Production-ready AI vs. experimental: What asset servicing actually needs
The conversation about artificial intelligence in asset servicing has reached a curious stage. Everyone discusses it. Most recognise its potential. Yet the gap between what’s being demonstrated in controlled environments and what functions in production operations remains substantial.
This matters because asset servicers face genuine pressure to modernise operations whilst preserving the accuracy and reliability their clients demand. Understanding the difference between AI that impresses in presentations and AI that performs consistently in live operations isn’t academic. It determines whether technology investment delivers business value or creates new problems to manage.
The experimental vs. production distinction
Much of what gets labelled as AI in financial services exists in a state between proof-of-concept and full production deployment.1 This technology works under specific conditions, with curated data sets, and with significant human oversight. This represents important progress, but it’s not the same as production-ready capability.
Production-ready AI exhibits specific characteristics. It handles real-world data quality issues without failing or requiring constant intervention. It operates reliably across the full range of scenarios it encounters, not just the common cases. It provides transparency about what it’s doing and why, enabling appropriate oversight. This requires constrained AI agents designed for collaboration rather than autonomous operation, agents that can say: ‘I’ve found an anomaly; here are three ways to resolve it; I recommend option B for these reasons.’ It integrates with existing operational workflows without requiring those workflows to be redesigned around its limitations.
The distinction isn’t whether the underlying algorithms are sophisticated. It’s whether the complete solution functions reliably in the messy reality of production operations.
What matters in asset servicing operations
Asset servicing presents specific requirements that shape what constitutes useful AI capability. Without understanding these requirements, organisations risk investing in impressive technology that fails to address actual operational needs.
Data reality. Asset servicing operations work with data of varying quality and structure. Counterparty data arrives in inconsistent formats. Client-provided information contains errors. Reference data requires constant updating. Production-ready AI must handle this reality rather than assuming clean, consistent inputs.2 Indeed, experience suggests that 80% of any AI project is fundamentally a data project; there is no viable AI strategy without an equally robust data strategy.
Accuracy requirements. Financial operations demand extremely high accuracy. An AI system that’s right 95% of the time might be impressive by general standards but creates serious problems in asset servicing and cannot be allowed to produce hallucinated outputs. The cost of errors, regulatory consequences, client impact, operational disruption, means accuracy thresholds must be substantially higher than in many other applications.
Auditability and explainability. Operations teams need to understand what AI systems are doing and why.3 When an AI system performs a task or produces a result, the logic behind it must be traceable. This isn’t only regulatory necessity; it’s operational practicality. Without explainability, teams can’t validate outputs, investigate discrepancies, or maintain confidence in the system.
Integration with human workflows. AI in asset servicing will reshape how people work, handling routine elements whilst escalating exceptions and complex cases. It won’t replace humans entirely; it will augment their work.4 This requires constrained AI agents designed for collaboration rather than autonomous operation. The handoffs between AI and human decision-making need to function smoothly.
Regulatory compliance. Financial services operations exist within comprehensive regulatory frameworks. AI systems must support compliance requirements rather than creating new risks. This influences everything from data handling to decision documentation to system governance.
The three components of practical AI
Understanding where AI delivers genuine production value in asset servicing requires moving beyond abstract maturity models. Three distinct components define how practical AI capability is deployed today.
Research. AI accelerates the gathering, synthesis and interpretation of information that operations and investment teams depend on. Rather than replacing analytical judgement, it handles the time-intensive work of processing large volumes of structured and unstructured data e.g. market data, regulatory updates, counterparty information, surfacing relevant insights for human review. Critically, the value of research agents multiplies when their outputs are contextualised against actual portfolio holdings. A regulatory change or credit event is abstract information until it is mapped to specific fund exposures, concentration risks and client positions. Research agents operating within a portfolio-aware data environment can make that connection automatically: identifying which funds hold affected instruments, quantifying potential impact and flagging exposures that warrant immediate attention. This is where constrained AI agents operate most confidently today, augmenting expertise rather than substituting for it.
Systematic oversight. AI acts as a systematic check across processes where errors carry the highest consequences. It monitors outputs, flags anomalies and presents options for resolution. Humans retain decision authority. The AI ensures nothing falls through the gaps. In practice, this means continuous validation running across every fund and every position simultaneously, at a frequency and consistency that manual checking cannot sustain.
Automation. For well-defined, rules-based processes with clear parameters, AI can handle end-to-end execution with minimal human intervention. This is where efficiency gains are most tangible when the conditions are right: well-defined processes with clear parameters and sufficient data quality to support end-to-end execution.
The challenge isn’t pursuing maximum automation. It’s deploying AI that reliably augments human capability and building the trust and governance infrastructure that allows automation to expand responsibly.
Evaluating AI capability
When asset servicers evaluate AI solutions, several factors distinguish production-ready capability from experiments.
Operational track record. The right question is not whether a vendor has years of production history, few do, but whether their approach to deployment is sufficiently rigorous to trust. Does the vendor follow a structured progression from proof-of-concept to controlled pilot to full production? Are handoffs between AI and human review well-defined at each stage? Can the solution be stress-tested against your actual data and edge cases before go-live?
Error handling and edge cases. How does the system behave when it encounters situations outside its training? Can it recognise when confidence is low and escalate appropriately? Does it fail safely or produce incorrect results? The behaviour at the edges matters as much as performance on common cases.
Integration architecture. How does the AI connect with your existing systems and workflows? Does it require extensive custom development? Can it operate with your current data structures? Integration complexity often determines whether impressive capability becomes practical value.5
Transparency and monitoring. Can you observe what the AI is doing? Are actions and results traceable? Can performance be monitored over time? Production AI requires ongoing oversight, which demands visibility into system behaviour.
Human interaction model. How do operational teams interact with the AI? Is the interface intuitive for people without data science expertise? Can business users understand and trust the outputs? The most sophisticated AI delivers limited value if operational teams can’t work with it effectively.6
Governance framework. What controls exist around the AI’s operation? How are changes to models and rules managed? What approval processes govern its activity? Production deployment requires formal governance that experimental systems can bypass.
The implementation question
The critical question facing asset servicers isn’t whether AI has potential. It’s whether specific AI capabilities are ready for production deployment in their specific operational contexts.
This requires moving beyond general enthusiasm about AI’s possibilities to disciplined evaluation of solutions against specific requirements. What problem does this AI solve? How does it integrate with current operations? What accuracy levels does it achieve consistently? How do operational teams interact with it? What happens when it encounters exceptions?
The servicers who will extract genuine value from AI won’t necessarily be the earliest adopters of experimental technology. They’ll be the ones who identify production-ready capabilities that address real operational challenges and deploy them effectively.7
Technology sophistication matters less than operational impact.8 The most advanced AI delivers little value if it can’t function reliably in production. Conversely, relatively straightforward AI that augments human capability consistently and transparently can transform operational economics.
The next stage of AI deployment in asset servicing requires this level of discrimination: understanding what production-ready means, recognising it when you see it, and implementing it where it delivers genuine operational value.
This article is part of our asset servicing insight series
Citations
1 0% report “extensive” AI use; 62% are running pilots/POCs; integration remains “moderate” (38%) or “emerging” (31%). Deloitte 2025 AS Survey, p.25–26
2 83% cite interoperability as their top data challenge; only 25% feel confident their tech addresses data governance. Deloitte 2025 AS Survey, p.17–18
3 Regulatory/legal hurdles are now the #1 AI concern (62%, up from 0% in 2023). EU AI Act will require documentation and risk classification of AI in financial services. Deloitte 2025 AS Survey, p.29–30; EY Top 5 Trends, p.6
4 Top use cases are augmentation-focused: automated data processing (92%), compliance (69%), client servicing (62%). EY recommends AI to “enhance and not replace human oversight.” Deloitte 2025 AS Survey, p.27–28; EY Top 5 Trends, p.7
5 92% say legacy systems constrain innovation. Legacy tech is the #1 F2B delivery challenge. 79% of Luxembourg firms cite manual workflows as their key operational hurdle. Deloitte 2025 AS Survey, p.13–14; EY Benchmarking 2025, p.6
6 64% report a skills gap in emerging tech. Lack of digital talent identified as a key setback for digitalisation progress. Deloitte 2025 AS Survey, p.27–28; EY Benchmarking 2025, p.14
7 >78% expect AI to enhance efficiency and reduce costs, but only 33% fully fund digital initiatives (down from 52% in 2024) — a widening aspiration-execution gap. Deloitte 2025 AS Survey, p.29–30; EY Benchmarking 2025, p.8
8 Luxembourg digital maturity unchanged at 2.3/5. Only 21% report most processes digitalised (down from 41% in 2024) despite continued investment. EY Benchmarking 2025, p.7, 10
