AI in custody operations works through automation of the three functions most exposed to failure as settlement windows shrink. Reconciliation exception management, settlement instruction validation, and corporate actions data extraction are still predominantly manual at most institutions, and each is a direct source of settlement failures, exception backlogs, and client reporting delays. This article explains where AI has produced the clearest results, what the honest limits are, and how an operations team should sequence adoption.

TL;DR

• Settlement compression, from T+2 to T+1 globally and T+0 in India, has turned acceptable processing delays into same-day failures, making AI in custody operations an urgency question rather than an innovation one.
• The strongest evidence sits in three areas in sequence: reconciliation exception management, settlement failure prediction, and corporate actions data ingestion.
• Most institutions are running AI pilots in isolation. The gap between pilot and production is primarily a data quality and governance problem, not a technology shortage.

Why do custody operations still carry so much manual work?

The structural reason is architecture. Custody technology was built over decades by layering systems designed for specific markets, asset classes, and regulatory regimes. The result is an operational environment where data arrives from dozens of sources in inconsistent formats, position records sit in one system and cash records in another, and reconciliation requires a team manually cross-referencing outputs before the working day ends.

Deloitte’s Global 2025 Asset Servicing Survey, covering firms representing approximately 75% of global assets under administration, found that 92% of asset servicers say their ability to innovate is partially constrained by legacy systems. Interoperability across platforms and silos was cited by 83% as their single biggest data management challenge.

The client pressure compounds the operational one. In 2021, 40% of firms considered real-time data analytics and insights a high priority. By 2025 that figure had reached 94%, and it is now treated as a baseline requirement, not a differentiator. Fee compression adds a third dimension, with 81% of asset servicers reporting that fee pressure from asset managers is being passed directly to them, squeezing the cost side of a business whose volumes are growing.

These pressures together explain why AI in custody operations has moved from exploratory to operational across the securities services industry.

Where has AI in custody operations produced the most evidence?

Three areas stand out, and they are not equally mature.

Reconciliation is where AI in securities custody has the longest track record. Matching internal position and cash records against external custodian, prime broker, and clearing house statements generates large exception volumes and consistent patterns. Supervised machine learning identifies those patterns, learns from historical matching decisions, and reduces the time operations teams spend on routine breaks. SIFMA’s April 2026 analysis confirmed that agentic AI systems are now orchestrating end-to-end post-trade reconciliation, automatically matching trades across fragmented systems, investigating discrepancies, and escalating only true exceptions, with compliance agents handling regulatory reporting alongside.

Settlement instruction management is where AI has produced its most commercially urgent results. According to S&P Global Market Intelligence’s March 2026 announcement, invalid standing settlement instructions contribute to nearly one-third of all settlement failures globally, and up to 40% of those instructions are still processed manually by financial institutions. AI addresses SSI failures in two ways. The first is validating instructions against custodian network data before they cause a fail. The second is predicting which instructions are at risk before the cut-off window closes. Both are central to custody automation in a T+1 environment.

Corporate actions is the area with the clearest gap between the scale of the problem and current automation levels. Announcements arrive in multiple formats, across multiple channels, often in multiple languages. Natural language processing and optical character recognition now extract and classify this data, reducing the volume of manual revalidation that has historically made corporate actions one of the highest-cost functions in asset servicing.

Of the three, AI in securities custody has the most validated track record in reconciliation, where task volume is high, success criteria are clear, and historical data is rich. Settlement prediction and corporate actions ingestion are maturing fast, but reconciliation is where the evidence is deepest.

How does settlement compression change what AI must do?

The shift from T+2 to T+1 converts what were manageable processing delays into same-day failures. A reconciliation break, a mismatched SSI, or an unprocessed corporate action that was a two-day problem under T+2 becomes an end-of-day failure under T+1. The operational model that absorbed T+2 volumes without AI is not designed for what follows.

India made this real before anyone else. It was the first country to implement T+1 rolling settlement, and SEBI subsequently introduced T+0, the world’s fastest settlement cycle, covering 500 securities from early 2025. Foreign portfolio investment into Indian equities now exceeds $1 trillion, which means the operational stakes of a failed settlement are substantial. SEBI’s updated custodian regulations, issued in March 2026, and its May 2026 AI governance circular, addressed to every regulated entity in the Indian securities market including custodians, make clear that governance of AI-assisted decisions is now a regulatory expectation, not a future consideration.

Across APAC, Asia holds the world’s largest gross international investment position, with 41% allocated to US markets, placing the region at the centre of cross-border investment flows. T+1 discussions are active across APAC markets, following North America’s transition in May 2024.

S&P Global’s assessment of the T+1 environment leaves no ambiguity. Manual SSI handling is unsustainable. The same logic applies to manual reconciliation and manual corporate actions processing. There is simply less time to catch and repair a failure after it occurs.

What are the limits of AI in custody operations?

AI in custody operations is not a replacement for operations judgment, and treating it as one is where most pilots fail to scale.

Regulatory and legal complexity is the most frequently cited obstacle. Deloitte’s 2025 survey found that regulatory and legal hurdles were the top challenge for 62% of firms launching AI pilots in asset servicing, up from 0% in 2023. This reflects the pace at which regulators have moved from principles to specific expectations. A firm cannot automate a decision it cannot also explain.

Data quality is the second hard constraint. The same survey found that 54% of firms cited inconsistent or poor-quality data as a major obstacle to scaling AI from pilot to production. AI learns from historical records. When those records are fragmented, inconsistent, or ungoverned, the model inherits those defects.

Three areas of custody remain outside reliable AI automation today. Exception escalation requiring regulatory interpretation stays with experienced operations staff. Valuation of non-standard or illiquid instruments depends on judgment that structured data cannot replace. Counterparty dispute resolution requires relationship context that AI systems do not hold. These functions become more valuable, not less, when the high-volume pattern-rich work around them is handled by AI.

Custody automation performs well where tasks are high-volume, historically documented, and pattern-rich. It performs poorly where tasks require regulatory judgment, novel conditions, or multi-party context that does not exist in structured form. That boundary is the honest one, and knowing it is what separates a sustainable deployment from a stalled pilot.

How should a custody operations team sequence AI adoption?

Most operations teams attempting AI in custody operations start with too broad a scope and stall at the pilot stage. Deloitte’s 2025 survey found that 46% of firms cited scaling from proof of concept to production as a top challenge. The core problem is rarely resources. It is data dependency. Each stage of AI adoption in custody builds on the data quality established by the stage before it.

The Custody Operations Readiness Sequence moves through four stages, each building on the foundation of the one before.

1. Reconciliation exception management. AI matches positions and cash against external records, learns from historical break resolution, and surfaces exceptions by priority. This stage delivers the fastest operational return and produces the clean, governed position data that later stages depend on.
2. Settlement failure prediction. With clean position and SSI history from Stage 1, predictive models score instruction-level fail risk before the cut-off window. Operations teams act on alerts rather than failed settlements. A reactive function becomes proactive.
3. SSI validation and enrichment. AI validates standing settlement instructions against custodian network data, flags stale or mismatched instructions before they reach the settlement window, and enriches records using custodian reference data. The failure history from Stage 2 improves model accuracy over time.
4. Corporate actions data ingestion. NLP and OCR process unstructured announcements, producing structured records that reduce manual revalidation and downstream errors. This stage depends on a clean instrument master, which the data governance work of the earlier stages produces.

The order is not arbitrary. Firms that began at Stage 4, drawn by the visibility of the corporate actions problem, found the underlying data infrastructure was not ready. Firms that started at Stage 1 and moved in sequence built the foundation that made every subsequent stage work faster. Starting with reconciliation is also where the return on operations time is fastest, which makes the investment easier to sustain. An AI custody platform designed for modular deployment makes the sequence tractable without requiring a full system replacement at the outset.

What do AI-enabled custody operations look like in practice?

For institutions in India, the Middle East, and APAC that are entering the custody business or modernising existing operations, urgency and opportunity arrive together.

The urgency is regulatory and operational. SEBI’s framework now expects custodians to govern AI-assisted decisions explicitly, and settlement cycles in India are the world’s fastest. Institutions that absorb regulatory change through configuration rather than manual rework carry a structural advantage as requirements continue to evolve.

The opportunity is architectural. Institutions entering the custody business today are not carrying a three-decade legacy architecture. They can build the governance layer alongside the automation layer, rather than retrofitting one onto the other. That is the difference between custody technology that grows with the business and technology that constrains it.

This is where a composable AI custody platform earns its value. eMACH.ai Custody is built on this premise, connecting settlement, reconciliation, corporate actions, billing, and client reporting in one governed architecture. It includes predictive alerts for settlement failures, AI-assisted corporate actions processing, and smart reconciliation across disparate data sources. Institutions can begin with the operational functions where they need support most and expand module by module, without a full platform replacement.

The honest qualifier applies here too. No platform resolves unclear data governance, and operational efficiency from AI depends on the data quality invested in before the AI is deployed. That sequencing principle holds whether you are entering the custody business for the first time or modernising a business you have run for decades.

Summary

AI in custody operations delivers its most durable value in reconciliation exception management, settlement failure prediction, and corporate actions data extraction, the three functions most exposed to settlement compression. The evidence for each is real, and so are the limits. Data quality determines whether AI in custody operations adds value or amplifies existing gaps. Sequence the adoption in the order the data supports. Start with reconciliation, then settlement prediction, then SSI enrichment, then corporate actions. That order is what distinguishes institutions that have moved from pilot to production from those still refining their approach.

Frequently Asked Questions