The problem with ‘human in the loop’ AI? Often, it’s the humans

A flurry of recent studies and corporate moves suggests AI is moving from augmentation toward autonomous value creation: a Vals AI legal‑research test found several AI tools outscored median human lawyers (lawyers 69% vs. ChatGPT 74%, Midpage 76%, Alexi 77%, Counsel Stack 78%), and firms such as Quinn Emanuel report senior professionals extracting outsized productivity gains by vetting AI drafts; similarly, an advertising trial found fully AI‑generated creatives lifted click‑through rates by 19% though disclosure that ads were AI‑made cut purchase intent by roughly a third, creating regulatory and brand‑risk tradeoffs. Market signals underscore rapid commercialisation—Google plans ads in Gemini in 2026, President Trump signalled limited exports of Nvidia H200 GPUs to approved Chinese buyers (triggering U.S. political pushback and likely Chinese procurement restrictions), while startups and partnerships are drawing large capital and enterprise commitments (Unconventional AI raised $475m at a $4.5bn seed valuation; Glean hit $200m ARR; Anthropic/Accenture will train 30,000 employees). At the same time, industry efforts to standardize agent protocols and Google's Titans/MIRAS work on longer‑term model memory aim to reduce integration friction and expand use cases, even as growing skepticism at NeurIPS about LLMs as a path to AGI could shift long‑term R&D and investment priorities.

Analysis

A series of recent studies and corporate moves indicate AI is shifting from augmentation toward autonomous value creation in knowledge industries. Vals AI’s legal-research benchmark showed several AI tools outscored median human lawyers (lawyers 69% vs. ChatGPT 74%, Midpage 76%, Alexi 77%, Counsel Stack 78%), and law firms such as Quinn Emanuel report senior lawyers derive outsized productivity by vetting AI drafts rather than originating research. Academic trials show domain variance: an advertising experiment found fully AI-generated creatives increased click-through rates by 19% but disclosure that content was AI-made cut purchase likelihood by roughly one-third, posing clear brand and regulatory tradeoffs for marketers.

Commercial signals underscore rapid monetization: Google plans to add ads to Gemini in 2026, Unconventional AI raised $475m at a $4.5bn seed valuation, Glean reached $200m ARR (up from $100m nine months prior), and Anthropic’s three-year partnership with Accenture will train 30,000 employees—evidence of enterprise demand and large capital flows. Geopolitical risk is visible: the US decision to allow H200 exports to approved Chinese buyers triggered bipartisan pushback and Beijing may impose purchase justifications or public-sector bans, complicating NVDA’s China opportunity.

Ecosystem and technology developments create both opportunity and uncertainty. Major vendors forming an Agentic AI Foundation and Google’s Titans/MIRAS memory architecture point to integration and capability advances, while growing skepticism at NeurIPS about LLMs as a path to AGI could reallocate long-term R&D spending and affect valuations dependent on continued scaling assumptions.

Key Decisions for Investors

• Favor modestly increasing exposure to enterprise AI platform winners and systems integrators showing traction on deployments and training commitments (e.g., companies involved in Anthropic/Accenture-style deals),
• Monitor NVDA-related geopolitical and regulatory headlines closely and consider position hedges given the political backlash to H200 exports and potential Chinese procurement limits,
• Tilt toward vendors and service providers that contribute to or adopt emerging agent standards and long-term memory architectures, as they are likeliest to win enterprise integration work,
• For consumer-facing and advertising-exposed portfolios, pressure-test revenue assumptions against potential disclosure requirements and brand risk since AI-generated ads raise CTR but materially depress purchase intent when disclosed