Version 2

Goldman Sachs projects 24x AI agent token growth by 2030 [1][2], McKinsey projects agents could mediate $3-5 trillion in global retail commerce over the same period [4], and OpenAI's internal data shows non-developer agent use grew 137x for individuals and 189x for organizations since August 2025 [3]. The dominant near-term application remains software development: OpenAI's Codex now accounts for roughly 99.8% of the company's internal output tokens [3]. Enterprise cost pressure is the main counter-signal, with Microsoft and Uber both reported to be reconsidering expensive agent deployments [1][2] — though Microsoft simultaneously expanded Copilot in Excel into a finance workflow system with external data connectors and audit trails [10]. DeepMind has added computer use to Gemini 3.5 Flash, making enterprise automation a three-way competitive race among OpenAI, Google, and Microsoft [6].

Multiple research houses now project multi-trillion-dollar stakes for agent adoption, but the economics remain unsettled: enterprises are hitting real cost ceilings even as vendors expand product capabilities. Whether agent cost structures improve fast enough to match projected demand — or whether gains stay concentrated in software development — determines how broadly the agentic thesis materializes across the economy.

Goldman Sachs Research's projection that AI agent token usage will grow 24-fold by 2030 [1][2] has become the anchor figure for the agentic AI investment thesis, and it is now reinforced by quantitative evidence from multiple directions. OpenAI's June 25 economic research paper reported that Codex, its coding agent, went from below 10% to approximately 99.8% of the company's internal output tokens in under a year [3]. Non-developer use of agents at OpenAI grew 137x for individuals and 189x for organizations between August 2025 and the publication date, and 70.2% of sampled users submitted at least one request representing more than one hour of equivalent human work [3]. Heavy users managed five or more concurrent agents simultaneously, with the 99th-percentile user running approximately 71 hours of agent work per day [3]. McKinsey added a separate demand-side projection: AI agents could mediate between $3 trillion and $5 trillion of global consumer retail commerce by 2030, with the highest-level scenario involving AI agents negotiating directly with retailer AI agents on price and shipping [4]. Eric Schmidt's framing — that founding an agentic AI company is the clearest near-term path to profit — circulated widely and is broadly consistent with these projections, though he also acknowledges that as all businesses build competing agents, margin compression follows [5].

On the supply side, the competitive landscape for enterprise agent infrastructure widened on June 24 when DeepMind integrated computer use directly into Gemini 3.5 Flash, previously only available as a standalone product [6]. The integration targets long-horizon enterprise automation tasks including continuous software testing and professional knowledge work. DeepMind paired the launch with two enterprise safeguards: explicit user confirmation gates for sensitive or irreversible actions, and automatic task halting when indirect prompt injection is detected [6]. This positions Google alongside OpenAI and Microsoft in offering enterprise-ready agentic infrastructure, each with different safety and auditability approaches.

The cost-pressure counter-narrative has some corroboration. Microsoft and Uber are reported to be reconsidering expensive AI agent deployments, with multiple news articles describing AI agents as costing more than equivalent human workers [7][8][9]. Against that, Microsoft has also expanded Copilot in Excel into a finance workflow system, adding connectors to licensed data providers including FactSet, Morningstar, PitchBook, and S&P/Kensho, and introducing SKILL.md-defined repeatable workflows for finance tasks such as DCF models and variance analysis [10]. The Plan mode requires Copilot to declare intended edits before execution, and a Show Changes feature labels Copilot edits alongside human edits to create an audit trail [10]. Whether this product expansion reflects a different budget decision than the reported pullback, or whether the two reports describe different deployment contexts at Microsoft, is unresolved.

One concrete business decision attributed to agent adoption: Notion announced it will shut down Notion Mail across all platforms on September 22, 2026, stating that most of its users rely on AI agents for electronic correspondence rather than email clients [11]. The product had been built primarily by former Skiff employees following Notion's February 2024 acquisition. The claim is worth noting as an early instance of a company citing agent adoption as the direct cause of a product discontinuation, though the Ars Technica report covers it without endorsing the justification.