the full research paper behind the great repricing: two organizational archetypes, what transplants and what doesn't, the cost model with every assumption's basis, and the four findings that would falsify the thesis.
We believe that in a world of abundant, rapidly cheapening machine intelligence, companies whose operating cost is built on human coordination will not be able to compete on cost with companies whose operating cost is built on compute — and that because markets reprice to the cost structure of the cheapest capable producer, most coordination-built companies will not survive the repricing.
We argue it by comparing two archetypes. The legacy archetype: a large, mature enterprise — thousands of employees, work coordinated through meetings, status chains, and approval hierarchies. The AI-native archetype: a real company with a small team — a handful of employees where the incumbent staffs hundreds — directing AI agents that hold real operating jobs, coordinated by machines against a shared ledger. The legacy archetype is characterized from published research; the AI-native archetype from a documented cost model; the competitive consequence is tested against SEC filings of two incumbents in one consumer market (Match Group/Tinder and Grindr). We state every estimate’s basis, show the dials, and close with four findings that would falsify the thesis.
Three kinds of evidence: published research (cited in Sources), public filings, and a documented cost model whose parameters come from current vendor pricing and measured production-scale prompt and usage data. Modeled figures are labeled and stated as magnitudes. The archetypes are patterns: no specific company is described or identifiable, and qualitative characterizations draw on the authors’ professional experience across large organizations only at the level of pattern, never instance.
This is not a controlled experiment; real companies sit between the poles. The claims are about cost structure and operating mechanics — checkable now — not about which companies win.
| AI-native archetype | Legacy archetype | |
|---|---|---|
| Humans | A handful of employees; tens at large scale | Thousands |
| “Staff” | AI agents with charters, autonomy levels, computed scorecards | Layers of roles coordinated by meetings and status chains |
| Coordination | Machine-enforced claims, handoffs, release authority on a shared ledger | Meetings, status chains, approval layers — ≥25% of supervisory time on synchronization (Bain) |
| Status latency | Seconds | Days to weeks |
| Decision record | Durable by default | Reconstructed from memory and minutes |
| Quality posture | Machine gates that can’t be skipped; defects can’t close without evidence | Quality debt paid as remediation after the fact |
| Operating cost | Compute-dominated at every scale | Payroll-dominated at every scale |
The dimensions collapse into one finding: the archetypes pay to know things in different currencies — payroll for human synchronization, or compute.
In the AI-native archetype a role is a charter plus an autonomy level plus a computed scorecard, held by software: a release captain owns the deploy lane, a fixer drains the defect queue, a finance loop tracks spend against hard caps. Promotion up the autonomy ladder is earned on scorecard evidence; demotion on a breach is automatic. Minting a role-holder costs the time to write its charter, so no seat stays vacant and no one triple-hats.
In the legacy archetype, role-minting runs through hiring, promotion, or consultants — slow enough that vacancy and overload are the steady state; key-person concentration is among the most commonly logged risks in enterprise programs of every kind.
The legacy coordination bill is measured. Bain (25 corporate audits + an EIU survey of 300+ executives): the average company loses more than 20% of its productive power to organizational drag, with ≥25% of a typical line supervisor’s time consumed by unnecessary meetings and e-communications. Hamel and Zanini, cited in the same work, put US bureaucracy cost above $3 trillion a year. And it compounds: time to complete critical business tasks (hiring, IT projects, B2B contracting) grew significantly 2010–2015.
The AI-native archetype pays the same tax in compute at machine latency: work claimed in a ledger before it starts, handoffs as structured records verified against the work, status written where work happens, machines that escalate on recurrence instead of waiting. None of this is exotic; it’s what a small team directing many agents is forced to build, and what an organization of thousands is structurally able to avoid building.
Published averages are self-reported estimates, and many legacy organizations beat them. AI-native organizations fail in new ways (blind metrics, runaway loops); the advantage is a faster, cheaper failure-metabolism, not immunity. And no AI-native consumer company has yet demonstrated durable success at incumbent scale — the claims here are cost mechanics, not market outcomes.
Ten opportunities scored on impact and difficulty (1–5), where “adopt” means default operating mode, not a pilot:
| # | Opportunity | Impact | Difficulty | Quadrant |
|---|---|---|---|---|
| 1 | Living knowledge base with a freshness contract | 4 | 2 | Quick win |
| 2 | Decision capture with durable approvals | 4 | 2 | Quick win |
| 3 | Governance overhead measurement and caps | 3 | 2 | Quick win |
| 4 | Ledger-based status instead of collected status | 5 | 3 | Strategic bet |
| 5 | Escalation-by-default automation | 4 | 2–3 | Quick win |
| 6 | Machine-gated quality lifecycle | 5 | 4 | Strategic bet |
| 7 | AI agents holding real jobs | 5 | 4 | Strategic bet |
| 8 | Autonomy ladder with never-silent gates | 4 | 3 | Strategic bet |
| 9 | Humans-operate-agents talent model | 5 | 5 | Moonshot |
| 10 | Compute-priced cost structure | 5 | 5 | Moonshot |
Three observations:
1. The quick wins are real — and they are not the moat. Items 1, 2, 3, 5 are weeks of policy plus thin tooling; they make an organization better-informed within two quarters. They change nothing about cost: the status report gets rendered instead of authored, and its authors are still employed, still coordinating.
2. Difficulty is organizational, everywhere. Each strategic bet de-jobs the constituency that performs the function manually — usually the constituency assigned the rollout. Base rates: 30% transformation success (BCG, n=900); 95% of enterprise GenAI pilots with no measurable P&L impact against $30–40B invested (MIT, 2025). Enterprises aren’t failing to buy AI; they’re failing to convert it into operating structure.
3. The ladder can’t be skipped. #10 requires #9, which requires trustworthy agents and a ladder for trusting them (#7, #8), which require machine-enforced quality (#6), which requires written organizational context (#1, #2 — agents run on written knowledge). Each rung takes quarters at enterprise scale, and the bottom rungs aren’t practices at all — they’re starting conditions.
| Match Group | Grindr | AI-native entrant (illustrative — modeled at Grindr’s scale) | |
|---|---|---|---|
| Revenue | $3,487M | $439.9M | ~$440M (illustrative: Grindr’s monetization applied to 15M MAU) |
| Employees | ~2,200 (after a 12% cut) | ~151 | ~15–30 (modeled) |
| Total operating cost | $2,615M | ~$244M implied (revenue − adj. EBITDA $195.6M) | ~$80–130M today, falling with inference prices (modeled, ex-growth-marketing) |
| Profitability | $872.5M operating income | 44.5% adj. EBITDA margin | implied ~70–80% operating margin (illustrative) |
| Monthly actives | 13.8M payers portfolio-wide | 15.0M MAU | 15M (by construction) |
| Cost per MAU-month | — | ~$1.36 | tens of cents |
| Cost per payer-month | ~$15.79 (vs $20.72 revenue/payer) | — | — |
| Revenue per employee | $1.59M | $2.91M | ~$15–30M (illustrative) |
The entrant column is a modeled archetype placed at Grindr’s scale and monetization for like-for-like comparison — not a real company’s results; its cost basis and assumptions are §3.2. Grindr is the bar — the leanest that legacy-structured looks in this market. Any argument that survives Grindr survives everyone.
Payroll scenario. Anchors: Grindr ~1 employee per ~100K monthly actives; the category leader ~1 per ~6K payers; WhatsApp, pre-agents, ~1 per 7.6M (55 employees, 420M users, 2014). We assume 1 human per 500K–1M users — 10x Grindr’s leverage, well short of WhatsApp’s: ~15–30 people at Grindr’s scale, ~50–100 at category-leader demand, for trust-and-safety, legal, payments, partnerships, and operating the agents.
Cost mix, evidence first. Published measurements: compute is 57–70% of total spending at leading AI companies (Epoch AI) and 40–50% of revenue at AI-first startups — companies with staffs many times larger than this archetype. Our leaner model lands at the top of the range, 70–90% of serve-cost; we treat the published band as the floor, the model as the payroll-light ceiling, and show below that the thesis holds across the range. Growth marketing is excluded from serve-cost on both sides (both pay market rates for attention; it returns in 3.6).
Result, in magnitudes. At product parity the modeled cost is tens of cents per user-month all-in — already below the leanest incumbent’s ~$1.36. At AI-rich depth (far more inference per user — the product worth building), per-user cost is comparable to the incumbent’s today and passes below as deflation and standard optimizations land: projected crossing ~2027, give or take a year on the dials. At Grindr’s scale: low-hundreds of $M/year unoptimized, under $100M optimized, plus ~$6–12M payroll — a 2–3x gap today. Real, but survivable. The finding is the derivative.
Frontier-equivalent inference deflates ~10x/year (a16z: ~1,000x over three years), with Epoch AI measuring 9x–900x across tasks, median 50x. Against the two mixes:
Two-year projection at Grindr’s scale, payroll included:
| 2-yr inference deflation | Modeled entrant, cost/yr at 15M MAU | Incumbent trajectory (FY2026 guidance implies ~$311M) | Gap |
|---|---|---|---|
| Bearish: 3x total | ~$35–40M | ~$300M+ | ~8x |
| Base: 9x | ~$20–25M | ~$300M+ | ~13x |
| Historical: 100x | ~$10–15M | ~$300M+ | ~25x |
Robustness: re-run at the bottom of the published mix range (50% inference) and the two-year gap compresses to ~3–5x — still widening yearly. The thesis needs the direction of the curves, not the top of the range.
Distribution adds a structural line: the category leader’s cost of revenue is dominated by app-store fees (~27 points of revenue); a web-billed entrant pays 3–6. Rules are loosening; the default rail still costs ~20 points the entrant never pays.
The divide has a behavioral mechanism too: the archetypes get paid back differently for the same token.
In the legacy archetype, inference is a cost bolted onto an operating model that can’t convert it: 95% of pilots without P&L impact (MIT); 42% of companies abandoned most AI initiatives in 2025, up from 17%, with an average 46% of proofs-of-concept scrapped before production (S&P Global, n>1,000, cost among the top obstacles); budgets up while, per Forrester, most organizations can’t demonstrate sustained return. The rational response — visible in caps, per-seat token rationing, deferred spend — is throttling. When tokens don’t return, throttling is correct.
In the AI-native archetype, inference is the product, so every token faces a margin test, and the incentives run the other way twice: spend concentrates where tokens earn, and waste is hunted structurally (caching, routing, lean context as survival disciplines). Same input, opposite management problem — which is the deepest reason “just adopt AI” does not close the divide. The divide is not who pays less for tokens; it is who gets paid back for them.
Breakeven floors: Grindr needs ~$1.36/MAU-month; the modeled entrant needs tens of cents today, on the order of a dime at two years. To stay margin-whole at half its monetization, Grindr must cut ~50% of cost — at 151 people. To break even at the entrant’s two-year floor: ~89%. Category-leader equivalents: ~$1.3B and ~$2.3B of a $2.6B base.
Cutting is not transforming: remove a third of the people and the coordination model remains, minus its operators. The base rate for far shallower change is 30% (BCG). The category leader’s most aggressive recent action was a 12% reduction — while deploying 108% of free cash flow into buybacks and dividends as paying users declined. Rational margin defense, and precisely the strategy that spends the transformation window: transformation is cheapest today (the incumbent is strongest) and most painful today (it cannibalizes a working business); every quarter of waiting raises the price and shrinks the payer.
Consumer marketplaces are liquidity businesses: cheap compute conjures no users, network effects defend the core, and the entrant bids in the same acquisition auctions (the category leader spends $626M/year on S&M). What the structure buys: a far larger share of revenue available for growth at equal margin; profitability at scales incumbents can’t staff; and survivability — with non-payroll costs in the hundreds of dollars a month and a handful of people, an entrant can’t realistically die of burn rate, so the incumbent faces a permanent population of small, patient challengers.
The model can also be wrong in the entrant’s disfavor: deflation could stall; richer product could eat savings; per-user AI costs are chronically underestimated (our own model’s history includes a >10x upward revision on measurement — why figures here are magnitudes). At the bearish setting of every dial the gap is still ~3–5x and widening; the dials are real.
Claim 1 — the operating-model difference is structural. Confidence: high. The legacy side is independently documented (>20% drag, ≥25% supervisory synchronization time, remediation-dominated quality spend); the AI-native side is constructible today by a small team at non-payroll costs in the hundreds of dollars a month.
Claim 2 — the top of the stack transplants; the bottom doesn’t. Confidence: moderate-to-high. Knowledge, capture, and escalation practices transplant in weeks. The talent model and cost structure are starting conditions atop an unskippable dependency ladder, and the base rates (30% transformation success; 95% of pilots without P&L impact; 42% abandonment) describe failure at far shallower change.
Claim 3 — the gap widens on its own, and incumbent incentives accelerate the loss. Confidence: moderate; this is the projection. 2–3x today at equal scale; several-fold to an order of magnitude in two years depending on mix and dials; and every incumbent behavior visible in filings (margin defense, capital return, single-digit trims) defends the present rather than paying for the transition while it’s cheap.
The verdict, as we’d bet it: in the abundance era, coordination-built companies will not compete on cost with compute-built companies, and most will not survive the repricing — managed decline run quietly to completion, not bankruptcy on a date. A minority will buy years by getting radically lean and adopting what transplants. The starting condition — a cost base riding the deflation curve — cannot be adopted. It can only be founded.
Review: re-run against FY2027 filings and production-scale AI-native economics, mid-2028.
A large company pays hundreds of people and dozens of meetings a week to move information from the person who has it to the person who needs it in time for the decision. An AI-native company implements that same sentence as a ledger write. Price the difference honestly, compound it through several more years of inference deflation, and the arithmetic is not close.
Disclosure. This paper is not investment advice and expresses no view on any security. The authors hold no positions, long or short, in any company named. Statements about named companies are conditional arithmetic on their public filings; forward-looking statements are the authors’ opinions, with falsification criteria stated above.
All external figures verified July 2026. Primary or primary-equivalent (fetched directly): Match Group and Grindr results/filings figures, Bain (book first chapter read directly), a16z, Epoch AI price-trend data, CISQ report figures, The Conversation (WhatsApp). Via named secondary coverage (marked below): S&P Global (CIO Dive), Epoch AI compute-share data (Visual Capitalist), Forrester (Computerworld), MIT NANDA (legal.io summary). Derived ratios (cost per user/payer, revenue per employee, implied operating cost) were re-computed from the underlying filing figures before publication.