To MCP or Not to MCP: That Is the Question
Ghosts and apparitions haunt every data leader. Some appear in noisy, inconsistent data, while others manifest in technologies that promise clarity but often deliver confusion. Especially with AI, buzz cycles heat up fast and seem to promise more than they can deliver.
It can feel at times like being a Hamlet figure, torn between choices, uncertain what’s real and what’s an illusion.
Model Context Protocol (MCP) presents another such choice. It has attracted a lot of buzz since Anthropic launched it in November 2024. But a clear view of whether to MCP or not to MCP takes some of the theatrics out of the decision. Approaching it with focus also spares firms from the tragedy of missed expectations.
MCP: teaching AI to speak data
MCP servers provide a simple and elegant solution to a potentially challenging problem in AI. While large language models (LLMs) excel at understanding natural language, they work by predicting sequences of words. They can struggle with structured data in enterprise applications. Because it is harder to predict the next value in a database column, they are less naturally suited for querying structured databases.
MCP helps bridge that gap by letting AI “speak” to enterprise data securely and directly. Behind the scenes, it orchestrates back-and-forth access to an application. End users can engage with the underlying data in a conversational interface. MCP also streamlines the messy integration work between AI models and databases that currently fragments AI adoption.
But even though MCP simplifies the process of closing this gap, it does not emerge as a magic fix to every data challenge. What it does is simpler. It provides a lightweight layer that helps LLM applications, such as chatbots and AI agents, parse and relate data. They can use tools or call functions because they pull from a catalog that registers which inputs an application expects and which outputs it can return.
What the data doesn’t say
In investment management, data always has two layers: what’s visible in systems, and what’s implied by the data’s meaning, lineage, and context. For example, a PMS might indicate a holding of 100,000 shares of XYZ Corp. But that doesn’t tell you what the position means or what choice it reflects. Was it a discretionary trade or done to track indexing or concentration choices? For hedging or for alpha? Was it booked intraday or at close? Was it adjusted after a corporate action? Those answers live in the surrounding context. And without that context, even clean data looks like a lingering trace of something that happened in the past.
Firms also must consider the perennial reality check of their data architecture. Having data by function is a common scenario, but it also creates silos and adds complexity. Each layer imposes distinct constraints, including structure, timeliness, and auditability. For example, data can straddle organizational boundaries.
From dialogue to discipline
Fragmentation constrains how AI can interact with your data. If you want a complete picture of anything from intraday risk to quarterly P&L, the value of bringing together multiple data points into a conversational response to a chat query depends on which data layers you are touching.
The challenge goes beyond truisms about data readiness and extends deep into firms’ approach to data governance. For data about assets and transactions, definitional consistency is essential. Firms need:
- Common identifiers and taxonomies across portfolios and entities
- Semantic alignment between risk, performance, and accounting data
- Lineage and governance that explain every change made
- Access and entitlement controls that match users’ fiduciary responsibilities
Without those disciplines, people can’t be certain what MCP-enabled outputs mean or whether they are valid. They also set the stakes for when to MCP and when not to.
The stakes may not be as tragic as Hamlet’s existential angst of whether to be or not to be. But getting the choice between whether to MCP or not to MCP wrong can have unintended consequences. Luckily, teams building investment technology can rely on some valuable rules of thumb to answer the question.
When to explore
In general, using MCP makes sense for lightweight, distributed access to unstructured data, such as knowledge systems, documents, and communications. Imagine a repository of deal documents or trustee notices for a pool of loans, for example. With the right MCP server, AI can enable users to explore and consider context without re-platforming.
When you’re connecting to unstructured or semi-structured sources, MCP servers can accelerate development. Secure, temporary access to data can take place without deep system integration. The use cases tend to be ad-hoc data or context retrieval, e.g., ontology search, cross-referencing, or workflow enrichment. However, teams should monitor context window consumption as connected MCP servers scale — loading numerous tool definitions and passing large intermediate results can quickly increase latency and costs,i particularly when orchestrating across dozens of data providers in capital markets workflows.
For example, asking an AI to “list differences in all OTC confirmations from Q3 2024” would benefit from an MCP server that lets users search context (which asset classes, where is data located) and retrieve trade confirmation data (document repository or email attachments). MCP's value amplifies when workflows require coordination across multiple data sources, rather than querying a single governed system. When done well, it saves operations teams tangible effort and represents the “right” way to leverage MCP, focusing on narrow data points instead of massive transformation projects.
When to integrate
However, curated data integration can be a better approach for centralized, modeled, governed data, optimized for consistency, traceability, regulatory assurance, and analysis response time. For example, books and records, i.e., historical datasets that contain the official record of fiduciary, financial, or risk-related activities, would benefit from centralized data services.
When you need regulatory-grade inputs and outputs that you can reconcile and audit, or when you need mathematical precision across systems, platform-led, tightly integrated systems are critical for business.
Similarly, if you have legacy applications without modern endpoints, MCP is not designed to handle consuming their data or functionality without additional complex implementation. For example, responding to a regulatory inquiry about fee discrepancies that require data from your legacy fund structure would push MCP against pragmatically unsolvable issues such as incomplete data models, undocumented databases, or obsolete file formats, which could result in incomplete or misleading regulatory responses that create compliance risk.
Not to be, or not only
Unlike Hamlet, you have the luxury of taking a hybrid approach without feeling haunted. Understanding MCP’s value depends on how you plan to use it. The goal of today’s data leader is to balance semantic control and conversational access.
While MCP will likely become a standard expectation for vendors and platforms, the future lies in using MCP’s flexibility alongside robust platform-level governance.
Rather than bypassing the walls that naturally build up in investment data over time, MCP makes the most sense as part of the ongoing discipline of governing and using data from front to back office and beyond.
Authored By
Dmitry (Mitya) Miller
Dmitry (Mitya) Miller is a Senior Vice President and General Manager of Aquata, Arcesium’s comprehensive self-service data platform purpose built for the investment management industry. Mitya is responsible for overseeing all aspects of the Aquata business, including P&L ownership, customer base growth, customer delivery and engagement, and product roadmap.
Share This post
[i] Anthropic, November 2025. https://www.anthropic.com/engineering/code-execution-with-mcp
Subscribe Today
No spam. Just the latest releases and tips, interesting articles, and exclusive interviews in your inbox every week.
