Divergent Compute.AI Economic Think Tank

Markets / Industries / Financial Services

Industry view · Financial Services

The first industry where AI does the work, not just the assisting

Banking sits on one of the largest generative-AI prizes in the economy — McKinsey puts the annual productivity opportunity at $200-340 billion, equal to 9-15% of operating profit. Yet as of early 2026 only about 23% of banks have moved past pilots into production, leaving the value mostly unbanked.

$200-340B
Annual gen-AI value potential in banking
McKinsey
$1-1.5B
Annual AI business value, JPMorgan
JPMorgan (D. Pinto)
3B+
Erica client interactions since 2018
Bank of America
23%
Banks in production, not pilots
Accenture

01 · The thesis

The value is real; the moat is data and distribution, not the model

Financial services is the cleanest test case for AI's economic claims. The work is text-heavy, rule-bound, and expensive — research synthesis, KYC, fraud scoring, advisor prep, code. McKinsey's $200-340 billion estimate is concentrated where labour cost is highest: middle and back office. JPMorgan already books $1-1.5 billion of annual value from its LLM Suite, now used by 230,000+ employees. These are not slide-deck numbers; they show up in headcount guidance from Citi, Wells, and Bank of America.

The defensible advantage is not the foundation model — every bank rents the same GPT-class engines. It is proprietary transaction data, regulated distribution, and trust. Nubank's nuFormer cut projected credit risk by 70% on comparable segments because it trains on 120 million customers' transaction histories. BlackRock's Aladdin sits on the risk plumbing for tens of trillions in assets. The losers are firms whose only asset was being the human interface between customer and a commoditising back end.

1Data & Models

Proprietary transaction data is the real moat

Banks own decades of labelled financial behaviour that no model lab can replicate, and the winners are turning it into purpose-built foundation models.

Nubank (nuFormer / Hyperplane), BlackRock, JPMorgan
2Risk & Fraud

Real-time scoring is already production-grade

Generative and behavioural AI now adjudicate fraud in tens of milliseconds across global card networks, the most mature deployment in the industry.

Mastercard Decision Intelligence Pro, Visa, Feedzai
3Advisor & Front Office

Copilots reached near-universal advisor adoption

Wealth and private-bank advisors now use AI assistants daily to retrieve and synthesise internal research, compressing hours of prep into seconds.

Morgan Stanley, JPMorgan Connect Coach, BlackRock Aladdin
4Customer Service

The automation frontier that overreached

Chatbots can deflect two-thirds of contacts, but Klarna's public reversal showed quality limits where nuance and complaints meet a model.

Klarna, Bank of America Erica, Nubank
5Software & Ops

Agentic AI starts replacing the engineer, not aiding them

Banks are deploying autonomous coding agents and back-office agents for KYC/AML, the workflows automation could never previously touch.

Goldman Sachs (Devin/Cognition), Citi
Pace of AI disruption by stage — Divergent Compute assessment

02 · Public players & exposure

Who routes through, who gets routed around

We plot the listed players on two editorial axes — how exposed each is to AI disruption, against how ready its data, brand and position are to be the answer. The figures in the table are sourced; the placement is our read.

Positioning — editorial assessment, not a sourced metric. Bubble = approximate relative scale.
CompanyStanceThe sourced fact
JPMorgan ChaseJPMAI compounderEstimates $1-1.5B in annual AI business value; LLM Suite now used by 230,000+ employees (JPMorgan).
Morgan StanleyMSAdvisor edgeAI @ Morgan Stanley Assistant reached over 98% advisor-team adoption; document access rose from 20% to 80% (Morgan Stanley / OpenAI).
BlackRockBLKInfrastructure moatAladdin launched Gen-AI Auto Commentary in Oct 2025 with Morgan Stanley as first client; platform underpins tens of trillions in assets (BlackRock).
MastercardMAFraud railsGenerative AI lifted fraud-detection rates by up to 300%; 42% of issuers saved $5M+ in two years (Mastercard).
VisaVNetwork scaleVisa Advanced Authorization helps prevent an estimated $28B in fraud annually at 76,000+ transactions/second (Visa).
Bank of AmericaBACScaled assistantErica surpassed 3 billion client interactions, now averaging 58M+ per month and resolving 98% of inquiries (Bank of America).
Goldman SachsGSAgentic pilotDeployed Cognition's autonomous engineer Devin alongside ~12,000 developers; GS AI opened to 46,500+ staff (Fortune / Goldman Sachs).
NubankNUAI-native banknuFormer transformer cut projected credit risk by ~70% on comparable segments; serves 120M+ customers (Bloomberg / Building Nubank).
StripeSTRIPAI-economy tollProcessed $1.9T in 2025 (+34%), benefiting from AI-sector customers including Nvidia and Microsoft; valued at $159B (Stripe / CNBC).
KlarnaKLARWalked it backAI assistant handled two-thirds of chats (work of 700 agents), but Klarna reversed course in 2025 and rehired humans on quality (OpenAI / CX Dive).
The map is Divergent Compute’s editorial positioning, offered as a lens, not a measurement. Every figure in the right-hand column is drawn from a named source — see Sources.

03 · The two clocks

The spend, and the payoff

Three timers running against the incumbents

Source: JPMorgan, McKinsey, Mastercard, company disclosures, 2024-2026

Headcount is the clearest clock. Citigroup is moving toward roughly 20,000 job cuts by 2026 (about 8% of staff), with CEO Jane Fraser citing automation and AI; Bank of America, Citi and Wells Fargo have all guided to lower headcounts for the year (Banking Dive / Fortune).

Adoption is racing ahead of governance. 91% of banking executives call AI a strategic priority but only 23% have reached production (Accenture), while the EU AI Act's high-risk rules for credit-scoring systems bite in August 2026 — a compliance wall arriving before most deployments are mature (Accenture / EBA).

The customer-service frontier has a hard ceiling. Klarna's AI deflected two-thirds of chats and saved an estimated $40M, then it rehired humans in 2025 after complaints about generic answers — a reminder that the last 20% of cases resists full automation (Twig / CX Dive).

04 · Private flagships

The AI-native challengers

The companies attacking this industry AI-first, with disclosed funding where available:

JPMorgan LLM Suite

In-house gen-AI platform

Won American Banker's 2025 Innovation of the Year; scaled from zero to 200,000 users in eight months and saves staff an estimated 3-6 hours per week.

Public (JPM); part of an ~$18B annual tech budget

Goldman Devin

Autonomous software engineer

Cognition's AI coder deployed into Goldman's engineering org, targeting 3-4x productivity versus prior tools, with plans to scale to hundreds or thousands of instances.

Cognition raised at a reported ~$10.2B valuation (2025)

Nubank nuFormer

Transaction foundation model

Billion-parameter transformer built on the 2024 Hyperplane acquisition; lifted credit decisioning AUC and cut projected risk ~70% on comparable segments.

Public (NU); Hyperplane acquired July 2024

Mastercard Decision Intelligence Pro

Generative fraud model

Scores transactions in ~50 milliseconds using 'heat-sensing' behavioural pathways; up to 300% improvement in detection rates network-wide.

Public (MA)

Ramp

AI-native finance ops

Expense and spend-management platform automating back-office finance; raised repeatedly through 2025 as agentic finance tooling drew premium checks.

$32B valuation, Lightspeed-led round (2025)

Feedzai

Fraud / RiskOps platform

AI fraud and financial-crime detection sold into banks and acquirers; representative of the fraud-infrastructure layer attracting fresh capital.

$75M raise (2025)

05 · Signals

What moved, and what to watch

2025

Gen-AI value goes from deck to disclosure

JPMorgan publicly attributes $1-1.5B annual value to AI tools, the first major bank to put a hard number on the return.

2025

Customer-service AI hits its ceiling

Klarna reverses its AI-first support strategy and rehires humans, the year's clearest cautionary tale on over-automation.

Jun 2025

Agentic coding enters the bank

Goldman deploys Cognition's Devin into engineering and opens its GS AI platform to 46,500+ employees.

Oct 2025

Data vendors weaponise AI

S&P Global ships ChatIQ and Document Intelligence in Capital IQ Pro; BlackRock adds Gen-AI commentary to Aladdin with Morgan Stanley first.

Aug 2026

The compliance wall arrives

EU AI Act high-risk obligations for credit-scoring systems take effect, forcing explainability and audit trails onto live models.

06 · The exposure read

Who’s defensible, who’s at risk

AI rewards clean, structured advantage and punishes friction. The line runs through who owns the data, the brand and the customer — and who is merely a step the technology can route around.

Defensible

  • Scale incumbents with proprietary data — JPMorgan, Bank of America and Nubank can train on customer behaviour rivals cannot buy, turning AI into a compounding cost and risk advantage.
  • Infrastructure and rails owners — BlackRock's Aladdin, Visa and Mastercard sit at chokepoints; AI deepens their moats rather than threatening them, and fraud AI alone is saving issuers millions.
  • The AI-economy toll-takers — Stripe monetises the AI build-out itself, processing $1.9T in 2025 with Nvidia and Microsoft among new customers.
  • Pick-and-shovel vendors — S&P Global, fraud platforms like Feedzai, and finance-ops tools like Ramp sell AI into every institution regardless of which bank wins.

At risk

  • Pure human-interface roles — junior analysts, customer-service agents and routine coders face the sharpest exposure; Citi is cutting toward 20,000 roles and BofA, Citi and Wells have all guided headcount down.
  • Over-automators who cut too fast — Klarna's reversal shows that gutting human support ahead of model maturity destroys service quality and brand trust.
  • Pilot-stuck laggards — with only ~23% of banks in production, firms still trapped in proofs-of-concept risk a widening productivity gap as leaders compound.
  • High-risk model deployers without governance — credit-scoring and pricing systems face EU AI Act explainability and audit obligations from August 2026, and ungoverned hallucination-prone tools invite regulatory and litigation risk.
The banking AI story in 2026 is not whether the value is real — JPMorgan, Mastercard and Nubank have shown it is — but who captures it. The dividing line runs through data ownership and distribution: institutions that own proprietary transaction data and regulated reach turn AI into a compounding moat, while those whose value was being the human layer over a commoditising back end are on the wrong side of the curve. The constraint now is execution and governance, not capability.

Sources

Where this comes from

McKinsey — Capturing the full value of generative AI in banking  ·  JPMorgan — LLM Suite named 2025 Innovation of the Year  ·  AI News — JPMorgan AI strategy and $1-1.5B value  ·  Bank of America — Erica surpasses 3 billion interactions  ·  Morgan Stanley / OpenAI — AI evals and advisor adoption  ·  Fortune — Goldman Sachs deploys Devin AI engineer  ·  American Banker — Goldman staff write a million gen-AI prompts a month  ·  BlackRock — Aladdin Wealth launches AI commentary tool  ·  Mastercard — AI helping banks save millions on fraud  ·  Building Nubank / Bloomberg — nuFormer and AI credit models  ·  Klarna / CX Dive — AI assistant reversal and rehiring  ·  Stripe — 2025 annual update and $159B tender valuation  ·  Fortune / Banking Dive — Citi job cuts and AI restructuring  ·  S&P Global — ChatIQ and Document Intelligence in Capital IQ Pro