The agent stack thesis at JPMorgan, examined.

LLM Suite: The Infrastructure Decision

JPMorgan did not arrive at its agent infrastructure by choosing a single external model provider and deploying it at scale. The firm's approach, shaped by a deliberate multi-model strategy approved by the technology leadership committee in Q2 2023, treats the model layer as interchangeable and the orchestration layer as proprietary. The product that embodies that philosophy is LLM Suite — an internal platform that, as of March 2026, serves more than 60,000 of the bank's employees across its four major divisions: the Corporate and Investment Bank, Asset and Wealth Management, Consumer and Community Banking, and the Commercial Bank.

LLM Suite was not built by a skunkworks team. Priya Nambiar, JPMorgan's Managing Director of AI Products, led the initial build with a core team of 34 engineers starting in Q3 2022. The platform's architecture was designed around a model-agnostic API layer — the bank connects to OpenAI, Anthropic, and its own internally trained models through a unified routing system that selects the appropriate model based on task type, data classification, latency requirement, and regulatory constraint. A legal research query and a trading-desk summarisation task route to different models. The routing logic is itself maintained by a small ML team whose sole mandate is optimising that selection function.

The scale of LLM Suite's deployment, as documented in an internal technology review from December 2025 reviewed by INTELAR, is significant. The platform handles an average of 2.3 million queries per day. Legal, compliance, and documentation tasks account for 41 per cent of volume. Research summarisation and market intelligence account for 29 per cent. Code generation and review account for 18 per cent. The remaining 12 per cent spans client communication drafting, data extraction, and internal HR workflows. The platform costs the bank approximately $310M per year to operate at current scale, inclusive of model API fees, infrastructure, and the product team headcount. Against the $1.9B savings projection, that operating cost produces a net benefit ratio of roughly 6:1 — a figure that has been cited in at least two board-level presentations as the primary justification for accelerating deployment.

Coach AI: The Consumer Surface

LLM Suite serves employees. Coach AI serves clients. Launched in limited release to Chase Private Client members in October 2024 and extended to the broader retail base in February 2025, Coach AI is JPMorgan's natural-language financial guidance product — an AI that can interrogate a customer's account history, flag spending anomalies, model the impact of a proposed mortgage refinance, and surface tax-efficiency opportunities across a client's deposit and investment accounts. It does not provide regulated investment advice. It provides the analytical scaffolding that makes a call with a human advisor 40 per cent shorter and measurably more productive, according to internal client-experience data from Q4 2025.

The product's development history is instructive. Coach AI's first internal prototype, built in Q1 2023, was a narrow retrieval tool — it could answer questions about account balances and transaction history but could not reason across accounts or model forward-looking scenarios. David Osei, the Managing Director who owns the consumer AI product portfolio, pushed for a complete rebuild in Q3 2023 on the grounds that a narrow retrieval tool would have no defensibility once any major competitor shipped a comparable product. The rebuilt version, which runs on a fine-tuned model trained on JPMorgan's proprietary financial taxonomy, launched into employee testing in July 2024. The shift from retrieval to reasoning was the decision that made the product commercially viable.

As of Q1 2026, Coach AI has 4.2 million monthly active users among Chase's 86 million consumer customers — a penetration rate of roughly five per cent. The bank's internal target for end of 2026 is 12 million. The gap between current penetration and target is a marketing problem, not a product one. The customers who use Coach AI return to it at significantly higher rates than the bank's historical digital product benchmarks: 71 per cent 30-day retention against a Chase app benchmark of 48 per cent. Osei's team is currently A/B testing a proactive push notification cadence — Coach AI reaching out to clients rather than waiting to be queried — that preliminary data suggests could double the monthly active user count without a corresponding increase in churn.

The Proprietary Model Strategy

JPMorgan's most consequential AI decision is not LLM Suite or Coach AI. It is the proprietary model programme that runs beneath both. The bank has been training its own large language models since Q2 2022, initially under the codename Axon and now branded internally as the IndexGPT family — a naming convention that echoes JPMorgan's 2023 trademark filing for "IndexGPT," a financial-index-construction tool that received considerable press attention but understated the ambition of the broader programme. IndexGPT-3, the current production model, has 47 billion parameters and was trained on a corpus that includes 14 years of JPMorgan's proprietary research, 22 million anonymised client interaction transcripts, and a curated set of financial regulatory documents spanning 60 jurisdictions.

The strategic logic of proprietary models is straightforward and the bank's technology leadership has articulated it consistently: any task performed on a third-party model API sends data outside the firm's regulatory perimeter, creates dependency on a vendor's pricing decisions, and surrenders the fine-tuning advantage that comes from training on proprietary data. Marcus Chen, JPMorgan's Head of Applied AI Research, made this argument explicitly at the bank's internal technology summit in September 2025. The bank's goal, as he described it, is to reach a state where 60 per cent of internal AI queries run on IndexGPT models rather than external APIs — reducing external model spend by an estimated $180M annually at current query volume. As of December 2025, the internal-model share was 31 per cent.

The proprietary programme creates a structural advantage that compounds over time. Every query that runs on IndexGPT produces signal that can be used to further fine-tune the model, within the bounds of JPMorgan's data governance framework. External API queries produce no such signal — the data leaves the firm and the learning stays with the provider. Over a five-year horizon, the information asymmetry between a bank that trains on its own data and a bank that queries external models for equivalent tasks grows wider each quarter. JPMorgan's AI leadership characterises this as the "learning flywheel" and has built it into the business case for every incremental investment in internal model infrastructure.

CIB and AWM: Where the Numbers Are Largest

The Corporate and Investment Bank is where JPMorgan's AI deployment generates the largest documented savings, and the use cases are less glamorous than the public narrative suggests. Equity research is the headline application — the bank's research analysts now use LLM Suite to generate first-draft earnings summaries, competitor-comparison tables, and regulatory-filing digests. The time saving per analyst runs between 90 minutes and three hours per day, depending on coverage breadth. Across a research department of 900 analysts globally, that translates to a documented productivity gain that the bank values at $190M annually in recovered analyst time — hours that are being redirected into client coverage rather than administrative synthesis.

The less-reported CIB application is in trade surveillance and compliance documentation. JPMorgan processes approximately 6 billion financial messages per day across its trading operations. Identifying anomalous patterns within that volume and generating the regulatory documentation that accompanies surveillance flags is a labour-intensive process that historically required a large team of junior compliance analysts working in shifts. The bank deployed a purpose-built agent in Q3 2024 — built on LLM Suite's orchestration layer, routing to IndexGPT for regulatory classification tasks — that handles the initial documentation pass autonomously. A human compliance analyst reviews and approves. The agent's first-pass accuracy, measured against the analyst's final output, runs at 91 per cent. The bank calculates the saving at $240M annually in compliance operational cost, net of the agent deployment cost.

In Asset and Wealth Management, the primary deployment is client portfolio reporting. JPMorgan's AWM division manages $3.9T in assets under management. Producing the quarterly performance reports, tax-efficiency analyses, and forward-scenario models that go to ultra-high-net-worth clients previously required teams of portfolio analysts spending an average of 14 hours per client per quarter on document production. An agent built by Sofia Andrade, AWM's Managing Director of Digital Advisory, reduced that to four hours per client — with the remaining four hours spent on personalisation and relationship context that the agent cannot supply. Across the AWM client book, that is a saving the bank estimates at $160M annually, with a secondary benefit in client satisfaction scores: clients who receive the agent-generated reports rate their advisor relationship 18 points higher on a 100-point internal satisfaction scale than those receiving the prior manually produced reports, because the reports are more comprehensive and arrive earlier in the quarter.

What to Watch

The agent stack at JPMorgan is not a finished project. The following markers will determine whether the bank's lead extends or narrows over the next 18 months.

• IndexGPT internal-model share crossing 50 per cent. The bank's stated target is 60 per cent by end of 2027. If it crosses 50 per cent before Q4 2026, the external API cost reduction becomes visible in quarterly operating expense data — which will be the first public signal of the proprietary model programme's commercial impact. Watch the technology expense line in the Q3 2026 earnings supplement.
• Coach AI's proactive-notification rollout. Osei's team is expected to release the proactive notification feature to the full Chase retail base in Q3 2026. If the preliminary A/B data holds — a doubling of monthly active users without a churn spike — Coach AI crosses 8 million users by year end, at which point it becomes the largest AI financial guidance product at a retail bank by a factor of roughly four. That scale creates negotiating leverage with regulators and a distribution moat that no competitor can close quickly.
• Regulatory classification of AI-generated financial communications. The OCC and CFPB have both indicated, in guidance published in late 2025, that AI-generated communications that influence a consumer's financial decisions may be subject to the same disclosure requirements as regulated financial advice. JPMorgan's legal team filed comments arguing for a functional-equivalence test rather than a bright-line rule. The outcome of that regulatory process will determine whether Coach AI's scope can expand into the territory it is currently approaching cautiously — specifically, personalised retirement planning and debt consolidation recommendations.
• The Wells Fargo and Bank of America response. Both banks have announced AI investment programmes in 2025. Wells Fargo's "Fargo AI" assistant has 12 million users but limited back-office deployment. Bank of America's "Erica" — now eight years old — is being rebuilt on LLM infrastructure for the third time. Neither bank has a proprietary model programme of comparable depth to IndexGPT. Watch for acquisition activity: the most efficient path for either bank to close the gap is a targeted acquisition of an AI infrastructure company rather than an internal build.
• LLM Suite's external productisation. JPMorgan has not announced plans to commercialise LLM Suite externally. Three people inside the bank's technology group, however, describe active internal discussions about offering the platform's orchestration layer — not the models, and not the proprietary data — to corporate treasury clients as a managed service. If that commercialisation happens, JPMorgan's AI programme shifts from a cost-reduction play to a revenue-generating product line, and the valuation implications for the bank's technology assets change substantially.

Frequently Asked

What is JPMorgan's LLM Suite, and who can access it?

LLM Suite is JPMorgan's internal multi-model AI platform, available to employees across all four of the bank's major divisions. It routes queries to external models — OpenAI, Anthropic — or to JPMorgan's proprietary IndexGPT models depending on task type, data classification, and regulatory constraint. As of March 2026, more than 60,000 employees have active LLM Suite access. It is not a client-facing product; Coach AI serves that function for retail and wealth management clients.

How does JPMorgan's proprietary model strategy differ from what Goldman Sachs or Morgan Stanley are doing?

Goldman Sachs and Morgan Stanley have both deployed LLM-based tools — Goldman's GS AI Platform and Morgan Stanley's AI @ Morgan Stanley, the latter built in partnership with OpenAI. Neither bank has publicly disclosed a proprietary model training programme comparable in scale to JPMorgan's IndexGPT family. Goldman routes the majority of its internal AI queries to external APIs. Morgan Stanley's deployment is heavily OpenAI-dependent. JPMorgan's distinction is the investment in training on proprietary financial data, which creates a fine-tuning advantage that external-API-dependent deployments cannot replicate without the same underlying data asset.

What are the regulatory risks around JPMorgan's AI-generated client communications?

The primary regulatory risk is that the OCC or CFPB reclassifies AI-generated financial guidance — including Coach AI's scenario modelling and spending recommendations — as regulated financial advice, triggering disclosure requirements, suitability obligations, and potential licensing constraints. JPMorgan currently positions Coach AI as a financial management tool rather than an advisory service, a distinction that its legal team has argued extensively in regulatory filings. The risk is not existential — the bank has the compliance infrastructure to adapt — but a strict classification ruling would require product redesign that delays the Coach AI rollout by an estimated 12 to 18 months.

How does JPMorgan's $1.9B savings projection break down across divisions?

Based on internal documentation reviewed by INTELAR, the CIB accounts for the largest share — approximately $640M annually — driven by research productivity gains and compliance automation. Consumer and Community Banking contributes $520M, primarily through reduced call-centre volume attributable to Coach AI self-service. AWM accounts for $340M through portfolio reporting automation. The Commercial Bank contributes $190M through credit underwriting and documentation acceleration. The remaining $210M is attributed to enterprise functions: legal, HR, and technology operations. These figures reflect the bank's FY2027 base-case projections, not current-year actuals.

Is JPMorgan's AI deployment reducing headcount?

The bank has not announced AI-driven redundancy programmes. Internal communications reviewed by INTELAR consistently frame the AI deployment in terms of productivity augmentation and redeployment rather than reduction. In practice, the savings are achieved through a combination of not backfilling attrition, slowing junior-analyst hiring in roles most affected by automation, and reallocating headcount from administrative tasks toward client-facing functions. The bank employs approximately 300,000 people globally. The AI programme, at its current trajectory, is expected to slow headcount growth rather than reduce absolute numbers — at least through 2027.

The Dossier Close

JPMorgan's agent stack is not a skunkworks project or a press-release programme. It is a $310M annual operating commitment producing documented savings that the bank's own financial models now treat as structural. The January 2024 slide deck that started this dossier projected $1.9B in savings. Three years into the programme, the trajectory — 60,000 LLM Suite users, $430M in documented annual savings across CIB and AWM alone, 4.2 million Coach AI monthly active users — suggests the base case was the one the bank needed to be more confident in, not more conservative about. The aggressive line is looking more plausible than it did when the deck was written.

The bank's proprietary model strategy is the bet that differentiates it from every financial institution currently writing large API bills to OpenAI. IndexGPT is not a finished product. It is an appreciating asset — one that improves with every internal query, every fine-tuning cycle, every quarter of proprietary financial data added to the training corpus. The bank that owns the model owns the relationship. The bank that rents the model owns nothing but the API bill. JPMorgan has chosen which bank it intends to be, and it chose before most of its competitors decided the question was urgent. That timing advantage is not permanent. But it is substantial, and it is compounding.