The Intern in the Machine: Why the AI Revolution Is (For Now) a Comedy of Errors.

~Sumon Mûkhöpadhuæy

When Microsoft AI CEO Mustafa Suleyman recently suggested that large-scale automation of white-collar work is just 12 to 18 months away, boardrooms across the world sat up straighter.

It sounded like the beginning of the end for routine corporate jobs. But when you zoom in on how today’s so-called “agentic AI” actually behaves in real environments, the picture is far less cinematic — and far more chaotic.

The revolution is coming. But right now, it’s tripping over office chairs.

The Claude Coworker Experiment: When AI Met the Desktop

In late October 2024, Anthropic unveiled its much-talked-about “Computer Use” feature — a version of Claude that could control a computer like a human: moving the cursor, clicking buttons, switching windows, and filling forms.

The promise was bold: AI that doesn’t just advise — it acts.

The Reality? Think brilliant intern. Zero focus.

During one recorded demo, Claude was asked to plan a sunrise hike. Midway through the task, it quietly abandoned the workflow and began browsing photos of Yellowstone National Park. Not because it was relevant. Because it got distracted.

In other runs, the agent:

• 🔹 Got trapped in endless UI loops
• 🔹 Misread pop-ups as new instructions
• 🔹 Clicked “submit” on half-filled forms
• 🔹 Froze when screens didn’t behave exactly as expected

It wasn’t malicious. It was simply… clueless in complex real-world interfaces.

---

The Immediate Reality: The Crisis Is Postponed

For sectors like Indian IT — which briefly trembled when agentic AI hype hit headlines — there’s a short-term sigh of relief. The great job wipeout is not happening tomorrow.

As of early 2026, even with the release of Claude Opus 4.6, AI agents still suffer from two fatal weaknesses:

• Latency: They often think too slowly for high-speed operational work.
• Reliability: While autonomous agents are hitting 90-95% accuracy in controlled benchmarks, that 5% "hallucination gap" is a dealbreaker for enterprise scale.

The Verdict: A 10% mistake rate in coding is annoying. A 10% mistake rate in banking systems, payroll, or compliance is a lawsuit.

But the Future? Uncomfortably Uncertain

Here’s where the comedy slowly turns into a thriller. What looks clumsy today is improving at exponential speed. The "Agentic Shift" is moving from assistive tools to autonomous teammates.

The Indian IT business model is structurally vulnerable because it has scaled for decades on a simple equation:

More people = more billable hours = more revenue.

Why the Business Style Must Change:

The Old Model	The Agentic Threat
Arbitrage: Selling cheap junior manpower.	AI handles 80% of junior tasks for near-zero cost.
Headcount Growth: Hiring thousands to scale.	Revenue detaches from human payroll.
Maintenance Teams: Large groups for routine work.	Clients prefer "good enough" in-house AI agents.

The #SumanSpeaks Verdict

The robots aren’t stealing jobs today. They’re too busy procrastinating on Yellowstone photos. But they are learning. With multi-agent systems growing by over 300% in the last few months, the "trajectory" is what should keep CEOs awake at night.

The danger isn’t sentient AI; it’s efficient AI. If IT giants continue selling “people hours” instead of outcomes, automation will quietly hollow out the very foundation of their business.

The warning light is already blinking. Is anyone restructuring fast enough to survive?