How many ATMs does it take to put a bank teller out of work?
In the early 1980s, the consensus answer was: many. Mainstream commentary predicted ATMs would eliminate one-third of bank teller jobs by the end of the decade. Automation was coming. The teller was a doomed profession.
Here is what actually happened. The number of bank tellers in the United States grew from roughly 500,000 in 1980 to over 600,000 by 2010 — a 20% increase, across the very period when ATMs were supposed to wipe them out. The economist James Bessen documented why. ATMs made it cheaper to operate a bank branch. Cheaper branches meant more branches. More branches needed more tellers — though the work shifted from cash handling to relationship management, sales, and advisory.
The automation didn't replace the workers. It multiplied the surface area on which they could be deployed.
This pattern is so consistent across every major technology shift of the last 200 years that it has a name in economics: induced demand. When a thing gets cheaper to produce, total consumption rises faster than per-unit cost falls. And in almost every documented case, total employment in the affected industry expands.
The current AI labor narrative is built on the opposite assumption. Every week brings new layoff headlines: tens of thousands at Amazon, Microsoft, Salesforce, Meta, Google. The accompanying analysis is uniformly grim. AI is coming for your job. The white-collar bloodbath is here. Knowledge work is over.
I think this is the wrong story.
Not because AI isn't displacing jobs — it is. But because the displacement is the visible part of a much larger transformation, in which the creation of new knowledge work is going to dwarf the elimination by a factor that's hard to comprehend until you walk through the math.
The honest forecast: the next decade will produce the largest expansion of knowledge work in human history. The number of people doing high-end, software-adjacent, judgment-intensive work will not shrink. It will at minimum double, and plausibly grow ten to twenty times.
The current layoffs are not the beginning of contraction. They are the noise before the largest hiring wave the world has ever seen. And almost nobody — including the companies that will need to do the hiring — is preparing for it.
Let me show you the math.
I. The Historical Pattern Is Unbroken
Every major automation wave of the last 200 years has been predicted to destroy the labor market. Every one has expanded it.
The mechanical loom (1810s). Predicted to eliminate weaving as a profession. Actual outcome: textile employment in England grew roughly 4× over the next century. The price of cloth collapsed by 95%, demand exploded, and textiles became one of the largest employment sectors of the industrial economy.
The typewriter (1880s). Predicted to make clerks obsolete. Actual outcome: the typewriter created an entirely new profession — the typist — that grew to employ millions in the early 20th century. Total clerical employment grew, not shrank.
The ATM (1980s). As above. Bank teller employment grew 20% during the period of maximum ATM deployment.
The spreadsheet (1980s-90s). Predicted to eliminate bookkeepers. Actual outcome: bookkeeping employment in the U.S. grew through the 1990s and 2000s. The spreadsheet didn't replace bookkeepers — it allowed every business to produce 10× more financial reporting, hire actual finance teams, and develop entirely new disciplines (FP&A, controllership, financial planning) that didn't previously exist.
The internet (1995-2010). Predicted to eliminate retail, real estate agents, travel agents, and journalists. Actual outcome: e-commerce employment alone is now larger than physical retail employment was in 1995. Online travel created Booking, Expedia, Airbnb — collectively employing more people than the entire travel-agent industry ever did. (Journalism is the partial exception. A story for another piece.)
The pattern is so consistent that economists treat it as a working law: when a technology lowers the cost of producing X, total spending on X-adjacent work expands rather than contracts, because demand for X was previously suppressed by cost.
This is the most important framing for understanding what AI is about to do to knowledge work.
II. The SaaS Unbundling
Here is the specific math that is going to surprise people.
Today, the global SaaS industry is roughly $300 billion in revenue, served by approximately 30,000 SaaS companies, employing somewhere in the range of 1.5 to 2 million people. Salesforce, Workday, ServiceNow, Atlassian, Notion, Slack, Zendesk, Monday, Asana, Airtable — these companies exist because building software is expensive, and the only way to make it economical is to amortize the cost across thousands of customers paying monthly subscriptions.
That economic logic is breaking.
A founder with Claude, Cursor, and modern cloud infrastructure can ship in 6 months what a 50-person SaaS company shipped in 3 years a decade ago. The cost of producing a competent internal CRM, project tracker, support tool, or HR system has fallen by roughly two orders of magnitude in 36 months.
When the cost of building falls 100×, the build-vs-buy calculation flips for an enormous range of use cases. Companies who used to pay Salesforce $200 per seat per month for a tool that doesn't quite fit their workflow will instead spend 6 months and 3 engineers building exactly the tool they want. Companies who used to bend their processes around Jira will build Jira-shaped-for-them. Companies who paid for ten different SaaS tools to cover slightly mismatched needs will build five integrated ones that fit perfectly.
To be clear: SaaS doesn't die. It bifurcates. Horizontal, best-practice, system-of-record SaaS survives — Slack, Notion, the Salesforce-as-CRM core, Workday-as-HRIS. These products encode 25 years of accumulated wisdom plus network effects plus integration ecosystems. You don't rebuild those.
What unbundles is the long tail of vertical, process-specific, internal-workflow SaaS. The thousand tools that exist because somebody once needed a slightly different way to track project budgets in a construction firm or manage clinical-trial data in a CRO. That long tail comes home, gets built internally, gets shaped to the company.
Now run the numbers.
Before: 30,000 SaaS companies × ~50 people each = 1.5M people building horizontal software for 10M customer-companies.
After: 500,000 mid-sized companies globally × 5-10 internal software products each × 5-10 engineers per product = somewhere between 12 and 50 million people building internal software. Best estimate: 20-25 million.
That is a 10-15× expansion in people doing software-development work, just from the unbundling of SaaS into vertical, custom, AI-built internal tools.
But this is only the obvious move. The real expansion is bigger.
III. The Long Tail of Software That Doesn't Exist Yet
The deeper truth is that most of the software that should exist has never been written.
Every dental practice, law firm, manufacturing plant, restaurant chain, construction company, school district, and city government has a dozen processes that today live in spreadsheets, sticky notes, tribal knowledge, and human memory. None of these processes have ever been encoded in software, because no SaaS company would ever build for them — the TAM is too small, too specific, too weird.
Marc Andreessen's observation that "software is eating the world" was made in 2011, when software was eating roughly 10% of the world. The other 90% was uneconomical to build for.
AI changes the economics. When building software costs $50,000 instead of $5 million, the long tail of "software that should exist" becomes economical. The TAM of software development expands by an order of magnitude, perhaps two.
If today there are roughly 30 million people globally doing software-and-data work, the addressable expansion is to roughly 300 million within a decade. That sounds insane until you do the math: that is software employment growing at the same compound rate that internet-related employment grew between 1995 and 2010. We have lived through this exact expansion before. We just didn't call it expansion at the time — we called it "the dot-com era."
IV. The Compounding Effect on Sales, Marketing, Operations
In a world where building is cheap, distribution becomes the bottleneck.
If 5 million new niche software businesses get founded over the next decade — each addressing a specific vertical, geography, or workflow — each of them needs sales, marketing, customer success, support, operations, finance, legal. Building the product is the easy part. Getting it to the right customers, onboarding them, retaining them — that's the hard part.
The first-order effect of cheap building is more buildings. The second-order effect is a massive demand surge for everyone who isn't a builder.
Conservatively: 5 million new software businesses × 10 non-engineering people per business = 50 million new commercial and operational knowledge-work jobs over the decade.
These won't all be human. AI will eat large parts of SDR work, content marketing, tier-1 support. But "AI eats some of these jobs" is not the same as "these jobs disappear" — it means each human in these roles becomes 5-10× more productive, supervises a fleet of AI agents, and the total economic surface expands faster than the productivity multiplier. Net result: more humans doing this work, with each one operating at a higher level.
V. The Real High-Demand Roles Won't Be What You Think
Here is the contrarian forecast worth tattooing on your wrist.
The highest-demand role of the next decade will not be AI engineers. There are perhaps 50,000 people in the world today who can train a frontier model from scratch. That number will grow modestly, but frontier-model development is becoming concentrated in a handful of labs, not distributed across the economy. AI engineering will be a high-prestige, narrow profession — not a labor-market answer.
The highest-demand role will be the unglamorous one nobody is talking about: data migration, integration, and reconciliation specialists.
When 100,000 companies build their own internal Salesforce, the moment they cross 18 months of operation they will all face the same problem: their data needs to migrate. To a new schema. To a new tool. To a merger partner. To a regulatory format. To an integration with a vendor they didn't anticipate. To a backup. To a reporting layer.
Today, "Salesforce data migration" alone is a $5 billion services industry, supported by a few thousand specialists and a few hundred consultancies. Multiply the number of bespoke systems by 100×, and the migration market grows accordingly. Plus the integration market. Plus the reconciliation market. Plus the schema-evolution market.
The clean, easy work — writing the application code — will be done largely by AI. The dirty, messy, judgment-heavy work — figuring out how this customer's data maps to that schema, what to do with the 8% of records that don't fit, how to handle the historical state of orders that were created under a different process three years ago — will require humans for a long time.
The plumbers of the AI era will out-earn the architects.
Other roles in the same category — undersupplied, unglamorous, about to explode:
- Domain experts who can articulate how their industry actually works well enough to encode it in software (the underwriter who really understands cargo insurance pricing; the floor manager who really understands the chemical sequence of a particular plastics line).
- Quality and evaluation engineers who can build harnesses for AI outputs in regulated industries.
- Distribution and growth operators — because every founder is now a builder, but few are sellers.
- Trust, security, and compliance specialists — because every company is suddenly building software with limited internal expertise.
If you are advising a 20-year-old today on what skill to build, "learn to migrate data" is a more defensible answer than "learn to train models."
VI. The Honest Steelman: The Transition Will Be Brutal
It would be dishonest to write this piece without engaging the bear case.
Yes — large numbers of people in current SaaS roles will be displaced. The 50-person customer success organization at a mid-size SaaS company is genuinely at risk, because the customer they were serving is going to build their own version of that company's product. The displacement is real, and brutal in the short run.
Yes — the transition will not be smooth. The historical expansions cited above (looms, ATMs, spreadsheets) happened over 15-30 year arcs, with significant individual hardship along the way. The transition from horse-and-buggy to automobile employment in the early 20th century created millions of net jobs in the long run, but a generation of stable hands had a very bad decade.
Yes — the geographic and skill distribution of new jobs will not match the displaced ones. A 45-year-old SaaS account executive in Dublin does not seamlessly become a data migration specialist in Bangalore.
The expansion is real and large. The disruption is also real and painful. Both are true.
The argument here is not "everything is fine, nothing to see here." It is "the aggregate forecast is dramatically more expansionary than the news cycle is communicating, and the policy and business question is how to manage the transition, not whether the contraction is real."
If you are a 50-person SaaS company, you have 24-36 months to figure out where your moat actually is. If it's a process you've encoded that customers can't replicate, you survive. If it's a thin wrapper over capabilities AI has now commoditized, you don't.
VII. The Structural Problem: Old Models Can't Absorb the New Demand
Here is the part almost nobody is thinking about.
If the forecast above is even directionally right — even half-right — the working population of high-end knowledge workers needs to roughly double, perhaps grow 5-10×, within a decade.
There are not enough people available under the current employment model to fill that demand.
Today's knowledge work runs on a 20th-century industrial assumption: one person, one company, one role, one full-time exclusive engagement. That model worked when demand for any specific specialized skill was lumpy enough to fill a 40-hour week at a single employer. It does not work when 500,000 mid-sized companies each need a senior data migration specialist three days a month, or a fractional security architect for 20 hours a quarter, or a compliance reviewer for 8 hours per regulatory cycle.
The math forces a structural shift. There are exactly three options:
a) Companies dramatically lower their hiring bar and accept much less expert work. Unlikely — the work is too consequential.
b) AI absorbs all the expansion and humans stay flat. Possible for some categories, but data migration, domain expertise, and trust work all require human judgment that is far from automatable.
c) The unit of organization shifts from "employee" to "fractional, multi-company, deliverable-priced specialist."
The third option is the only one that scales.
This is why the fastest-growing executive roles in the United States today are fractional ones — fractional CFOs, CMOs, CTOs, GCs. It is why platforms like Toptal, A.Team, and Andela have grown 10-30× over the last decade. It is why the consulting industry is bifurcating into two extremes: very expensive McKinsey-tier strategy work on one end, and on-demand specialist platforms on the other.
The middle is hollowing. The new structure will be small core teams of full-timers, surrounded by a much larger network of specialists deployed on demand across multiple companies. The Virtual Delivery Center model — pods of skilled professionals offering capacity to multiple clients simultaneously, with shared infrastructure for delivery, governance, and trust — is the natural endpoint of the math. Not because it's a nice idea, but because the demand expansion of the next decade cannot be served by any other structure.
The companies that figure this out early will have an enormous structural advantage. The companies that try to hire 20× more full-time employees will simply not be able to staff the work.
VIII. The Layoff Cycle Is Noise
Step back from the news cycle for a moment.
In 1991, IBM laid off roughly 20,000 people in a single quarter, then more the next year. The press declared the end of corporate America. Five years later, the dot-com hiring boom began, and IT employment in the United States grew faster than at any point in its history.
In 2008-2009, the financial crisis eliminated 8.7 million jobs in the U.S. labor market. The press declared a permanent contraction. Within five years, employment had recovered. A decade later, the U.S. labor market was tighter than at any point since World War II.
In 2022-2025, tech layoffs eliminated roughly 500,000 positions across the major software firms. The press has declared the end of knowledge work.
The third episode rhymes with the first two. Layoffs are visible, dramatic, and easy to cover. Hiring is invisible, distributed, and boring — until you look back five years later and realize the workforce has expanded, and the people doing the work are doing different work than they were before.
If you are reading this and worried about the layoffs in your industry, the answer is not panic. It is to position yourself for the thing the layoffs are obscuring: the largest expansion of knowledge work in human history is starting now, and the bottleneck is going to be people, not technology.
The companies that will dominate the next decade will not be the ones that built the model. They will be the ones that figured out how to organize the workforce to use it.
A billion knowledge workers is not a forecast. It is an arithmetic floor.