Daily TEA – ATMs Didn’t Kill Jobs (iPhones Did), Legal Agents, and AI Tax Season
paradigm replacement, Zoom agentic AI, Harvey legal agents, AI tax returns, HBR last mile
Hello, dear TEA-mates! Here is what you need to know today.
1. 🏦 Why the ATM Didn’t Kill Bank Tellers, But the iPhone Did
David Oks dismantles the favorite politician talking point—”ATMs didn’t eliminate tellers, so AI won’t eliminate jobs”—and reveals it’s only half the story. ATMs were task automation within an existing paradigm: they cut tellers per branch from 21 to 13, but cheaper branches meant more branches, so total teller employment actually increased from the 1970s through 2010. The real killer was the iPhone. Mobile banking made physical branches irrelevant entirely. Bank teller jobs collapsed from 332,000 in 2010 to 164,000 by 2022—a 51% wipeout. Bank of America alone shed 84,000 employees and closed 40% of its branches. The lesson: task automation encounters friction within existing workflows and often creates new roles. Paradigm replacement makes the entire workflow obsolete. For AI, the “drop-in remote worker” framing is the ATM phase. The real displacement comes when fully AI-native organizations are built from scratch—not retrofitted into human-shaped holes. (Read More)
🫖 TEA For Thought: Automation is the ATM—once the paradigm shift completes, automation itself goes obsolete. Or to put it differently, automation is not a long-term business. Though one could argue: what is long-term in the era of AI?
2. 📹 Zoom Expands Agentic AI Platform With Workflow Orchestration
Zoom announced a sweeping expansion of its agentic AI platform, turning every meeting, call, and contact center interaction into a trigger for automated enterprise workflows. AI Companion 3.0—whose monthly active users tripled year-over-year in Q4 FY26—now supports no-code custom agent building with prebuilt templates for sales, IT, and marketing, plus 10 new enterprise data connectors including Salesforce, ServiceNow, Box, Google Drive, and OneDrive. Zoom Workplace gets a dedicated AI tab, AI-native Docs/Sheets/Slides that convert meeting insights directly into structured outputs, real-time voice translation in 5 languages, and deepfake detection. Zoom Phone (10M+ seats globally) adds agentic post-call automation—auto-drafted emails, summaries, and task execution—while AI Receptionist gains SMS. Zoom CX 3.0 introduces Expert Assist for real-time agent coaching and natural-language analytics. A new developer layer, Zoom AI Services, exposes speech, language, reasoning, and summarization APIs. (Read More)
🫖 TEA For Thought: Zoom wants to realize the “speak it, and it’s done” vision for enterprise workflows. Seems like it’s coming—every conversation becomes a workflow trigger, and the line between meeting and execution disappears.
3. ⚖️ Harvey’s Agent Builder: Vertical-Specific Agentic Intelligence
Harvey announced Agent Builder, evolving its Workflow Builder into a platform where legal teams create reusable, shareable agentic workflows tailored to their specific processes. The numbers show Harvey is already natively agentic: 400K+ daily agentic queries for contracting and document editing, 20M+ terms extracted in review tables, 445K+ Deep Analysis reports generated, and 25,000+ custom workflows already built by teams like GSK Stockmann (automating due diligence) and Ashurst (saving hours on lease summaries). Agent Builder enables template embedding for consistent outputs, organization-wide agent sharing, and scheduled background execution for tasks like contract expiration monitoring and compliance checks. The key innovation: agents now handle multi-step tasks with inference—they grasp objectives, ask for context, find optimal paths, and maintain human-in-the-loop checkpoints at critical moments. (Read More)
🫖 TEA For Thought: Verticals like Harvey are becoming agentic intelligence layers with embedded subject matter expertise. As Jack Dorsey said, the future of the company is the intelligence itself—a layer that can be built upon by other agents. Harvey, as the legal vertical, is already leading this game.
4. 🧾 Kyle Corbitt’s AI vs. Accountant Tax Return Experiment
Kyle Corbitt ran a real-world experiment: preparing his complex 2025 tax return simultaneously with a professional accountant and OpenAI’s Codex. His situation was non-trivial—7 income sources, 4 K-1s, a company sale with multiple compensation types, appreciated stock donations, crypto, mortgage, and partial ACA coverage. Codex’s approach: ingest prior returns plus a 10-minute voice dictation of life changes, then collaboratively research IRS rules, ask targeted clarification questions, and maintain a living README of every tax decision. When off-the-shelf tools proved inadequate, Codex built a custom Python tax engine from scratch in roughly 30 minutes. The accountant’s initial estimate came in $20,000 off—Codex caught a missed “adjustment escrow” payment from the company sale that even Corbitt had forgotten. After correction, both estimates matched within dollars. Corbitt open-sourced the tax engine on GitHub. (Read More)
🫖 TEA For Thought: Accountants’ jobs will face serious pressure soon. This is just the beginning—when an AI catches a $20K error the professional missed, the value proposition flips fast.
5. 🏢 The “Last Mile” Problem Slowing AI Transformation
Harvard Business Review’s Karim Lakhani, Microsoft’s Jared Spataro, and Harvard D³’s Jen Stave identify the real bottleneck in enterprise AI: not model quality or data, but organizational design. Companies are “pilot-rich but transformation-poor”—one global bank runs 250+ LLM apps, a payments network hit 99%+ copilot adoption, yet productivity gains get reabsorbed into legacy processes. Seven structural barriers: pilot proliferation without scaling paths, productivity gains that don’t translate to organizational ROI, legacy process debt that AI makes visible faster than it can be fixed, subject matter experts resisting knowledge externalization, governance breaking down with multi-agent systems, patchwork AI architectures, and the “efficiency trap” where cost-cutting framing triggers defensiveness. Solutions: clean-sheet process redesign, strategic knowledge capture (experts as AI architects), centralized agentic control planes, and human role evolution toward orchestration and interpretation. (Read More)
🫖 TEA For Thought: Getting from 90 to 99 might be as hard as from 0 to 90. Though the advancement of the models might close the gap faster than we thought. The last mile is organizational design, not technology.
Prompt Tip of the Day
Stop writing prompts from scratch—let the AI write the perfect prompt for you. Google DeepMind research found AI-optimized prompts outperform human-written ones by up to 50%.
“You are a top-tier prompt engineer and a [domain] expert. Goal: Create the best possible master prompt that generates [desired output]. Step 1: Ask me the must-have inputs (at least 12 questions) covering audience, goals, constraints, tone, proof points, and format. Step 2: After I answer, output a single COPY-PASTE MASTER PROMPT that uses placeholders like {variable} and produces [specific deliverable type].”
Feed this meta-prompt with your rough goal; it asks clarifying questions, then hands back a reusable template with placeholders—each iteration gets smarter as you refine it.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:
Twitter/X