Register B — Analytical · Primary: Bank Technology Officers, Chief Digital Officers · Secondary: Senior Marketing & Product Leaders · ~2,300 words
Generative AI in Banking 2026: What’s Actually Deployed, What Failed, and Why the Harvest Year Is Harder Than the Hype
The pilot era is over. Banks running production gen AI are pulling ahead. The ones still in POC mode are not catching up — they’re falling behind in a compounding way.
Seven trillion dollars a year flows through US commercial banks. For decades the core infrastructure handling that flow — the document processing, the compliance queues, the fraud review, the advisor call prep — ran on rules engines written by people who are now retired, plus armies of knowledge workers doing things that feel important but are mostly pattern-matching. Gen AI doesn’t care about any of that history. It’s just better at the pattern-matching.
But here’s what the listicle crowd keeps missing. This is not a 2023 story anymore. The EY-Parthenon survey of 100 senior retail and commercial banking decision-makers, conducted in early 2025, found that 77% of banks had actively launched or soft-launched gen AI applications — up from a fraction of that two years prior. EY-Parthenon, 2025 GenAI in Banking Survey, n=100 retail and commercial banks, senior decision-maker respondents What they also found: 56% of use cases target internal efficiency, not revenue. Front-office applications are just now reaching production. The money’s been going into the back office first, because that’s where the controls already exist to catch mistakes.
Goldman’s CTO Marco Argenti framed the trajectory last year in plain terms: 2023 was build. 2024 was experiment. 2025 was deploy. His call for 2026 was scaling and harvesting. That’s the frame that matters. Banks that deployed early and got the governance right are now harvesting. Banks still in pilot mode are harvesting nothing and watching the gap widen. Source: Goldman Sachs Institutional Client podcast, recorded May 24, 2025, Marco Argenti
So. What actually works in production. Where the documented failures are. And what the deployment data implies for the next twelve months.
The Three Use Cases With Documented Production ROI
Not six. Not twenty-five. Three. That’s where the production evidence is actually strong enough to hold weight in a board deck without a bucket of caveats. The rest are real — just less proven at scale.
1. Knowledge retrieval for advisors and analysts
Morgan Stanley is the canonical case here and it’s worth going into the actual numbers. The firm deployed its AI @ Morgan Stanley Assistant — a GPT-4-powered chatbot using retrieval-augmented generation against its internal research corpus — to its wealth management division in September 2023. By the time the Debrief tool launched in mid-2024, 98% of advisor teams were using the Assistant daily. Document access jumped from 20% of the firm’s knowledge base to 80%. Advisors report saving approximately 30 minutes of administrative work per client meeting. OpenAI case study on Morgan Stanley; Morgan Stanley press release, June 2024; CDO Magazine interview with Atul Dalmia, Chief Analytics Officer MSWM, October 2025
98% adoption in enterprise software is not a number you see. Ever. The reason it happened at Morgan Stanley is worth understanding: they didn’t build a generic AI tool. They built a tool that sits directly between an advisor and 100,000+ proprietary research documents they already wanted access to but couldn’t surface fast enough. The value was immediate, personal, and daily. JPMorgan has its own version — ChatCFO for its finance team — with JPMorgan’s leadership estimating the AI use case portfolio is valued at between $1 billion and $1.5 billion across the firm. Source: JPMorgan investor day, May 2025, per CNBC coverage Directional figure from company; no independent audit confirmed.
Second-order mechanism
The reason knowledge retrieval works in banking when other gen AI use cases stumble is that the failure mode is immediately legible. If the AI retrieves the wrong research document or misattributes a fact, the advisor sees it and corrects it before it reaches a client. The second-order problem — degraded outputs that look correct but aren’t — is partially mitigated by the human who’s checking the answer for professional reasons anyway. Contrast this with fully automated fraud detection or credit scoring, where a wrong answer has no immediate human in the loop to catch it before it causes harm. That asymmetry is why knowledge retrieval reached production first. The error is catchable.
2. Code generation and legacy modernization
Goldman Sachs is the extreme case. The firm deployed GitHub Copilot across its developer workforce years ago and documented productivity gains of around 20%. Then in July 2025, Goldman announced deployment of Devin — an autonomous coding agent from Cognition — across its 12,000 developers, projecting 3x to 4x productivity gains versus the Copilot baseline. CTO Argenti described this as the ability to manage entire software lifecycles autonomously. Source: Goldman Sachs announcement July 11, 2025; Lucidate Substack analysis July 22, 2025 — directional, not independently audited
The strategic logic is obvious once you know the constraint. Banks still run COBOL code from the 1970s and 1980s handling core transaction processing. The developer pool who can read and modify that code is literally dying — not metaphorically. Gen AI that can translate legacy code to modern languages, generate tests, and accelerate refactoring isn’t a nice-to-have. It’s a succession plan. Goldman also separately made headlines when CEO David Solomon noted that gen AI can complete 95% of an IPO prospectus in minutes — work that previously required six people over two weeks. Goldman Sachs, David Solomon remarks; reported widely including Financial Times, 2025 Solomon’s framing: “The last 5% now matters because the rest is now a commodity.”
3. Document processing and loan origination
Banks run on paper — or PDFs, which are just paper with worse search. Mortgage applications, KYC forms, audit reports, compliance filings. Most of this is still processed manually or through rigid OCR systems that break when a format shifts slightly. Gen AI that can read and extract from any document structure — scanned PDFs, typed forms, email attachments — and cross-reference against lending criteria, flag gaps, and generate preliminary assessments is demonstrably faster than the current process. Chatbot containment rates of 40–60% for document queries — meaning queries fully resolved without human escalation — are appearing in production deployments, per multiple vendor reports. Fluid AI deployment summary, April 2026; note: vendor-reported figures, treat as directional The FDIC confirmed in congressional testimony two weeks ago that banks are actively deploying gen AI specifically for loan applicant financial information summarization. FDIC Testimony, April 2026, FDIC.gov
“The deployment gap between knowledge retrieval and full process automation isn’t a technology gap. It’s a liability gap. Banks know exactly what happens when automated credit decisions go wrong.”
Editorial synthesis — sources: EY-Parthenon GenAI Banking Survey (2025), FDIC congressional testimony (2026), Goldman Sachs CTO Marco Argenti, May 2025
The Failure Case Nobody Wants to Lead With: Klarna
2023. Klarna decides to automate the equivalent of 700 customer service agents with an OpenAI-powered chatbot. By February 2024 the company claims its assistant is handling 75% of customer chats — 2.3 million conversations across 35 languages. The numbers sound incredible. They are, in the most technical sense, real: the bot handled that volume. And then.
By early 2025, customer satisfaction ratings had dropped, repeat inquiry rates climbed 25%, and CEO Sebastian Siemiatkowski told Bloomberg that “cost unfortunately seems to have been a too predominant evaluation factor.” Klarna started hiring human customer service agents again. Siemiatkowski, Bloomberg interview, May 2025; Fintech Weekly, May 12, 2025
The failure mechanism is specific and teachable. Klarna’s AI couldn’t initially explain why a customer was rejected from a service. It struggled with dispute resolution and fraud-adjacent inquiries that required judgment. It routed complex cases to human agents — but only after the customer had already been through a long, unproductive conversation. The escalation was a worse experience than just reaching a human in the first place. Twig.so analysis of Klarna customer service failures, March 2026
This matters for banking specifically. The transaction types that generate the most support volume are exactly the ones with the highest consequence — disputes, fraud, declined transactions. A chatbot that handles routine balance inquiries brilliantly and breaks on fraud disputes is not serving the customers who most need serving. That asymmetry is not a technology problem; it’s a deployment design problem. Klarna is now on a hybrid model. CMSWire interview with Siemiatkowski, May 2025 — note: the narrative around Klarna’s “reversal” is genuinely contested. Siemiatkowski pushed back on the “failed AI” framing, citing two-minute resolution times vs. 11 minutes for human agents and sustained 66% automation. The quality degradation in complex cases is documented; the aggregate scale story is more nuanced than most coverage acknowledged.
The lesson — which some banking technology leaders are applying — is not “don’t automate customer service.” It’s: automate the queries where AI demonstrably equals or outperforms human resolution speed and quality. Route everything else immediately. Don’t make the customer wait to figure out which bucket they’re in.
Cross-source synthesis — not present in any single cited source
Goldman’s internal knowledge retrieval achieved 98% adoption and demonstrable productivity gains. Klarna’s external-facing chatbot achieved volume but degraded quality on high-stakes interactions. The variable that separates these outcomes is not the AI model — both used GPT-4-class systems — it’s the consequence asymmetry of errors. Morgan Stanley’s advisors checked AI outputs before using them professionally, creating a human catch layer. Klarna’s customers had no equivalent layer and experienced the failures directly. EY-Parthenon’s finding that 56% of banking gen AI use cases target internal efficiency (not direct customer-facing revenue) may reflect this learned lesson more than it reflects technology limitations: banks are building where the error-catch layer already exists.
What’s Coming in the Back Half of 2026: Agentic AI and the Governance Problem
Here’s where it gets weird. The tier of use cases that matter in 2026 aren’t the ones everyone wrote about in 2023. Those — chatbots, document summarization, fraud pattern detection — are table stakes now. The frontier is agentic AI: systems that don’t just answer questions but execute multi-step workflows autonomously.
Goldman’s Devin deployment is the clearest example in production. Autonomous coding agents that don’t just suggest code — they write, test, debug, and deploy. Goldman is also reportedly experimenting with autonomous deal-sourcing: AI models that analyze regulatory filings, supply chain data, and macroeconomic signals to identify M&A targets before a human banker has been briefed. Chronicle Journal, January 15, 2026; directional reporting — Goldman has not publicly confirmed deal-sourcing deployment at production scale
The governance problem with agentic AI in banking is not hypothetical. KPMG’s survey of US banking executives found 68% planning gen AI deployment for regulatory compliance and risk avoidance. KPMG GenAI in Financial Services report; exact survey date not confirmed in available sources — treat as directional That’s the function where autonomous errors have the highest regulatory consequence. An agentic system that misfiles a suspicious activity report or incorrectly flags a transaction as compliant is not an inconvenience. It’s an exam finding.
The FDIC has been explicit about its posture: technology-neutral, open-minded, but with the expectation that banks manage associated risks in a safe and sound manner. FDIC Chairman Hill and FDIC testimony, April 2026, FDIC.gov The regulator is not standing in the way of gen AI deployment. It’s watching to see which banks build the governance architecture before expanding autonomous scope, and which ones expand first and bolt governance on after something breaks.
The EY-Parthenon finding that organizational governance is emerging as a “key element of success” for banking gen AI — not a constraint, a success factor — is the tell. Banks getting this right are treating model behavior documentation, audit trails, and human oversight as deployment prerequisites, not post-hoc compliance theater.
| Use Case | Evidence Level | Named Deployment | Documented Result | ⚠ Limitation |
|---|---|---|---|---|
| Knowledge retrieval (internal) | Strong | Morgan Stanley AI @ Morgan Stanley | 98% advisor adoption; 30 min saved per meeting; document access 20% → 80% | Results from a single firm with exceptional proprietary data density; replication at community banks or mid-tier institutions is unverified |
| Autonomous code generation | Directional | Goldman Sachs / Devin deployment | 3–4x projected productivity gain vs. prior Copilot baseline; 12,000 developers in scope | Productivity projection from Goldman’s own CTO; no independent audit. Earlier Copilot data (20% gain) from GitHub 2024 research, which showed wide variance by task type |
| Customer service automation | Moderate | Klarna (fintech, not a bank); various banks in pilot | 40–60% containment rates for routine queries; 2-min vs. 11-min resolution for handled cases | Klarna documented quality degradation on complex/high-stakes queries; satisfaction dropped before hybrid model was implemented; banking regulatory requirements around dispute resolution add constraints absent in BNPL context |
| Document processing / loan origination | Directional | Multiple unnamed deployments; FDIC-confirmed category | Faster preliminary assessment; format-agnostic extraction | No named bank has published before/after throughput data with disclosed methodology; vendor-reported figures dominant in available sources |
| Compliance and AML monitoring | Directional | Category named by KPMG; 68% of execs plan deployment | Faster anomaly detection; AI-generated summaries for compliance analysts | Fully autonomous compliance decisions create regulatory exposure; human review requirement likely makes cost asymmetry unfavorable vs. assisted-AI model in near term |
The Finding Every Competitor Missed: Why 2026 Is Harder Than It Looks
Everyone’s writing about what gen AI can do in banking. Nobody’s writing about why the scaling year — which Goldman specifically called 2026 — is harder than the deployment year.
Here’s the thing that’s structurally weird about this moment. The EY-Parthenon data shows 61% of surveyed banks already report substantial impacts from gen AI. The investment has shifted: 65% of gen AI spend now comes from IT and technology budgets, down from 27% in 2023 that came from broader corporate strategy funding. EY-Parthenon, 2025 That’s a pragmatic shift. IT owns the tools now. The question is whether the business units know how to use them at scale.
The deployment-to-scaling gap is real and not well understood by most banking leadership teams. Deploying gen AI to 10,000 Goldman employees is a different problem from sustaining quality at 10,000 users for two years. Models drift. Data quality erodes. The monitoring stack designed to catch system failures wasn’t designed to catch AI output quality degradation that looks like normal performance variance. Banks learning this the hard way in 2026 will be the Klarnas of the next news cycle — except in credit underwriting or compliance, where the consequence is an OCC finding rather than a bad Glassdoor review.
“The model doesn’t fail loudly. It drifts quietly into producing outputs that look right until someone pulls the audit trail and realizes the last three months of decisions rested on degraded signal.”
Editorial synthesis — sources: Goldman Sachs CTO Marco Argenti (2025); EY-Parthenon GenAI Banking Survey (2025); Fluid AI deployment analysis (2026)
The McKinsey $200–$340 billion annual value potential for banking — that number is real and it’s cited everywhere. McKinsey Global Institute, exact report date varies across citations; figure described as potential, not current; treat as directional forward projection What McKinsey didn’t emphasize in the headline: that range assumes banks can sustain deployment quality at scale, govern agentic systems effectively, and close the talent gap on AI orchestration. Three things that are currently unevenly distributed across the industry. The gap between $200 billion and $340 billion is basically the gap between banks that govern well and banks that don’t.
57% of planned banking AI use cases are being built through external partnerships rather than internal development. EY-Parthenon, 2025 That’s the pragmatic choice. It’s also a governance complication — vendor model updates can change output behavior without the bank’s full awareness. A bank that outsourced its credit scoring AI layer to a vendor and didn’t build version-pinning and output monitoring into the contract is flying blind on model drift. This is not a hypothetical edge case.
“Scaling is where the governance architecture either holds or doesn’t. Pilots are forgiving. Production at 10,000 users for 24 months is not.”
Editorial synthesis — sources: EY-Parthenon (2025); Goldman Sachs CTO statements (2025–2026); FDIC testimony (April 2026)
What To Do With This
Two audiences here. The technology decision-makers who own the deployment. And the marketing and product leaders who have to figure out how to position and build on top of whatever the technology team delivers.
For: Bank Technology Officers / Chief Digital Officers
The question isn’t whether to deploy. It’s whether your monitoring architecture will catch drift before a regulator does.
The 2026 deployment problem isn’t building gen AI applications — you’ve done that or you’re doing it. The specific thing that creates regulatory exposure over the next 18 months is deploying without model output monitoring that catches quality degradation independent of system-level health metrics. Your observability stack tells you if the model is returning responses. It probably doesn’t tell you if the responses are as accurate this month as they were six months ago.
Build output quality tracking before expanding to high-consequence use cases (credit, compliance, fraud decisioning). This means periodic human review of AI outputs against ground truth, not just hallucination detection. If you’re using a vendor model, put version-pinning and behavior-change notification requirements in the contract before renewal. One model update can shift output distributions in ways that look fine in aggregate and are wrong in the tail cases — which in banking are often the legally significant cases.
The FDIC’s “technology-neutral” posture is not a green light for autonomous AI decision-making without audit trails. It’s a signal that the regulator will assess your governance architecture as part of how they evaluate the deployment. Build the audit trail first, not as a documentation exercise after the fact.
Stop using pilot containment rates as the primary success metric for customer-facing deployments. Klarna’s bot handled 75% of chats. Customer satisfaction still fell. Containment rate and resolution quality are different metrics. Track both.
For: Senior Marketing & Product Leaders
Your AI roadmap depends on data quality infrastructure your team probably didn’t build and can’t control.
The personalization and next-best-action use cases — the ones that drive CLV and product attachment — are not blocked by AI model capability. Morgan Stanley’s advisors are already generating personalized recommendations that account for individual client context. The blocker, for most banks not at Goldman or Morgan Stanley scale, is first-party data quality. AI personalization trained on clean, unified customer data is transformative. AI personalization trained on fragmented CRM data, inconsistent transaction categorization, and siloed product data produces confident-sounding outputs that are wrong about individual customers in ways that erode trust instead of building it.
The 12–18 month first-party data rebuild that your data engineering team has been requesting is not a tech project. It’s your AI personalization roadmap. If that rebuild isn’t funded in your annual plan, your gen AI personalization initiatives in 2026 are building on sand, and you’ll be explaining to your CMO why the AI-driven campaign underperformed relative to the vendor pitch deck.
The EY-Parthenon data showing 65% of gen AI spend now in IT budgets, down from broader corporate strategy funding, means the product team’s leverage on gen AI investment allocation is compressing. Make the case for data infrastructure funding using revenue-per-customer modeling, not technology framing.
Don’t pitch AI-powered personalization as a differentiator to customers without legal review of your consent architecture. The LinkedIn GDPR finding — consent ruled not “freely given, sufficiently informed, specific, or unambiguous” — was specifically about behavioral targeting consent. Pre-checked boxes and bundled consent fail that standard in EU jurisdictions and are increasing regulatory targets in US state privacy frameworks. A personalization program built on inadequate consent is a liability, not a product feature.
The McKinsey number — $200–$340 billion in potential annual banking value — will probably prove accurate in aggregate over the decade. The question is which banks capture it and which ones fund the vendor fees, build the technical debt, and hand the ROI to the institutions that got the governance right in 2026. That’s not a technology question. It’s a management one. And management questions are the ones that don’t resolve by adding another POC to the portfolio.
Continue Reading on AI Earner Hub
Generative AI in Financial Services 2026 →
AI Governance in Banking: The 2026 Playbook →
Agentic AI: From Hype to Production →
Detecting and Preventing Model Drift in Financial AI →
AI Business Success Stories 2026
Generative AI Freelance Jobs: Unlock High-Paying Opportunities in the Booming AI Gig Economy