Bank of America tells AI skeptics they are judging the technology at the wrong scale

Bank of America has entered the debate over whether artificial intelligence is delivering real economic lift, publishing a blunt message for people growing skeptical of the AI boom: the problem is not that AI cannot improve productivity, it is that observers are “thinking too small” about when and how the benefits should show up. The note, cited in recent market coverage, frames AI as a special kind of technology, one that can transform a wider share of work than earlier innovations, but whose macro impact may arrive after delays and organizational frictions play out.

In the argument reported by TheStreet and Fortune, Bank of America’s economists point to a disconnect. At the task level, they say AI is already raising output for certain types of work. But the economy-level footprint, measured through gross domestic product and pay, appears close to negligible. The coverage describes an estimate of around 0.1% for AI’s contribution to growth so far, compared with global growth near 3.5%, which would not be noticeable in everyday wages or portfolio performance even if individual productivity gains are real.

Bank of America’s mechanics, as summarized in the reporting, start with a cost-and-availability constraint. The analysis suggests AI can already transform about 20% of workplace tasks, but only 23% of those are cheap enough to automate at today’s prices. It estimates that automated tasks save roughly 27% of labor costs, and that labor is about half of business costs. When those components are combined, the article says the resulting theoretical productivity ceiling is about 0.66% before friction, with the bank’s model compressing that down toward the near-0.1% aggregate figure being observed.

The bank’s more bullish case is tied to the idea that AI’s payoff may follow a familiar technology adoption path often described as a “J-curve,” with improvements accelerating after early implementation hurdles. The reporting attributes the “10x larger” framing to a research approach associated with economist Philippe Aghion and co-authors, which the story says models how falling inference costs and expanding economic feasibility can compound productivity gains over time. Fortune and the linked academic work both describe inference costs as dropping rapidly, with the reported pace of halving every few months functioning as a central assumption behind the long-run math.

Not everyone agrees with the long-term productivity story, and Bank of America’s message comes as other analysts warn that the capital spending cycle around AI is already priced for perfection. In coverage of a separate strategy note by Panmure Liberum strategist Joachim Klement, the AI investment boom is compared to the dot-com era, with a claim that hyperscalers’ planned data-center investment is large relative to the economic returns needed to justify it. The same reporting flags concerns beyond macro timing, including hallucinations in large language models as well as the threat from smaller, specialized models that can perform tasks more cheaply.

Bank of America’s stance on AI is also consistent with how large banks are operationalizing the technology, not just debating it. In an investor presentation filing, the company described “Erica” as an AI assistant used for self-service across the firm, reporting tens of millions of active users and billions of interactions since launch. The same materials describe CashPro Chat and Erica’s handling of a large share of interactions, plus a generative-AI platform used to search, summarize, and synthesize internal research and market commentary, as well as AI tools for software developers and customer-service work. Separately, the company’s Code of Conduct updates emphasize responsible use of AI, requiring the use of approved tools and insisting on human oversight of AI outputs used in decisions or client communications.

Still, important uncertainties remain. The market coverage does not provide the full underlying research note, and it does not spell out all assumptions that would govern whether the modeled productivity path can materialize in time. It also does not address how much of the current 0.1% measurement gap reflects limitations in today’s AI deployment versus limitations in how GDP captures quality improvements. In that sense, the debate may be less about whether AI can improve work, and more about timing, diffusion, and whether organizations can translate task-level gains into durable aggregate growth.