MIT created duplicate AI workers to tackle thousands of different tasks. The verdict? Most of the time AI is still just ‘minimally sufficient’

AI models achieved a "minimally sufficient" score (7 on a 1–9 scale) for roughly 65% of primarily text-based tasks as of late 2025; MIT estimates model capability gains of up to ~11 percentage points per year and projects 80–95% of such tasks will reach that threshold by 2029. However, models never exceeded a 50% probability of attaining "superior" (9) quality and continue to fail on multi-step, creative, or high-precision work, with real-world fabrications reported in Deloitte and media outputs—implying continued need for human oversight and limited near-term displacement of skilled roles.

Analysis

The near-term market consequence is not a binary replacement of headcount but a structural reallocation of labor toward verification, orchestration and high-stakes oversight. Firms that monetize trust — legal/financial research vendors, evidence-grade content providers, and enterprise QA marketplaces — will see step-function demand for premium, auditable outputs and R&D budgets reallocated from pure model access to human-in-the-loop systems over the next 12–36 months.

Second-order supply effects: cloud/GPU capacity remains necessary but no longer sufficient; customers will buy integrated stacks (model + human QA + provenance) rather than point-inference services, widening moats for incumbents that can bundle data, compliance and workflow (think long sales cycles, sticky ARR). Conversely, pure-play content mills and low-quality AI tooling face margin pressure as clients internalize the cost of downstream correction and legal exposure, turning what looked like a cost-savings into a new operational line item.

Key catalysts to watch are two-fold and asymmetric: a single large hallucination or regulatory enforcement action could reset adoption timelines within weeks, while a genuine model-architecture breakthrough would accelerate displacement but likely concentrate value with platforms that capture verification flows. Position sizing should reflect this binary: favor cash-flowing incumbents and human-in-the-loop platforms with 12–36 month horizons, underweight speculative AI-native media/content names and hedge with cybersecurity exposure to protect against adversarial exploitation risk.

Key Decisions for Investors

• Long Thomson Reuters (TRI) — 12–24 months. Buy equity on any pullback; rationale: recurring revenue and demand for auditable research/precedent tools should allow 20–35% upside as firms pay for verified outputs. Risk: regulatory/tech disruption could cap gains; use a 10–12% stop-loss or size to 3–5% of tech exposure.
• Long NVIDIA (NVDA) with short-dated covered calls — 6–12 months. Stay long for continued GPU demand from training/inference; sell 1–3 month calls 5–10% OTM to finance carry and blunt volatility. Reward: capture secular capex; risk: multiple contraction from macro; covered calls improve risk-adjusted return.
• Long Upwork (UPWK) or Fiverr (FVRR) — 3–9 months. Exposure to elastic human-in-the-loop labor demand as companies outsource verification/QC; buy stock or 6–12 month calls to play volume rebound. Risk/reward: high variability — set a 15% stop and target 30–50% upside if adoption accelerates.
• Long CrowdStrike (CRWD) via a 6–12 month call spread — 6–12 months. Cybersecurity spend will rise as AI increases attack surface and fraud vectors; use call spreads to limit premium outlay. Expected payoff: asymmetric upside with limited cost; tail risk is enterprise budget cuts but near-term catalyst is elevated security incidents.