AI Client Service in Securities Services

Share this email with a colleague now so they’re on the same page. We only grow through word of mouth and your referrals - Brennan

Table of Contents

• The Challenge

• How AI Addresses This

  • The Technical Approach

  • What is Retrieval-Augmented Generation?

• Real-World Applications

• The Agentic Evolution

• Implementation Considerations

• Measuring Success

• Key Terms

• Sources

Based on available disclosures, every major US custodian has deployed AI internally to help staff handle client queries faster or do other simple tasks under strict governance and controls. Almost none have let the AI respond to clients directly. The adoption barrier is not the technology. It is whether your firm trusts AI to communicate on its behalf. For broader context on AI across custody functions, see The Complete Guide to AI in Global Custody.

The Challenge

The majority of inbound client queries are factual lookups: position balances, settlement status, fee calculations, corporate action deadlines. Each requires an analyst to search across custody, accounting, corporate actions, and settlement systems, compile the answer, and respond. The work is high-volume, manual, and largely identical from one query to the next.

The problem is getting worse. Assets under custody are growing faster than firms can add headcount. BNY manages $59.3 trillion in assets under custody and administration, up 14% year-on-year (Q4 2025 earnings). Citi's securities services assets grew 24% to $31 trillion (Q4 2025 earnings). Fee compression limits the business case for proportional staffing increases. Post-T+1, settlement-related queries are now time-critical in ways they were not under T+2.

The client-side conversation has barely started. A Coalition Greenwich survey of 121 institutional investors found that two-thirds of UK institutions have not discussed AI with their asset managers or service providers. The gap between what is technically possible and what clients are prepared to accept remains wide. Regulators are now requiring the very transparency that research shows erodes client trust, a bind explored further below.

How AI Addresses This

The Technical Approach

AI client service combines several capabilities. Natural language processing classifies incoming queries by intent and extracts key entities (account numbers, security identifiers, date ranges). According to John Snow Labs benchmarks, NLP achieves 92% accuracy in extracting key information from financial documents, compared with 74% for conventional rule-based methods. Retrieval systems then pull the relevant data from structured sources: positions, transactions, corporate actions, settlement instructions. A generation layer drafts the response.

Citi's approach illustrates the pattern. Its NLP models are trained on SWIFT MT599 free-format messages to classify client inquiries and generate draft responses for agent review (Citi press release, 2025). The sequence is: classify the query, retrieve the data, draft the response.

What is Retrieval-Augmented Generation?

Retrieval-augmented generation (RAG) is an architecture that makes this feasible for client communications. Rather than generating answers from training data (which risks fabrication), RAG retrieves factual data from authoritative systems before generating a response. Every answer is grounded in the system of record.

This matters because unconstrained large language models are not accurate enough for consequential monetary-impact type client communications without extensive testing and guardrails. Academic benchmarks show LLMs struggle beyond 50% accuracy on complex real-world financial tasks (Finance Agent Benchmark, arXiv). RAG does not eliminate errors, but it constrains the AI to verified data rather than statistical prediction. Broadridge's OpsGPT uses RAG for settlement fails research, trained on data from systems handling $10 trillion in daily trades (Broadridge press release, 2025). The same architectural pattern applies to client service.

Real-World Applications

Custodian disclosures reveal a consistent pattern: AI is deployed to help staff, not to replace them in client interactions. The data retrieval patterns mirror those used in AI Reconciliation in Custody Operations, but applied to client queries rather than break resolution.

Citi operates the most visible example. CitiService Agent Assist provides real-time guidance to customer service agents during client interactions across 47 countries. It surfaces procedural information, generates call transcripts, and creates after-call summaries, reducing average handle time and improving first-contact resolution (Citi press release, 2025). The client speaks to a human. The AI supports the human behind the scenes. This is Level 2 on the autonomy spectrum.

BNY Mellon has deployed 125+ AI use cases and over 130 "Digital Employees" through its Eliza 2.0 platform (Q4 2025 earnings). None are client-facing. The firm has explored "Client Co-pilots" but has not deployed them, with industry commentary citing data privacy and multi-tenant security concerns as blockers. 125 use cases, and none face the client directly.

JPMorgan is the closest to a client-facing deployment. Its Payments Virtual Assistant, launched in November 2025, supports natural language queries for Commerce Center reporting (J.P. Morgan Developer Portal). It is the only confirmed live client-facing AI tool among the major custody banks, and it is in payments reporting, not securities services.

Based on available disclosures, no major custodian has deployed AI that communicates directly with securities services clients at scale.

The Agentic Evolution

The autonomy spectrum helps explain where client service AI stands and where the resistance lies. For the broader shift from automation to autonomy across post-trade, see Agentic AI in Post-Trade: From Automation to Autonomy.

Based on available disclosures, most custodians operate at Level 1-2 for client service. Citi's Agent Assist is Level 2: the AI drafts, the human reviews and responds. The critical shift to Level 3 is where AI handles routine queries directly, with humans monitoring outcomes and managing escalations. An agent at this level might autonomously confirm a settled cash balance but instantly route any query containing "dispute" or "missing" to a human supervisor.

The jump from Level 2 to Level 3 is where the trust barrier bites hardest. Internal AI tools carry low reputational risk. An analyst who receives a bad AI suggestion simply ignores it. Client-facing AI carries the risk of embarrassing the firm in front of its most important stakeholders. The technology for AI client service exists today. Whether your firm will trust it to speak to clients on its behalf is a different problem entirely.

Implementation Considerations

The technology described above exists. So why is most client service still handled manually?

The 5C Adoption Friction Model identifies where resistance emerges. For client service, the dominant friction point is Credibility, not Capability.

Credibility (trust in AI outputs) is the primary barrier. Client service is the firm's public face. A wrong answer to a client is not just an operational error; it is a relationship risk. Experimental research demonstrates the problem: in an experimental setting with general participants, when subjects learned that AI was involved in generating financial content, trust scores dropped from 5.08 to 4.18 on a seven-point scale (Schilke and Reimann, Organizational Behavior and Human Decision Processes, 2025). A separate Brunswick Group survey of 100 institutional US equity investors found that only four in ten trust AI summaries as much as analyst-written content (Brunswick Group 2026 Investor Survey). Teams will manually revalidate every AI-drafted response until they trust it, which defeats the efficiency gain.

Clarity (understanding what the AI actually does): "The AI looked it up" is not sufficient for institutional clients managing billions in assets. Control (who is accountable when it goes wrong): if the AI gives a client incorrect position data, accountability frameworks are largely undefined. Capability (skills to work alongside AI): teams need to interpret confidence scores and manage hybrid workflows of AI-handled routine queries alongside human-handled complex ones.

Consequences (acceptable risk exposure): ESMA guidance (May 2024) expects firms to disclose AI use in client interactions. FINRA Rule 2210 applies the same communications standards to AI-generated content as human-generated. Disclosure itself reduces trust, creating a compliance-credibility bind.

Measuring Success

Client service AI should be measured on outcomes, not deployment.

Query resolution time is the headline metric: reducing routine queries from hours to minutes. First-contact resolution rate measures whether queries are resolved without escalation. Client satisfaction scores (NPS or equivalent) capture whether faster responses actually improve the relationship. Cost per query should track headcount reallocation, not headcount reduction; the goal is freeing experienced staff for complex work and time-critical T+1 exceptions. Error rate comparison between AI-assisted and manual processes is the credibility metric: until AI-assisted error rates are demonstrably lower, teams will not trust the system.

The firms that solve the trust problem first will have a structural advantage in client retention after assets under custody at the two largest firms grew 14-24% year-on-year in 2025. Client service headcount cannot keep pace.

The disclosure paradox makes this harder than it sounds. Regulators are increasingly expecting firms to disclose when clients interact with AI. But research shows that disclosure itself reduces trust. Firms must build AI that is trustworthy enough to survive the transparency it is legally required to provide.

Is your firm at Level 1, where AI helps analysts research answers? Or is it ready for Level 3, where agents handle routine queries directly and humans focus on the work that actually requires judgment?

I write about AI adoption for operations leaders at Get AI To Work. If you have read this far, the technology is not your problem. Subscribe for the part that is.

Key Terms

Agentic AI: AI systems that can reason, plan, use tools, and take actions with minimal human oversight. Unlike traditional AI that makes predictions or generative AI that creates content, agentic AI can complete multi-step tasks autonomously within defined guardrails.

AI Agent: A software system that uses AI to perceive its environment, make decisions, and take actions to achieve goals. In client service, an agent might receive a query, investigate across multiple systems, draft a response, and escalate only genuinely complex issues without human intervention.

Human-in-the-Loop: An AI design pattern where humans review and approve AI decisions before execution. Common in current client service AI, where AI drafts responses but humans review and send them. The dominant pattern at Level 1-2.

Guardrails: Constraints and boundaries that limit what an AI agent can do. In client service, guardrails might include approved response templates, mandatory escalation triggers for certain query types, or value thresholds requiring human review.

Tool Use: The ability of an AI agent to interact with external systems, APIs, and data sources. Enables client service agents to query custody platforms, access position data, retrieve corporate actions information, or check settlement status as part of answering client queries.

Autonomy Level: The degree of independence an AI system has to make decisions and take actions. Ranges from fully human-controlled (Level 0) to fully autonomous (Level 4). Most custody client service AI currently operates at Level 1-2.

Natural Language Processing (NLP): AI techniques that enable computers to interpret and extract meaning from human language. In client service, NLP classifies query intent, extracts key entities (account numbers, securities, dates), and enables systems to understand unstructured client requests.

Retrieval-Augmented Generation (RAG): An AI architecture where the system retrieves factual data from authoritative sources before generating a response. Reduces hallucination risk by grounding answers in verified data. Critical for client communications where every answer must be factually correct.

Straight-Through Processing (STP): Automated processing of transactions or queries from start to finish without manual intervention. In client service, STP would mean a query received, processed, and answered without human involvement.

Assets Under Custody/Administration (AuC/A): The total market value of financial assets held or administered by a custodian on behalf of clients. A key scale metric; as AuC/A grows, client service query volumes grow proportionally.

Sources

Custodian Disclosures

1. BNY Mellon Q4 2025 Earnings (January 2026): Eliza 2.0 platform, 125+ AI use cases, 130+ digital employees, $59.3T AuC/A.

2. Citi Q4 2025 Earnings (January 2026): Securities services record revenue, 70%+ AI adoption, $31T AuC/AUA.

3. Citi press release (2025): CitiService Agent Assist deployment in 47 countries, SWIFT MT599 NLP, Pega Innovation Award.

4. JPMorgan Q4 2025 Earnings (January 2026): 1,000+ AI use cases targeted, LLM Suite for 50,000 employees.

5. J.P. Morgan Developer Portal (November 2025): Payments Virtual Assistant launch.

Vendor Disclosures 6. Broadridge press releases (January 2024, May 2025): OpsGPT capabilities, RAG architecture, $10T daily trade data.

Industry and Academic Research 7. Coalition Greenwich: Institutional Investing in the AI Era (121 institutional investors surveyed across Asia/Europe). 8. Brunswick Group: 2026 Investor Survey (100 institutional US active equity investors, November-December 2025). 9. Schilke, O. and Reimann, M.: "The Transparency Dilemma: How AI Disclosure Erodes Trust", Organizational Behavior and Human Decision Processes, Vol. 188, May 2025. 10. John Snow Labs benchmarks: NLP accuracy for financial document extraction (vendor research). 11. Finance Agent Benchmark (arXiv): LLM accuracy on real-world financial tasks.

Regulatory 12. ESMA: Public Statement on AI and Investment Services (May 2024). 13. FINRA: Regulatory Notice 24-09 (July 2024).

Article prepared February 2026. Company-reported figures have not been independently verified.