Anthropic's research shows that AI can already do a huge portion of many jobs; its top economist talks about how that could shape the future of work

Anthropic’s enterprise-usage report finds coding has ~94% theoretical AI exposure but only ~30% observed task adoption, with 30–40% of Claude conversations coding-related concentrated in occupations that represent ~3% of the workforce. The research shows AI exposure is highly uneven across occupations and that widespread displacement is not evident in official BLS statistics yet. Anthropic recommends monitoring real-world adoption, investing in complementary data and workflows, and upskilling to preserve roles that require expert evaluation and contextual knowledge.

Analysis

AI-driven automation will be highly non-linear: productivity gains will concentrate in narrow, repeatable tasks first, producing 20–50% time-savings inside those workflows while leaving adjacent, judgment-heavy work intact. That creates a rising premium on “evaluation” skills—employees who can validate, curate, and direct models—so expect wages and hiring for those roles to outpace average white‑collar hiring by 1.5x–2x over the next 12–36 months.

The biggest commercial winners are likely to be the invisible plumbing and integration layers that remove organizational friction: cloud infra, data lakes/warehouses, observability, and secure content plumbing (IAM, vector DBs). Corporates will shift capex from bespoke model builds to standardized data stacks and governance; anticipate a 6–24 month wave of migration projects and recurring revenue growth for firms that reduce integration cost by >30%.

Key downside catalysts are structural, not model-quality: slow data modernization, expensive workflow rewiring, and political/regulatory pushback if displacement concentrates geographically. Any of these can stall adoption for 12–24 months and favor incumbents with sticky contracts. Contrarian angle: market attention is on flashy LLM apps, but durable alpha will come from owning the companies that turn raw models into reliable, auditable business processes.

Key Decisions for Investors

• Long SNOW (Snowflake) — 12–24 month horizon. Buy a defined-risk call spread (e.g., buy 1y–2y ATM call, sell a higher strike) to capture accelerated enterprise data migrations; target asymmetric payoff 2–4x if Snowflake sustains 5–10% incremental ARR growth from AI workloads. Hedge with modest put protection if cloud spend stalls.
• Long MSFT (Microsoft) — 12–18 months via call spread or long-dated calls. Microsoft’s diversified cloud + enterprise hooks make it a lower-volatility way to play infrastructure monetization; expect steady upside if cloud AI integrations drive +100–200bps incremental Azure growth. Risk: muted if macro weakens enterprise IT budgets.
• Pair trade: Long NOW (ServiceNow) / Short CTSH (Cognizant) — 6–18 months. ServiceNow benefits from workflow automation; regional labor-arbitrage players are exposed to process automation. Target 1.5–2.5x risk/reward; exit or flip if outsourcers show rapid AI-service revamps within two quarters.
• Event trade: Buy ACN (Accenture) on any >6% pullback ahead of next guidance — 6–12 months. Accenture can capture transformation spend and convert project dollars into fractional recurring revenue; reward comes if consulting budgets reallocate to AI modernization. Risk: project delays or client austerity compress margins.