Milos Maricic: listen for the AI number on next week’s bank calls

Bank of America CEO Brian Moynihan cited 130–140 AI projects costing $400 million and generating an $800 million benefit, but it remains unclear whether the figure will appear on the bank’s earnings call. A review of 919 calls at 60 financial firms found only S&P Global had quantified AI savings or revenue; JPMorgan has also not repeated its CEO’s cited $2 billion benefit on earnings calls. The commentary warns that AI returns may take time, accrue to a limited group, erode, and face rising costs, urging investors to seek verifiable figures rather than assume near-term gains.

Analysis

AI is more likely to show up first as gradual expense leverage than as a near-term revenue catalyst. For banks, lower servicing and back-office costs could improve efficiency ratios, but that benefit is only investable when it is net of model, integration, control and employee-transition costs. A headline savings figure without a recurring run-rate, baseline and attribution method should not earn a multiple premium.

The second-order risk runs through AI suppliers: if large users optimize token consumption and switch among vendors, model usage can grow while pricing power and revenue per unit weaken. That makes adoption metrics a poor proxy for supplier economics. Incumbent banks may still gain from proprietary data and distribution, while smaller competitors risk a scale disadvantage; however, third-party tools and supplier flexibility could narrow that gap.

The article’s earnings-call catalyst is dated. Treat the next disclosure as a verification exercise, not as an upcoming event: determine whether benefits are realized in reported expense or revenue, recur, and exceed fully loaded costs. Over 1–3 months, unsupported AI claims are a credibility and sentiment risk, but bank fundamentals remain more directly driven by credit, rates and capital markets. Over 6–18 months, evidence of sustained productivity could support modest operating leverage; compliance expense, implementation delays or rapid imitation could absorb it. The contrarian point is that weak disclosure does not prove weak productivity—but it does argue against paying today for unverified, durable AI rents.

Key Decisions for Investors

• Do not initiate a broad bank short on this evidence alone. For BAC, treat any repeated savings claim as a diligence trigger, not an earnings upgrade, until management reconciles it to recurring reported expense or revenue and discloses the all-in cost base.
• Keep BAC event exposure small until checking subsequent calls and filings; the article’s cited call dates are historical. If a new disclosure is specific and financially reconciled, consider a modest BAC-over-XLF position, with the thesis invalidated if reported expense trends do not confirm the claimed productivity over the following quarters.
• Avoid using AI adoption or token-volume growth as a long signal for model suppliers. Prefer evidence of customer retention, pricing and gross-profit contribution; supplier-agnostic optimization would falsify a thesis based primarily on rising usage.
• Monitor bank efficiency ratios, operating expense guidance and any quantified AI savings net of implementation and control costs. Without that evidence, no standalone trade is warranted; credit quality, deposit costs and market-sensitive revenues remain the more actionable bank drivers.