Silicon Valley tech leaders panic online about NYC's socialist mayoral nominee, but 200 executives who met Zohran Mamdani in person tell a different story. The geographic divide reveals how distance distorts political risk assessment in tech.
Good Morning from San Francisco, Enterprise AI just got a reality check. MIT studied 300 deployments and found a 95%
Nvidia's new B30A chip for China reveals how Washington is monetizing export controls. The Blackwell-based processor delivers half the power of flagship models while requiring 15% revenue sharing with U.S. government—turning restrictions into tolls.
MIT study reveals 95% of enterprise AI pilots fail to boost revenue despite billions invested. The twist: companies spend on flashy sales tools while back-office automation delivers real savings. One firm saved $8M with an $8K tool.
💡 TL;DR - The 30 Seconds Version
👉 MIT study of 300 AI deployments finds only 5% of enterprise pilots achieve rapid revenue growth despite billions in investment.
📊 Companies allocate 50% of AI budgets to sales tools but MIT found highest returns in back-office automation eliminating outsourced contracts.
💰 One firm saved $8 million annually by deploying an $8,000 AI tool to streamline administrative work.
🤝 Partnerships with specialized vendors succeed 67% of the time versus 33% for internal development projects.
🏢 Companies avoid layoffs by canceling outsourced contracts and not backfilling vacant administrative and support positions.
🚀 Success requires narrow focus on specific workflows rather than broad AI deployment across multiple business functions.
Budgets chase sales; savings hide in the back office.
Billions have rushed into generative AI. The surprise is where value isn’t showing up. According to MIT’s State of AI in Business 2025, only about 5% of pilots deliver rapid revenue gains. Most efforts stall before they hit the P&L in a measurable way.
MIT’s NANDA team draws a bright line between consumer-grade chatbots that feel productive and enterprise systems that must learn, adapt, and persist inside workflows. The report calls this the “learning gap.” Tools that don’t remember context or mesh with existing processes can dazzle in demos yet fail in production. The result is lots of activity, little impact.
The winners look different. Startups that pick one painful use case, ship quickly, and partner with distribution channels show outsize results. MIT profiles companies that went from zero to meaningful revenue in a year by narrowing scope and integrating deeply rather than chasing general-purpose magic. Focus beats breadth.
Enterprises are spending in the wrong places. Roughly half of gen-AI budgets flow to sales and marketing tools, where attribution is murky and uplift is hard to prove. The stronger returns, MIT finds, come from back-office automation: eliminating BPO contracts, cutting agency spend, and streamlining administrative work. One firm reportedly saved $8 million annually with an $8,000 tool. Quiet savings, real cash.
That mismatch explains the “looks great, pays late” problem. Front-office pilots produce glossy dashboards and exec demos. Back-office projects lack sizzle but drop costs fast. Most companies still optimize for visibility over verified value. Fixing that bias is the low-drama way to put AI on the income statement.
Adoption strategy is the strongest predictor of success. Purchasing from specialized vendors and forming tight partnerships succeeds about 67% of the time, MIT reports. In-house builds succeed roughly one-third as often. That’s awkward for financial services and other regulated sectors that are investing heavily in proprietary systems this year. The odds aren’t with them.
Why the gap? Vendors that live inside a single workflow can ship domain-specific memory, connectors, guardrails, and change-management playbooks. Internal teams face a moving stack, brittle requirements, and governance cycles that outlast the technology half-life. When the tooling itself changes monthly, velocity matters. So does choosing a problem narrow enough to finish.
This isn’t a layoff wave. It’s a slow rebalancing. Companies are canceling outsourced work and not backfilling administrative and support roles when people leave. Executives in tech and media—two sectors furthest along—expect lower hiring volumes over the next two years, not mass firings today. Cost curves shift without the headline shock.
That pattern is politically simpler and financially cleaner. Ending a contract is easier than pink slips. It also concentrates impact in standardized, lower-priority tasks, exactly where automation shines. The risk is complacency: organizations undercount productivity gains when “shadow AI” handles work outside sanctioned tools, masking upside and hiding failure modes.
Three moves recur in the success cases. First, reallocate budget toward back-office automation with line-item savings you can audit. Second, empower line managers to own deployment and metrics; central labs can coach, but adoption lives on the front lines. Third, prefer vendors that ship workflow memory, fine-grained permissions, and embedded controls over generic chat interfaces. Those details decide whether pilots stick.
This is unglamorous work. It’s also how AI starts paying for itself. Treat gen-AI like any other operational change: define the unit of value, choose a bottleneck, measure before and after, and keep scope narrow until the savings show up in the monthly close.
MIT’s snapshot synthesizes interviews, employee surveys, and analyses of hundreds of deployments, but it is still early-cycle data. Measurement remains hard in the front office, where attribution spans quarters and multiple touchpoints. And headline percentages can obscure variance across industries and use cases. The direction of travel is clear; the absolute numbers will move.
Why this matters:
Q: What exactly is the "learning gap" that's causing AI failures?
A: Generic AI tools like ChatGPT work well for individuals because they're flexible, but they fail in enterprises because they don't learn from company workflows or remember context between tasks. They can't adapt to specific business processes.
Q: Why do partnerships with vendors succeed twice as often as building AI internally?
A: Specialized vendors focus on single workflows and ship with built-in memory, connectors, and guardrails. Internal teams face moving technology stacks and governance cycles that outlast the technology's half-life.
Q: Which industries are seeing the most AI disruption right now?
A: Technology and media sectors show the clearest signs of AI impact. Over 80% of executives in these industries expect reduced hiring volumes in the next two years, making them the most advanced adopters.
Q: What is "shadow AI" and why does it matter?
A: Shadow AI refers to employees using unauthorized tools like ChatGPT for work tasks. This creates productivity gains outside official AI budgets and measurement systems, making it hard to track real AI impact.
Q: How are the successful young startups different from struggling enterprises?
A: Successful startups led by founders in their early twenties pick one specific pain point, execute quickly, and partner strategically with established companies. They focus narrowly rather than trying to solve everything at once.
Q: What specific back-office tasks are companies automating successfully?
A: Companies are eliminating business process outsourcing contracts, cutting external agency costs, and streamlining administrative work. Customer support and administrative roles show the earliest automation success.
Q: What timeline should we expect for broader job displacement?
A: MIT estimates 3% of jobs face near-term replacement risk, but 27% could be affected longer-term. Current impact focuses on not backfilling positions rather than layoffs.
Q: What makes financial services companies' approach particularly risky?
A: Many financial services firms are building proprietary AI systems internally in 2025, but MIT's data shows internal builds succeed only 33% of the time versus 67% for vendor partnerships in regulated industries.
Nvidia's new B30A chip for China reveals how Washington is monetizing export controls. The Blackwell-based processor delivers half the power of flagship models while requiring 15% revenue sharing with U.S. government—turning restrictions into tolls.
Margaret Boden couldn't use computers but became a leading AI philosopher. Her 1998 prediction—that AI would excel at generating ideas but struggle with evaluation—now explains why ChatGPT creates convincing misinformation and legal systems cite fake cases.
China staged its first humanoid robot olympics with 280 teams from 16 countries. Unitree dominated track events while robots stumbled through soccer and boxing. The spectacle revealed both genuine progress and persistent limitations in a $1 trillion robotics push.
Anthropic lets Claude Opus 4 end abusive chats—not to protect users, but potentially the AI itself. The company remains uncertain about AI consciousness but implements "model welfare" safeguards anyway. A precautionary ethics experiment.
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
