CFOs believe AI is paying off. Researchers aren’t so sure—yet

Companies reported AI-driven productivity gains averaging 1.8% in 2025, but revenue-implied gains were much smaller and the paper finds a roughly one-year lag (2025 reported gains align with 2026 revenue-implied gains). Gains are uneven across sectors—high-skill services (notably finance) show the strongest improvements while manufacturing, construction and low-skill services lag. CFOs and researchers emphasize ROI timing: AI benefits appear driven by efficiency and quality rather than capital spending, and require a disciplined multi-year (3–4 year) horizon to justify investments.

Analysis

The headline take-away is timing: reported productivity gains are earlier in the deployment cycle than realized revenue — think a 12–24 month monetization window driven by rollout, pricing experiments, and revenue recognition. That creates a transient mismatch where margins can improve (lower headcount, faster throughput) before the top line scales, compressing near-term ROI metrics but setting up multi-year cashflow improvements if companies can retain pricing or expand share.

Sector-level winners are those owning the stack that must be upgraded once — AI silicon, cloud infra, and embedding layer SaaS — because they capture both one-time capex and recurring ARR; losers are low-margin labor intermediaries whose unit economics are under direct threat from automation. Expect outsized capital spending on GPUs and networking (benefitting suppliers and data-center REITs) even as some professional services see a near-term implementation boost that could morph into disintermediation.

Key second-order effects: (1) Pricing power matters — firms that can re-price services (financials, high-skill B2B SaaS) convert productivity into revenue faster; (2) supply-chain winners will be semiconductor and storage suppliers with sticky order books, while commodity labor markets face downward wage pressure in specific cohorts; (3) regulation or model underperformance are plausible multi-year dampeners. Short-term (days–weeks) volatility will be driven by earnings language on AI payback; 12–36 months is the relevant horizon for profit-cycle realization.

Key Decisions for Investors

• Long NVDA (12–24 month horizon): asymmetric call-buy (e.g., Jan 2027 calls) to capture continued GPU-driven infra spend. Rationale: captures both training/inference demand and pricing leverage; downside is 30–50% implied-volatility decay and competition from custom silicon — size position to 2–4% NAV, target 2–3x upside, stop-loss at 35% of notional.
• Pair trade — Long MSFT vs Short RHI (12–36 months): MSFT benefits from embedded enterprise AI monetization and cloud pull-through while Robert Half faces secular displacement of junior analyst/clerical roles. Rationale: capture ARR expansion vs staffing compression; target 20–40% relative outperformance, limit loss to 15% on either leg.
• Short TTEC or similar call-center outsourcer (12–24 months): automation-first deployments will compress volumes for legacy contact-center providers. Rationale: high probability of unit-demand erosion and margin contraction; size as a tactical short (1–2% NAV), cover on any signs of large contract wins or multi-year outsourcing rebounds.
• Long Equinix (EQIX) or Data-Center REITs (12–24 months): play the infrastructure squeeze from accelerated AI workloads (higher density, more power). Rationale: inelastic demand for colo/connectivity with potential multiple expansion; expected total return 15–30% if capex remains elevated, risk includes capex normalization and rate shock — position as yield-plus-growth (2–3% deployment size).