AI job anxiety is becoming a startup thesis, but the hard part is political

In a new post on Keyvan’s blog, the argument is not that AI will neatly replace humans, or that large language models are close to human intelligence. The sharper claim is that much of salaried work was never designed around the full range of human ability in the first place. If a job mostly requires repeated classification, rewriting, routing, summarizing, checking, formatting, and coordination, then current AI does not need to be truly intelligent to change the economics of that role.

That distinction matters for founders, investors, and operators. The AI labor debate often gets stuck between two weak positions. One side treats automation as inevitable replacement. The other side argues that because models do not understand like humans, employment is mostly safe. Keyvan’s post sits in the more uncomfortable middle: AI may not be human-like, but many jobs are already machine-like enough for AI to absorb meaningful parts of them.

There is no funding announcement here, no investor syndicate, and no startup traction metric disclosed. That makes the post different from the usual AI-company launch story. Still, it points at one of the largest venture theses in software right now: startups are not only building tools for workers, they are increasingly building systems that compete with tasks, teams, and sometimes entire service categories.

The company-shaped version of this idea is already familiar. AI coding tools target junior engineering tasks, QA loops, documentation, and migration work. AI customer support companies target ticket triage, first-response handling, and knowledge-base retrieval. AI sales products draft outreach, summarize calls, update CRM fields, and identify follow-ups. AI legal and finance tools process documents, extract obligations, flag anomalies, and prepare first-pass analysis. The market positioning is rarely stated as “we replace jobs,” because that is commercially and politically risky. It is usually framed as productivity, coverage, response time, or margin expansion.

That framing is not false. It is incomplete.

The problem these companies solve is real: organizations run on an enormous quantity of repetitive knowledge work. A manager may call it operations. A worker may call it the job. An investor may call it workflow automation. A model sees tokens, patterns, and instructions. The same activity can look empowering from one angle and destabilizing from another.

The funding logic follows from that ambiguity. Investors like AI labor automation because the budgets are large and the pain is measurable. If software can reduce support cost per ticket, shorten sales cycles, speed up engineering review, or replace outsourced back-office work, the buyer can calculate return on investment quickly. That is why AI application startups often position themselves against headcount, agencies, business-process outsourcing, or legacy SaaS seats. The strongest pitch is not that the model is magical. It is that the customer already pays people or vendors to perform a repeatable process, and the AI system can do enough of it at lower marginal cost.

Keyvan’s skepticism is useful because it separates capability from ideology. He does not accept the claim that current LLMs are on a clean path to artificial general intelligence. He also rejects the comforting assumption that this limits their labor impact. A spreadsheet did not need to be an accountant to reshape accounting departments. A search engine did not need to be a librarian to change research habits. A large language model does not need human judgment to take over drafts, summaries, handoffs, and routine decisions where organizations already tolerate shallow reasoning.

For startups, that creates a narrow but valuable opening. The best AI companies in this category do not sell a vague promise of intelligence. They attack constrained work with clear inputs, clear outputs, and enough review mechanisms to keep failure costs acceptable. A support agent copiloted by AI is easier to sell than a fully autonomous support department. A contract review system that highlights clauses is easier to trust than one that negotiates alone. A developer tool that proposes code changes inside an existing workflow is easier to adopt than a black-box engineering replacement.

The trade-off is that partial automation can still change labor demand. If one person with AI can handle the throughput of three people in a narrow function, the company may grow faster without hiring. That does not show up as a mass layoff headline. It shows up as fewer entry-level roles, smaller teams, flatter service margins, and a higher bar for workers who once learned through routine tasks.

This is where the post becomes more than another AI opinion piece. Keyvan questions why the defense of human welfare is so often reduced to defending jobs. That argument is uncomfortable in a venture context because most startup economics are built around firms becoming more efficient. Founders pitch lower cost, higher throughput, and software-like margins. Workers experience the same shift as insecurity unless the gains are redistributed through wages, shorter working hours, public services, ownership, or some other mechanism outside the product itself.

That gap is not a model problem. It is an institutional problem.

A startup can build AI that removes drudgery, but it cannot by itself decide whether the saved time becomes leisure, profit, lower prices, or unemployment. A founder can claim that automation frees people for higher-value work, but that claim depends on training budgets, hiring practices, labor markets, and management choices. In many companies, “higher-value work” is not an available destination for everyone. Sometimes it is a slogan used while the org chart gets smaller.

The technical limits of LLMs still matter. These systems are brittle in unfamiliar domains, prone to confident errors, expensive at scale in some workloads, and dependent on good data access. They work best when the task can be verified, bounded, or corrected. They are weakest when the work requires accountability, long-horizon planning, deep domain causality, physical presence, or trust built through human relationships. That is why the near-term labor impact is likely uneven rather than universal.

But uneven does not mean small. Many knowledge-work roles are bundles of tasks. AI does not need to replace the whole bundle to change the job. If it removes 30 percent of the work, the role may become more strategic. If it removes the easiest 30 percent, it may also remove the training ground. If it removes the visible output but leaves humans responsible for errors, it may increase pressure rather than reduce it.

For investors, the opportunity is obvious but ethically loaded. AI companies that automate expensive workflows can create large businesses. The most credible ones will show traction through usage depth, retention, cost reduction, and workflow ownership rather than demo appeal. The weaker ones will confuse impressive generation with durable value. The market is already crowded with products that can draft, summarize, and chat. The defensible companies will own distribution, proprietary data loops, compliance pathways, or deeply embedded operational workflows.

For workers, the lesson is less tidy. Learning to use AI tools is rational, but it is not a full answer to the social question. If every worker becomes more productive, employers may need fewer workers for the same output. If only some workers gain access or training, inequality inside firms can widen. If AI handles routine tasks, early-career workers may lose the ladder that once helped them build judgment.

Keyvan’s post is valuable because it refuses a comforting binary. AI is not necessarily close to human intelligence, and jobs are not necessarily sacred. Both can be true. The startup ecosystem is now building businesses in the space between those statements. That is where the commercial opportunity sits, and also where the hardest questions are being deferred.