We've read about AI taking over low level programming, summarizing medical appts, and other things.
I've been chatting with my granddaughter who just graduated accounting and took a high level intern job at EY.
##################
EY's response to tools like Cowork has been to pivot away from selling "labor" and toward selling "outcomes."
- Outcome-Based Pricing: EY is leading a push among the Big Four to move away from billing by the hour. If a Claude Cowork agent can do a week’s worth of financial auditing in an afternoon, billing by the hour becomes a fast track to bankruptcy.
- The "Dumbbell" Workforce: EY is restructuring its staff into a "dumbbell" shape—heavy on senior strategic leaders and junior "AI orchestrators," while gutting the "middle" layer of associates who used to do the manual data-crunching that Cowork now automates.
The reason Wall Street is nervous for firms like EY is the democratization of elite labor.
- Audit & Tax: Historically, companies paid EY millions for "managed services" where humans organized messy financial data. Cowork allows a mid-level manager at a client company to point an AI agent at a folder of receipts and spreadsheets to perform the same data synthesis for the cost of a Claude Max subscription ($100/month).
##################
And I asked AI about agentic agents:
Agentic Agents: The "Digital Coworker"An Agentic Agent (like Claude Cowork) is an autonomous system that can navigate multiple environments, use multiple tools, and solve problems it wasn't specifically told to expect.
- How they work: You give a high-level objective, and the agent determines the steps, selects the tools, and self-corrects if things go wrong.
- The "Goal": "Research three competitors for our new product, find their pricing, and draft a 5-page comparison memo in Word."
- The Workflow (Autonomous):
- Plan: Decides it needs a web browser and a word processor.
- Act: Opens a browser, searches for competitors, and bypasses "cookie" pop-ups.
- Synthesize: Realizes one competitor doesn't list prices, so it searches for "leaked pricing" or "PDF brochures."
- Execute: Opens Word and writes the memo.
- Key Differentiator: Reasoning Loops. If the agent gets an error message, it doesn't stop; it thinks, "That didn't work, maybe I should try searching a different database," and tries again.
The Comparison at a Glance| Feature | Agentic Tool (e.g., Salesforce AI) | Agentic Agent (e.g., Claude Cowork) | | User Input | Specific Instruction ("Summarize this") | Broad Objective ("Organize my trip") | | Scope | Single App / Single Task | Cross-Platform / Multi-Task | | Autonomy | Follows a pre-defined path | Creates its own path | | Failure Mode | Stops and asks for help | Troubleshoots and iterates | | Wall Street View | An "Add-on" that justifies a subscription | A "Replacement" that makes subscriptions redundant |
Why this distinction "sent shockwaves"
The reason Wall Street panicked is that Agentic Agents can use Agentic Tools better than humans can.
If an agent (Claude Cowork) can open Salesforce, use the Salesforce "Agentic Tool" to get data, then open Slack and send a report, the customer no longer needs to pay for 10 "seats" of Salesforce for human employees to do that work. They just need one human to manage the Agent. |
