Asset managers have spent the last two years exploring what AI can do. However, there is a more important challenge that many firms have neglected: building the operational infrastructure required to make AI work reliably, at scale, and at a sustainable unit cost.
Across the industry, we see how asset managers are moving beyond isolated proofs of concept and beginning to embed AI into investment operations, compliance workflows and decision support systems. Expectations are rising quickly. Teams are under pressure to deliver faster insights, improve operational efficiency and respond more effectively to market volatility and regulatory scrutiny.
Yet, projects often stall when teams encounter fragmented workflows, disconnected systems and unstructured operational data that was never designed to support intelligent automation. In fact, Gartner reported that up to 30% of AI projects are abandoned within a year. Evidently, operating-model misalignment is a significant hurdle; models are only as effective as the infrastructure surrounding them, and many organisations lack the operational resilience required to get the most out of AI.
The hidden bottleneck: infrastructure, not intelligence
Most conversations around AI still focus on insights and outputs rather than the operational processes that sit underneath them.
In practice, operational teams face a different challenge entirely: work still arrives as PDFs, emails, spreadsheets, and inconsistent data feeds. Before AI can generate insights or take action, information must be interpreted, validated, and routed, which often requires employees to manually re-enter data, reconcile records, or validate unstructured content. These manual processes create friction and inefficiencies, which slow operations and create risk, particularly in high-volume environments where small delays or inconsistencies can quickly escalate.
Models are only as effective as the operational environments surrounding them, and as AI adoption accelerates, these underlying weaknesses only become more exposed.
Defining true AI-ready infrastructure
Through my work with asset managers, I’ve come to see AI-ready infrastructure resting on four interdependent pillars.
The first pillar hinges on governing how agents connect to data across the enterprise. The Model Context Protocol (MCP) addresses this by providing a standardised way for intelligent agents to access both structured and unstructured data, tools, and APIs. MCP has quickly become the industry standard for secure, reliable connectivity, replacing brittle, bespoke integrations with a governed, auditable layer. Major hyperscalers have already adopted it, and the ecosystem now includes over 16,000 production servers. For asset managers, MCP is the mechanism that transforms fragmented data estates into a substrate agents can safely reason over.
The next priority is ensuring Cloud-native scalability, which has become increasingly agent-native. Investment operations workloads such as batch cycles, market events, corporate actions, and client onboarding peaks are often unpredictable in frequency and volume. Legacy stacks were never built for this profile. Infrastructure must scale elastically, which in practice means being cloud-native by design. This includes dedicated agent runtimes that provide multi-tenant isolation, persistent memory, native observability, and built-in support for MCP.
Another essential element is embedding financial operations (FinOps), into AI infrastructure. AI relies on infrastructure with costs that can change quickly. Expenses from things like token usage, GPU processing, and moving data can turn a promising pilot into an unmanageable budget item. In fact, 72% of IT and finance leaders now describe GenAI spending as unmanageable, and 98% of FinOps professionals now manage AI spending, (up from just 31% two years ago). Without clear tracking, model selection policies, and budget guardrails, costs can quickly spiral. FinOps has become more than a back-office function, it is now also a core requirement of designing AI infrastructure.
The last and perhaps most critical pillar is building a broad, governed model ecosystem. No single AI model excels at every use case. Tasks such as reasoning, extraction, classification, summarisation, code generation, and reconciliation each require distinct capabilities. Therefore, AI-ready platforms must be model-agnostic by design: able to intelligently route tasks across both frontier and specialised models, apply policy guardrails, capture evidence, and ensure human oversight for high-stakes decisions. This diversity is essential for making AI outputs trustworthy across regulated workflows.
Embedding intelligence into workflows
When these four pillars are in place, the operational dynamic changes. AI interprets unstructured content as work enters the organisation, extracting relevant information, validating it against business rules and directing exceptions to the right teams before delays compound downstream.
This changes the operational dynamic significantly. Teams spend less time manually processing information and more time handling exceptions, oversight and judgement-based decisions. Operational bottlenecks become easier to identify earlier in the workflow. Auditability improves because actions and decisions are captured systematically rather than through fragmented manual processes.
From AI experimentation to operational resilience
Operational maturity requires infrastructure that supports AI on an ongoing basis, not just as an experiment. This means investing in reliable data pipelines, governed agent-to-tool connectivity, elastic agent-native runtimes, a curated multi-model ecosystem, and a FinOps discipline to ensure every automated decision remains economically viable.
Equally important is a shift in mindset around enterprise technology. While depth of functionality has traditionally been the main consideration, usability and workflow coordination are now just as critical. Even the most powerful systems can underperform if users cannot access information quickly or navigate workflows efficiently.
For mid-sized asset managers in particular, this change creates both pressure and opportunity. Large firms continue to invest heavily in AI transformation programmes, while smaller firms often benefit from greater agility and fewer legacy constraints. Mid-sized firms sit between those two realities. Their competitive advantage increasingly depends on how effectively they modernise operational infrastructure without creating additional fragmentation.
In this environment, firms must prioritise interoperability, workflow continuity, and real-time operational visibility, while reducing manual intervention and maintaining strong oversight. It also means recognising that operational resilience is closely tied to the quality of the user experience inside complex enterprise environments.
AI will continue reshaping asset management operations over the coming years. Organisations should focus on creating operational environments that turn intelligence into reliable execution. Many may be distracted by chasing the smartest model, but I have seen that it is more worthwhile to spend your time and resources building infrastructure that enables intelligence to run continuously, safely, and profitably.
