Three Years in Four Weeks

Three Years in Four Weeks

In late February 2026, Perplexity AI quietly published a blog post with a claim that should have set off alarms in every corporate office from London to Los Angeles. The company's new product, Computer for Enterprise, had been deployed internally as a Slack integration, with every employee in the same channel.

After processing more than 16,000 queries in four weeks, the system had, by Perplexity's own estimation, completed the equivalent of 3.25 years of human work and saved the company $1.6 million in labour costs. The benchmarks used to measure this output came from institutions including McKinsey, Harvard, MIT, and Boston Consulting Group.

Let that settle for a moment. Not 3.25 years spread across thousands of workers performing marginal speed improvements. The claim is that a single AI platform, running cloud-based workflows across roughly 20 frontier models, replaced years of the kind of cognitive labour that knowledge workers perform every day: querying databases, compiling reports, synthesising research, drafting analyses. The tasks that fill the calendars of financial analysts, marketing strategists, management consultants, and corporate researchers everywhere.

Perplexity's CEO, Aravind Srinivas, framed the ambition with characteristic directness. "What we are going to try to do is help businesses run as autonomously as possible," he said. On the question of AI displacing jobs, he offered a response that managed to be both provocative and revealing: "The reality is most people don't enjoy their jobs." His suggestion was that displacement could free people to pursue entrepreneurship and more fulfilling work. It is, to put it mildly, an incomplete answer to a question affecting hundreds of millions of workers worldwide.

The Machine That Writes the Queries

To understand why Perplexity's claims matter, you need to understand what Computer for Enterprise actually does. It is not a chatbot. It is not a search engine with a conversational veneer. It is an orchestration platform that routes tasks across approximately 20 AI models from multiple providers, including:

• Anthropic's Claude Opus 4.6 as its primary reasoning engine
• Google's Gemini for deep research
• OpenAI's GPT-5.2
• xAI's Grok

Each session runs inside its own isolated Firecracker virtual machine, ensuring data separation between users. The platform connects natively to the software stack that modern enterprises already run: Snowflake, Salesforce, HubSpot, Slack, Notion, GitHub, Gmail, Outlook, and more than 400 other applications through its connector ecosystem. Administrators can install custom connectors via the Model Context Protocol.

The system includes workflow templates for legal contract review, finance audit support, sales call preparation, and customer support ticket triage.

Here is the critical capability: Computer for Enterprise does not merely answer questions. It writes the database queries, executes them, and returns structured results. A financial analyst can ask for revenue broken down by vertical from Snowflake, and the system will compose the SQL, run it against the data warehouse, and present the findings. A sales team can simultaneously pull CRM data and competitive context. The AI handles the translation from natural language intent to technical execution and back again, collapsing what might take a human analyst hours into seconds.

Srinivas described the underlying philosophy on the social media platform X: "When AIs can orchestrate a file system with CLI tools plus a browser, AI essentially becomes the Computer, running things on the cloud as you sleep." He drew a distinction between traditional operating systems and what Perplexity is building: "A traditional operating system takes instructions; an AI operating system takes objectives."

The enterprise offering comes wrapped in the security apparatus that corporate procurement teams demand:

• SOC 2 Type II compliance
• SAML single sign-on
• Audit logs
• Sandboxed query execution
• GDPR and HIPAA compliance

Pricing runs at $325 per user per month for the Enterprise Max tier, or $40 per user per month for Enterprise Pro. Perplexity's annualised revenue reached approximately $148 million by mid-2025, with internal projections targeting $656 million by the end of 2026.

The company is candid about limitations. Factual hallucinations occur, particularly on niche topics or very recent events. The system occasionally generates broken URLs. External communications, whether emails or published content, should always be reviewed by a human before distribution. But the trajectory is clear, and the implications are staggering.

The Scale of What Could Be Lost

The question that Perplexity's announcement forces into the open is not whether AI can perform knowledge work. That debate ended sometime around mid-2024, when large language models began consistently demonstrating competence at research synthesis, data analysis, report writing, and code generation. The question now is what happens to the people who currently do this work for a living.

The numbers are sobering. According to Goldman Sachs research, generative AI could automate tasks equivalent to 300 million full-time jobs worldwide, with 26 per cent of office roles and 20 per cent of customer service positions highly exposed. In the United States alone, Goldman Sachs estimates that AI automation will ultimately displace roughly six to seven per cent of the workforce, equivalent to approximately 11 million workers.

The World Economic Forum's Future of Jobs Report 2025, drawing on perspectives from more than 1,000 leading global employers representing over 14 million workers, projects that 92 million roles will be displaced by 2030, though it forecasts 170 million new roles emerging for a net gain of 78 million jobs.

McKinsey's analysis adds another dimension. The consultancy estimated that today's technology could, in theory, automate approximately 57 per cent of current U.S. work hours. That figure does not mean 57 per cent of jobs will vanish. It means that across the entire working population, just over half of the hours worked involve tasks that a sufficiently deployed AI system could handle. McKinsey projects that 30 per cent of U.S. work hours could be automated by 2030, accelerated by generative AI's capabilities.

The disruption is already visible in employment data:

• There were 77,999 AI-attributed tech job losses in the first six months of 2025 alone.
• Employment in the computer systems design and related services sector declined five per cent since ChatGPT's release.
• Entry-level job postings dropped 15 per cent year over year.
• Employment among software developers aged 22 to 25 fell 20 per cent compared to their late 2022 peak.

According to research from the Dallas Federal Reserve, AI is simultaneously aiding existing workers and replacing others, with the wage data suggesting a complex and uneven transformation.

Certain roles face particularly acute risk:

• Data entry positions carry a 95 per cent automation risk.
• Customer service representatives face 80 per cent risk, because most inquiries are answerable from a knowledge base.
• Paralegals face an 80 per cent risk of automation by 2026, and legal researchers face a 65 per cent risk by 2027.
• An estimated 200,000 jobs are expected to be cut from Wall Street banks over the next three to five years, and as much as 54 per cent of banking jobs have high potential for AI automation.
• SSRN projections estimate that 7.5 million data entry and administrative jobs could be eliminated by 2027.

Seventy-five per cent of knowledge workers are already using AI tools at work, and nearly half started within the last six months. They report 66 per cent productivity improvements. But the question nobody wants to confront directly is this: if each worker becomes 66 per cent more productive, how many fewer workers does an organisation actually need?

The Cautionary Tale Already Playing Out

The corporate world is not waiting for the research to settle before acting. The global technology sector eliminated nearly 60,000 jobs in less than three months of 2026, according to layoff tracker TrueUp, which recorded 171 separate events affecting 59,121 workers since January. That pace, averaging 704 jobs lost per day, is running ahead of 2025, when 245,953 workers were let go across the full year. If it holds, total cuts could reach 265,000 by December.

A Resume.org survey of 1,000 U.S. hiring managers found that 55 per cent expect layoffs at their companies in 2026, and 44 per cent identified AI as a primary driver.

Some of the largest names in technology are leading the charge:

• Amazon confirmed 16,000 corporate job cuts in 2026 despite reporting record revenue of $716.9 billion the previous year, framing the reductions as a push to flatten management layers. Some of those roles are not being backfilled with humans; they are being backfilled with software.
• Block, the payments company formerly known as Square, slashed 4,000 roles in early 2026, nearly 40 per cent of its entire workforce.
• Ingka Group, the largest IKEA retailer, announced 800 office role cuts in March.

Perhaps the most instructive example comes from Klarna, the Swedish fintech company. In 2024, Klarna deployed an AI assistant that handled the equivalent workload of 700 full-time customer service employees. The company's headcount fell from approximately 7,000 in 2022 to roughly 3,000, and CEO Sebastian Siemiatkowski publicly championed the results.

But the strategy backfired. Customer complaints increased, satisfaction ratings dropped, and internal reviews revealed that AI systems lacked empathy and could not handle nuanced problem-solving. By early 2025, Siemiatkowski acknowledged that the company had overestimated AI's capabilities, stating bluntly: "We went too far." Klarna began rehiring human customer service staff, specifically targeting students, rural populations, and dedicated product users.

Klarna's reversal is a cautionary tale that speaks directly to Acemoglu's warnings about "so-so automation." The financial savings looked impressive on a spreadsheet, but the technology degraded the quality of the service it was supposed to improve. The question for every organisation evaluating tools like Perplexity's Computer for Enterprise is whether the same pattern will repeat across other domains: impressive benchmarks followed by the slow realisation that human judgement, context, and empathy were doing more work than anyone appreciated until they were gone.

The Uncomfortable History of "New Jobs Will Appear"

Every wave of technological disruption produces two competing narratives. The optimists point to history: the Industrial Revolution destroyed agricultural and artisan livelihoods but created factory work. The IT revolution eliminated typing pools and filing clerks but created entire industries around software, networking, and digital services. The pessimists counter that this time is different, that the pace and breadth of AI's capabilities outstrip anything that came before.

History offers both comfort and caution. During the first Industrial Revolution, the Luddites famously destroyed the mechanised looms that threatened their livelihoods in industrial Britain. Their fears were not irrational. While new manufacturing jobs eventually emerged, the transition period was brutal. Research from economic historians shows that average real wages in England stagnated for decades even as productivity rose. Eventually, wage growth caught up to and then surpassed productivity growth, but only after substantial policy reforms including labour protections and education acts.

The Second Industrial Revolution followed a similar pattern. Automation technologies increased the efficiency and scope of mechanised production, requiring fewer operators but more engineers, managers, and other new occupations. As automation created fewer middle-skill jobs than it made obsolete, the result was a hollowing out of the skill distribution in manufacturing, a pattern that persists to this day.

The robotics wave of the 1970s and 1980s displaced approximately 1.2 million manufacturing jobs globally by 1990. In the United States alone, robot-induced automation displaced 300,000 factory workers in the automotive sector. New jobs did eventually appear, but they required different skills, existed in different locations, and often paid different wages.

McKinsey's historical analysis offers a striking statistic: 60 per cent of today's U.S. workforce is employed in occupations that simply did not exist in 1940. That is genuinely encouraging. But it also means that 60 per cent of today's workers are in roles that their grandparents could not have trained for, because the jobs had not yet been invented. The lag between destruction and creation is where the human cost concentrates.

What makes the AI wave qualitatively different from previous automation episodes is its target. Earlier forms of automation primarily replaced physical labour and routine cognitive tasks: drilling, sewing, sorting files, calculating spreadsheets. AI encroaches on non-routine cognitive domains once thought uniquely human, including recognising images, drafting emails, drawing illustrations, synthesising research, and making complex judgements.

The Bipartisan Policy Center in Washington notes that AI is different because it can automate many tasks that do not follow an explicit set of rules and are instead learned through experience and intuition.

The pace compounds the challenge. Previous technological transitions unfolded over generations, allowing social institutions to adapt. The shift from agricultural to industrial employment in the United States took roughly a century. The transition from manufacturing to services took several decades. AI capabilities are advancing on a timeline measured in months.

Goldman Sachs models show that each one percentage point productivity gain from technology raises unemployment by approximately...