Cuando el flujo se posee de punta a punta, la IA resuelve 45% de tickets en la mediana y 70%+ en el tope — el apalancamiento aparece cuando el criterio se codifica.
When ATMs arrived, banks did not fire tellers — they opened more branches. The first wave of automation often grows the surface area of the job.
Los motores eléctricos llegaron en 1900 pero la productividad solo saltó cuando las fábricas se rediseñaron — el recableado toma una década, no un trimestre.
Spreadsheets didn't kill accounting — they killed the clerk and grew the auditor. Automation reshapes roles upward toward judgment.
Incluso software gratis, infinito y perfecto subiría el PIB solo ~2%: el valor no está en el código, sino en todo lo que el código toca.
Two ceilings are visible at once: ~63% of humanity is online and daily screen time has plateaued near 6.8 hours — new apps now steal attention, they don't add it.
Dos techos a la vez: ~63% de la humanidad está en línea y el tiempo de pantalla diario se estancó en ~6,8 horas — las nuevas apps roban atención, no la suman.
This is not the end of software companies. It is the end of companies whose only moat was that code was hard to write.
Este no es el fin de las empresas de software. Es el fin de las empresas cuyo único moat era que escribir código era difícil.
The research behind this deck
AI is collapsing the cost of code while repricing what a software company is worth. The winners are agentic infrastructure, governed workflows, and judgment; the losers are companies whose moat was that code was hard to write.
Key findings
Spreadsheets did not erase accounting: U.S. BLS projects accountants/auditors growing 5% (2024–34, +73k) while bookkeeping/accounting clerks decline 6% (–94k) — automation reshaped the role, not destroyed it. U.S. Bureau of Labor Statistics
Demirer et al. (2026) find AI sync agents raise lines of code by 741% and pull requests by 65%, but shipped releases rise only 20% — gains attenuate sharply down the production hierarchy. Noahpinion
Anthropic's annualized revenue surged from $87M (Jan 2024) to ~$45B by May 2026, anchored by Claude Code reaching $2.5B run-rate just nine months after launch. VentureBeat
Anthropic reports that code contributed per person rose from a pre-2025 baseline of roughly 1x to 8x by Q2 2026 as Claude became central to internal engineering workflows. Anthropic Institute
Despite ATMs growing from near-zero to 400k+ in the U.S., bank teller employment did not fall: tellers per branch dropped from 20 to 13 (1988–2004), but urban branches rose 43% — preserving teller jobs. IMF Finance & Development
Across RCTs at Microsoft, Accenture and a Fortune 100 firm (4,867 developers), Copilot raised junior/short-tenure developer output by 27–39% versus only 8–13% for senior/long-tenure devs. MIT Economics
The electrification analogy: U.S. manufacturing productivity barely grew in 1909–1919 despite dynamos, then accelerated by ~5.3 percentage points in 1919–1929 once factories adopted unit-drive layouts. Paul A. David / Stanford CEPR
DataReportal trend data show global daily digital screen time hovering around 6.6–7.0 hours since 2020, suggesting that consumer software growth increasingly reallocates attention rather than creating unlimited new hours. DataReportal
Gorgias reports that the median AI-touched ecommerce brand resolves 45% of tickets end-to-end, the top quartile clears 65%, and the highest-performing brands reach 70%+ AI resolution; enterprise discussions cite 60–80% resolution for the largest customers. Gorgias
Global internet adoption has flattened near saturation: from ~7% of the world in 2000 to ~63% in 2023 — leaving little fresh consumer attention for new mainstream apps to capture. Our World in Data
METR's 2025 RCT found experienced open-source devs using AI tools took 19% longer despite expecting a 24% speedup; a 2026 follow-up showed selection bias but only modest –4% to –18% point estimates. METR
Demirer, Musolff, and Yang find that AI coding tools raise commits far faster than shipped releases: autocomplete, interactive agents, and autonomous agents lift coding activity by 40%, 140%, and 180%, while the release effect remains much smaller. National Bureau of Economic Research
The argument
Anthropic went from $87M to ~$45B annualized revenue in 28 months, one of the fastest ramps ever recorded.
Inside the lab the loop is already turning: code contributed per Anthropic engineer rose ~8x by Q2 2026.
Supabase quintupled to a $10.5B valuation in 14 months as agents deployed the majority of its databases.
Median public SaaS multiples collapsed from 19x to ~3.4x, an >80% derating of the old moat.
Productivity leaks before shipping: lines of code rise 741%, pull requests 65%, releases just 20%.
Only 18¢ of every $1 spent on AI coding tokens reaches a shipped product; the rest is bugs, rewrites and review delay.
Copilot lifts junior devs 27–39% but seniors only 8–13%, compressing the skill premium that defined the industry.
History rhymes twice: ATMs preserved tellers, but mobile banking later cut them by 42% — the second wave is the dangerous one.
Software is only ~2% of GDP, so the durable company is the one that owns context, distribution and judgment, not just code.
This research is published in English and Spanish. Ver en español
