Specialized knowledge work is being eaten. Not augmented. Not assisted. Eaten.

Claude Code took the junior developer job. Junior dev postings dropped 60 percent between 2022 and 2024. Salesforce stopped hiring software engineers in 2025 and said AI productivity was the reason. The bottom of the ladder didn’t upgrade — it got removed.

Duolingo’s first 100 language courses took 12 years to build. Their AI built 148 more in under a year. The company went “AI-first” and stopped contracting humans for work the model could do.

Harvey ate the work junior associates used to do — contract analysis, due diligence, legal research — and reached an $11 billion valuation in four years.

Radiology is next. In April 2026, the CEO of NYC Health + Hospitals — America’s largest public hospital system — said he could replace a great deal of radiologists with AI today if liability law allowed it. The capability exists. Regulation is the only thing slowing it down.

The pattern is clear. Agentic AI is moving into domains where specialized knowledge workers once owned the information flow, the decision-making, the coordination. When an AI can do the work — not assist with the work, but do the work — the software that supported that work becomes obsolete.

This is happening everywhere.

And it’s starting in construction.

Project management software today is a database with a chart.

Procore. P6. Microsoft Project. They store information. They visualize it. That’s it.

But actual project management isn’t data storage. It’s information flow. A typical $100M commercial project moves 10,000 products, 25,000 action items, and coordinates dozens of subcontractors across site managers, superintendents, procurement, and owners.

That flow is 100 percent human-powered.

A PM is a human inbox. A superintendent is a human router. Decisions happen in people’s heads and get lost.

Agentic AI changes the shape of the work. It connects to every channel — email, WhatsApp, calls, texts. Captures every update the moment it happens. Routes it to the right person based on org structure and rules. Updates forecasts in real time. Adjusts sequences. Alerts before problems cascade.

Not software you license.

A worker you hire.

Why Now

Three pieces converged.

First, the agent frameworks are here. Open-source systems like OpenClaw make it possible to build autonomous agents that coordinate real-world work. Not chatbots. Not decision support. Agents that perceive, reason, and act.

Second, the build cost collapsed. Cloud infrastructure is ubiquitous. Large language models — Claude, GPT-4, Codex — dropped the cost of building complex software by an order of magnitude. What used to take twenty engineers two years now takes three engineers three months.

Third, autonomous decision-making under uncertainty is solved. Tesla’s Full Self-Driving handles hundreds of moving agents, real-time sensor fusion, and millisecond decision trees with a human supervising. Construction projects face the same class of algorithmic problems. The same approaches apply — with humans in the loop where judgment matters.

The conditions are right. The technology exists. The pieces are being built separately.

The integration is the opportunity.

We’re building it.

Pejman Golkar

I write weekly about agentic AI, construction, and the death of software-as-a-tool. Subscribe to get the next essay.