The six largest US banks reported $47.3 billion in net income for the first quarter of 2026. At the same time, they cut nearly 5,000 net jobs. Both numbers sit on the same line of the same quarterly report, and that line is increasingly being written in code.
The fact: a record quarter, a shrinking workforce
According to Bloomberg, mid-April earnings confirmed a dual movement. Wall Street delivered an exceptional profit quarter while headcount retreated at a pace not seen since 2023. The split is clear: Wells Fargo shed roughly 4,200 roles, Citigroup and Bank of America cut more than 1,000 each, while JPMorgan Chase and Morgan Stanley kept hiring. Net of those flows, the largest US banks lost nearly 5,000 jobs in three months, compared with 707 in the same quarter a year earlier — seven times as many cuts for a significantly better income statement.
Executives do not flag this as "AI layoffs", but their talking points point there. Charlie Scharf, CEO of Wells Fargo, touts 23 consecutive quarters of headcount reduction while "increasing tech investment, especially in AI". Brian Moynihan, running Bank of America, speaks of "real benefits today" from AI. On the other side of the Atlantic, BNP Paribas booked roughly 1,200 role reductions tied to the Axa Investment Managers integration in January, and Societe Generale confirmed 900 exits from its Paris headquarters. A Morgan Stanley study relayed by the Financial Times projects 200,000 fewer banking roles in Europe by 2030, with efficiency gains "of up to 30%" attributed to AI and digitisation, concentrated in central functions, back-office, risk and compliance.
AI is no longer a budget line option
For two years the banking narrative on AI read like a procession of pilot announcements. In 2026 the tone shifts. Wells Fargo now strings 23 consecutive quarters of headcount decline, nearly six full years without replacement hiring. When the same management expands the "technology and AI" line item, arithmetic says one thing: a portion of the human hours that came off the balance sheet will not come back in through another door. It is neither ideological nor forward-looking. It is P&L arithmetic that moved from industrial promise to visible cost item.
The European case rounds the picture off. The jobs Morgan Stanley is modelling are the ones an LLM plugged into internal databases handles best: data extraction, note drafting, first-line controls, regulatory summarisation. This is precisely the surface Anthropic is targeting with Claude for Financial Services, now connected to Databricks and Snowflake to unify market feeds and internal data. The vendor pitch has become much sharper in a year: they are no longer selling "an assistant", they are selling partial replacement of document production pipelines.
A real-time capital-labour trade-off
The classic reading of banking results pits the rate cycle against market volumes. That first driver is still strong: the ECB will publish its next decision on 30 April 2026, with the refi rate still at 2.15% and internal debate on a possible hike in June, a scenario CNBC reported on 16 April. But a third driver is gaining weight: the cost-to-income ratio being pulled down by software. If the Morgan Stanley range holds, 30% of retail and central-function operating costs avoided is the equivalent of 30% additional capacity in dividends, buybacks, or prudential reserves. The decision scales up.
This trade-off is now structural. A dollar invested in an AI pipeline no longer competes against a five-year ROI promise; line by line, it competes against a dollar of salary that hits the books every month. The differential tips the balance. On the supervisory side, the signal is just as loud: Banque de France and the EBA are mapping these scenarios, and the AI Act will, from August 2026, require an enhanced regime for "high-risk" AI systems, including credit scoring and insurance pricing. The risk and compliance functions thinned out by the cuts will simultaneously need to absorb a new regulatory framework on deployed models: fewer FTEs, more controls, supported by tools that are heavier to operate.
What Q1 really says
Q1 2026 is not a routine quarter. It is a pivot quarter. Executives do not title it "AI layoffs" because the phrase would trigger a political and union backlash, but the reporting structure says it for them. The old rule, that rising profits pulled hiring along, no longer holds at the same point of the curve. Monetary and prudential authorities will have to rule on what this discontinuity implies. Can under-capitalised control functions still cover an operational shock or a model shock? Regulators meeting at the IMF and World Bank spring meetings between 13 and 18 April already placed AI on the list of systemic risks to watch; the Irish Times reported on 17 April that those discussions were dominated by the capacity of the latest models to become a prudential issue of their own.
A second, more cyclical, implication: the cost-compression dynamic blurs. A bad market quarter used to absorb part of the shock on the results side via headcount reduction, betting on a later rebound. When a major share of that adjustment variable has already passed through automatable functions, banks hold less leverage for the next soft patch; the next compression will then reach for core functions or entire business lines. The elasticity of the model shifts, and with it the way a board should read a "good" quarter.
What this changes for my own path
I am joining AFD this autumn as an ALM apprentice, after being admitted to the MSc Data & AI for Finance from Albert School x Mines Paris-PSL. This trajectory hits the topic head-on. Asset-liability management is, by nature, a function that consumes a lot of data, a lot of rate scenarios and a lot of reporting. It is precisely the sort of function that a well-parameterised LLM can accelerate. I experience this on my personal Bloomberg Dashboard, where I use Claude to pull and comment on macro series before crossing them with my own notes and Python scripts.
This is not substitution, it is amplification: the ALM operator of 2026 will not be replaced by a model, they will be evaluated on what they produce with one. Banks cutting headcount today are probably trying to reach that second floor faster. From a student or candidate standpoint, the useful question is no longer "will AI take my job", but "which part of the balance sheet can I do more on with AI than without, and under which prudential constraints". That is exactly the grammar the MSc Data & AI for Finance aims to teach, and the grammar my ALM apprenticeship at AFD will force me to apply on real portfolios.
Take-away
A record profit and 5,000 fewer jobs on the same page of a quarterly report: this is the new grammar of large banks. What remains to be seen is how many quarters it takes before a regulator, a union, or a board of directors asks the forbidden question out loud: what share of these profits is actually produced by AI, and who receives that share?