, IP, asset, and accountability container around agents making external API calls.
Ismail’s diagnosis of current failure is equally categorical: “80-plus percent of AI projects in companies are failing miserably” because firms insert AI into workflows designed as human-to-human approval chains. His analogy is early television merely placing radio announcers on camera—the new medium automates old bottlenecks instead of exploiting a new operating model.

2. The new organization is a governed intelligence loop

The “organizational singularity” replaces hierarchy as the organizing principle with intelligence. Ismail places the massive transformative purpose, or MTP, at the center; DRIVE supplies the intelligence scaffold and SHAPE describes organizational operation. Crucially, the MTP stops being a wall poster and becomes a protocol that constrains both human and AI behavior.
The pair use Uber’s early surge-pricing behavior to illustrate why purpose needs operational boundaries: customers who repeatedly accepted surge pricing could receive a higher price than a cheaper customer standing beside them. Diamandis frames the need for feedback loops that test whether behavior remains “within the cone of the MTP”; Ismail says that ethical boundary is guided in the MTP architecture.
Borrowing from Boyd’s OODA Loop, Ismail defines six layers: purpose, sensing, interpretation, decision, orchestration, and learning. If a retailer’s competitor announces same-day delivery, agents detect it, assess which businesses are threatened, compare responses such as matching the service or buying a startup, orchestrate corporate development and legal work, then learn from earlier acquisitions.
Diamandis notes that strategy officers and marketers once spent months on such work. Ismail says agents can handle the layers, with a human review at the interpretation layer and senior people overseeing agents evaluating six strategic options; a manual operation that might take months can take hours or days. The endpoint is “recursive self-improvement at the workflow level”: an invoice-processing system repeatedly asks how to improve its own loop rather than merely automating fixed checkpoints.

3. Governance becomes machine-readable, continuous, and redundant

The intelligence loop sits inside “govern and assure”: trusted evaluation architecture, a searchable log for every agent, granular rollback, and a human review queue. Humans become dashboard supervisors, monitors, validators, exception handlers, problem solvers, and efficiency designers—not the people manually gathering and repackaging every input.
Each agent receives a passport-like bundle of metadata defining authorized actions, policy-controlled APIs, permitted data exposure, and liability boundaries. If behavior departs from policy, an oversight agent can stop it, notify a human, roll it back, and rerun the evaluation; this is Ismail’s answer to recent agents “doing crazy things,” including the Replit agent that deleted all the volumes of rental-car data.
His redundancy analogy comes from quantum computing: if 1,000 physical qubits are needed to make one logical qubit, relatively free agents can similarly supervise other agents. The additional oversight does not erase the stack’s economics because agents themselves are relatively free.

4. Learning speed becomes the moat as headcount ceases to be one

Ismail’s recurring CEO question is whether “a two- or three-person team with Hermes or OpenClaw” could disrupt a major business line. Diamandis adds the asymmetry: any “juicy margin” is exposed; in a big company, one of 20 people can kill an idea, while a startup needs one of 20 investors to say yes.
The remaining defenses are proprietary data that cannot readily be replicated; regulation, particularly in healthcare, though Ismail warns it can erode; deep customer relationships; brand; and unwavering purpose. The largest is an “intelligence moat”: if the company learns faster than everyone else, competitors struggle to catch its accumulated feedback loop.
Brand and MTP reinforce one another because brand carries the emotional connection to the end user. Ismail says companies should use agents to reinforce that position; Diamandis adds that a dedicated customer relationship can feed proprietary data.

5. Management shifts from coordination to accountability and judgment

C-suite leaders become accountability holders, evaluators, and dashboard overseers. Agents perform strategic analysis; executives contribute experience and “hit yes” or reject the recommendation. Ismail’s broader survival list includes curatorial judgment because “when execution is nearly free, judgment and taste become really important.”
Middle management changes most because its dominant job—collecting frontline data, repackaging it, and coordinating decisions upward—“drops about 90%.” Ismail allocates roughly 60% of total workforce compression to the middle, 20% to frontline roles, and 20% to the top, while redirecting retained people toward exceptions, design, and unsolved operational problems.
His estimate is that an average company could operate with 20% to 25% of its former workforce; later he gives a 10% to 25% range by industry. Marketing could approach 10%, physical operations nearer 25%, while work for Fermi America suggested a power plant might run with about 80 people instead of 800.
Diamandis raises the missing career ladder: without entry-level spreadsheet work, where does future senior management come from? Ismail expects “very active and aggressive apprenticeship programs,” pairing displaced managers directly with leaders such as the CFO. He anticipates guild-like learning and one manager per 20 “high-impact individual contributors,” versus today’s one-to-three or one-to-five ratios.

6. Transformation succeeds at the edge, not inside the cash cow

Ismail’s hardest prescription is: “You cannot change and fix and transform the existing company.” Disruptive work triggers the corporate immune system, so it must be built at the edge and become a new center of gravity. He says that after examining innovation inside roughly 250 Fortune 500 companies, he has “never ever ever ever” seen another method work.
His best specimen is Nespresso: Nestlé created it in 1976 but spent 10 years trying to operate it inside the parent despite a different brand, supply chain, delivery model, and customer proposition. Once separated, friction fell and it became one of Nestlé’s highest-performing lines of business. The same edge logic underpins Skunk Works and other protected teams discussed by the pair.
Ismail supplies firsthand confirmation: he tried to force such a transition inside one of his roughly 100-person organizations and could not, despite his authority, so he created a separate organization. Their governance condition is non-negotiable—the edge unit must report to the CEO, and a board disrupting its own business must explicitly support that CEO.
Ismail’s implementation starts with an AI-native digital twin: assign three to five “crazy young people,” pair them with builders and forward-deployed engineers, copy a defined workflow and fork its data, then run both systems in parallel. Only after recursive improvement and several additional weeks of quality comparison should the old workflow be deprecated and the next one migrated.

7. REWRITE turns the edge strategy into a measured migration

REWRITE begins with backcasting. Rather than extrapolating today’s trucking or retail company forward, management describes how an AI-native company would fulfill its MTP in the target future, then works backward to the required intermediate states. Ismail says this difficult act of abandoning inherited assumptions is unusually easy to explore with a large language model.
The company then scores itself across seven dimensions. Two examples are organizational drag—whether action passes through five or six approvals—and whether AI is a first-class organizational capability or merely an IT-injected tool. A chief AI officer and native capability score higher than a thin layer of AI over legacy operations.
Next comes workflow documentation, including tacit knowledge that an experienced employee performs but never records. Ismail says some firms are trying to “shadow” workers with agents, provoking an immune response: he cites 44% of Gen Z workers as sabotaging AI with bad information so it cannot later take their jobs. Retraining therefore belongs inside the migration, not after displacement.
Before digitizing, management should reduce a 10-step approval process to perhaps three steps. The target stack replaces siloed ERP data with an accessible data lake, permissions attached to each data object, custom AI-built applications and workflows, then agents above them. Ismail says this owned architecture threatens SaaS providers whose position depends on being embedded in the legacy stack.

8. The transition is already visible, but its endpoint is a moving target

Ismail estimates that a properly operating digital twin should deliver “100X or higher” performance per year—processing 100 invoices where the old system handled one, or reducing 100 days to one. Contact centers and marketing/content generation are the clearest sectors where he says the full progression from human-heavy, to AI-assisted, to AI-native is happening; the third EXO book itself took three months versus three years for the first and two-and-a-half years for the second.
Diamandis frames restructuring as a one- or two-year imperative, not a five- or 10-year issue. Ismail clarifies that his five-to-seven-year estimate concerns the surviving majority of companies completing the full transition, nested within a broader two-to-eight-year “turbulent transition.” A first engagement, by contrast, is intended to establish several working edge workflows in roughly 90 days.
What survives is an MTP encoded as protocol, the legal-accountability shell, proprietary intelligence, coordination protocols, and human judgment. What dies is the static org chart, five-year plan, quarterly review as the decision unit, and annual planning; the organization instead changes “like an amoeba” until “the organization itself becomes a protocol.”
Ismail cites Cognition Labs’ 73-fold ARR growth after going fully AI-native as an early signal, while acknowledging that excess profitability attracts immediate imitation. Diamandis connects that competition to demonetization and ultimately universal high income: companies may deliver 100 times more cheaply, but agents will attack any margin they expose.
The model extends beyond corporations. Ismail says Sheikh Mohammed wants 50% of the Emirati government run this way and points to golden visas processed in five hours; universities are also approaching his group as teaching moves from content toward execution, where an engineering credential could reflect what a student built rather than four years spent studying.
Because “every two, three days we’re learning new things,” the Organizational Singularity book is planned as a downloadable Claude skill rather than a static publication. That form embodies the thesis: the framework must learn continuously because “the organizational singularity is here. It’s just not evenly distributed.”

Curing Every Disease With Al by 2050 | Sam Rodriques, Edison Scientific Your Biggest Lever: Designing your AI Career for Maximum Impact, with 80,000 Hours founder Ben Todd