JPMorgan Chase spent $2 billion on AI in 2025 and saved $2 billion. Clean headline. Terrible benchmark — at least if you're running a $40 billion Canadian regional bank with one model risk officer and a compliance deadline closing fast.
The Reclassification Move Every CFO Should Steal
The most important thing JPMorgan CFO Jeremy Barnum said in 2025 wasn't the savings number. It was the accounting language. He stopped calling AI a technology investment and started calling it core infrastructure — the same reclassification banks ran with ATM networks in the 1980s.
That shift matters more than the 1:1 return ratio. When you reclassify AI as infrastructure, you stop measuring payback in quarters and start amortizing it like a data center. The ROI conversation changes entirely. You're no longer defending a project budget in front of a skeptical board; you're defending a capital asset with a 10-year depreciation schedule.
For Canadian CFOs watching this from the sidelines, the strategic move isn't to copy JPMorgan's spend — it's to copy the framing before OSFI's Guideline E-23 forces a different kind of conversation. E-23, published September 11, 2025 and effective May 1, 2027, requires every federally regulated financial institution to establish enterprise-wide AI model risk management. That compliance build has a price tag. If AI is still sitting in your budget as a series of experimental line items when that deadline arrives, you're going to be explaining both the compliance cost and the ROI shortfall simultaneously.
The Survivor Bias Baked Into the 7-Month Number
ChatFin's March 2026 analysis of 340 finance AI implementations puts the average payback period at 7 months, with 38% of implementations achieving payback in 4 to 8 months. The median 3-year ROI across implementations that reached production came in at 4.2x. Those numbers get cited constantly in board decks right now.
Here's what the denominator is hiding: Gartner warned that 30% of generative AI projects would be abandoned post-proof-of-concept by end of 2025. ChatFin's 340 implementations are, by definition, the ones that made it to production and got formally analyzed. The pilots that died in staging, the models that got shelved after the vendor raised prices, the automation workflows that broke when the core banking system updated — none of those are in the sample.
A CFO who builds their board presentation around a 7-month payback benchmark without accounting for that abandonment rate is setting themselves up for a difficult Q3 conversation. The honest math for a typical Canadian mid-market bank running its first enterprise AI program — factoring in failed pilots, E-23 compliance build, and the near-certain replatforming when the first vendor gets acquired — probably lands closer to 36 months.
According to McKinsey's 2025 Global AI Survey, 62% of financial services executives report difficulty quantifying AI returns. That's not a communication problem. That's a methodology problem, and the 7-month benchmark is making it worse by giving CFOs a number that sounds credible but doesn't survive contact with their actual implementation history.
What Bank of America's 20% Lift Actually Teaches Mid-Market CFOs
Bank of America spent $4 billion on new strategic technology initiatives in 2025, up 44% over the prior decade. The number that matters more than the headline spend: a 20% developer productivity lift from AI coding tools, applied across more than 130,000 employees.
Twenty percent productivity improvement on existing headcount is a number you can model at a $500 million institution. It doesn't require a $2 billion AI budget. It doesn't require reclassifying your entire technology stack as infrastructure overnight. It requires picking a high-volume, well-documented workflow — developer tooling, document review, back-office reconciliation — and running a disciplined pilot with a real control group.
According to the American Banker 2026 AI Talent Shift Survey, 79% of banking organizations increased company-wide AI spending by more than 10% in the past 12 months. The institutions getting the cleanest ROI signals are the ones who started narrow, measured rigorously, and expanded only after the first implementation survived a full business cycle.
The RPA lesson is still fresh enough to be instructive. Three years ago, Canadian banks ran the same playbook with robotic process automation — 18-month payback projections, consultant-led rollouts, aggressive board presentations. By 2023, a material portion of those bots were sitting idle because nobody had budgeted for maintenance when underlying systems changed. RBC and TD both had to restructure their automation governance after discovering that RPA without proper model risk oversight created operational fragility rather than efficiency. OSFI E-23 is, in part, a direct regulatory response to watching that cycle play out.
The E-23 Compliance Cost Nobody's Modelling
Here's what the McKinsey $200 to $340 billion annual value creation estimate for global banking doesn't include: the cost of complying with the regulatory frameworks being built specifically to govern the AI generating that value.
OSFI's Guideline E-23 requires model inventories, validation frameworks, and documentation protocols across every AI and ML model in production. For the Big Six, this is an integration project. For a mid-size Canadian bank or credit union, this is potentially an existential budget question.
According to OSFI and FCAC's September 2024 Risk Report on AI Uses and Risks, 75% of Canadian federally regulated financial institutions plan to invest in AI over the next three years. The same institutions now face a compliance build that requires dedicated model risk officers — a role that barely existed in Canadian banking five years ago and whose compensation is about to spike as the 2027 deadline approaches.
A senior compliance officer at a Vancouver-area credit union, who asked not to be named because their institution hasn't publicly disclosed its E-23 readiness timeline, put it plainly: "We're being asked to build the governance infrastructure for AI we haven't fully deployed yet, on a timeline set by regulators watching the Big Six."
The second-order effects are already in motion:
- Mid-size Canadian banks are likely to acquire AI governance startups rather than build compliance infrastructure from scratch before May 2027.
- OSFI E-23 compliance costs will accelerate credit union consolidation as smaller institutions cannot absorb model risk management overhead.
- Demand for model risk officers in Vancouver and Toronto will spike through 2026 and 2027, temporarily pushing compensation beyond what regional institutions can match.
- CFOs locking in AI vendor contracts before E-23 guidance is fully interpreted risk expensive renegotiations when compliance requirements clarify.
- Banks that reclassify AI as infrastructure will report lower discrete tech ROI, complicating peer benchmarking for institutional investors.
The regulatory fragmentation compounds this. Bill C-27, Canada's proposed Artificial Intelligence and Data Act, died on the order paper before the 2025 federal election. Quebec's Law 25 automated decision-making rules remain the only provincial private-sector framework expressly addressing AI. A national bank running models across provinces is simultaneously subject to OSFI's principles-based E-23 and Quebec's more prescriptive requirements — with no federal legislation to rationalize the overlap. That compliance friction doesn't appear in any of the global value creation estimates, and it disproportionately hits institutions with national retail footprints.
The Audit Question JPMorgan's Number Can't Answer
Banking sector AI spending reached $31.3 billion in 2024 and is projected to reach $84.99 billion by 2030, according to Statista and Juniper Research. Statista separately projects global financial sector AI spending will grow from $35 billion in 2023 to $97 billion by 2027, a 29% compound annual growth rate. The capital is moving regardless of whether the ROI methodology is sound.
The honest version of JPMorgan's $2 billion savings claim is this: it is a disclosure, not an audit finding. Until an external auditor signs off on the attribution methodology — which model drove which saving, over what time horizon, net of what counterfactual — every CFO benchmarking against that number is building their capital allocation case on a figure that has never been independently verified. JPMorgan's Chief Analytics Officer Derek Waldron told McKinsey in October 2025 that AI-attributed benefits have grown 30 to 40% year-over-year since the program's inception. That trajectory is credible. The precise dollar attribution is not auditable from the outside.
This is not an argument against AI investment in banking. The efficiency case is real. The productivity gains at the developer and analyst level are measurable and repeatable. The back-office automation gains are documented — Canadian banks achieved 64% AI penetration in back-office automation by 2023, high relative to their asset base, partly because the Big Six operate on more consolidated core banking platforms than the fragmented U.S. regional system.
The argument is simpler: the CFO who walks into a board meeting with a 7-month payback projection sourced from a survivor-biased sample, benchmarked against an unaudited headline from the largest bank in the United States, and without a line item for OSFI E-23 compliance, is not doing their job. The ones who build the honest 36-month model — with abandonment rates, compliance costs, and infrastructure reclassification baked in — are the ones who will still have their jobs when the next rate cycle compresses margins and the AI budget becomes the most visible line item to cut.
