Summary
Model Context Protocol (MCP) is emerging as a critical technology for private market firms seeking to operationalize AI. By connecting large language models to investment systems, data platforms, and workflows, MCP enables agentic AI to automate tasks, improve data integration, and accelerate decision-making across private market operations.
Financial services CTOs have a recurring AI dream: enabling department leaders to ask plain-English questions about their portfolio data. This dream is now within reach, thanks to the rapid advance of agentic AI and Model Context Protocol (MCP) servers.
MCP is both a protocol for searching data and a mechanism of agentic action. According to a January 2026 survey by Stacklok, a security platform built for MCP environments, only 3% of financial services organizations were using MCPs on a broad production basis. It wouldn’t be surprising if, in the seven months since that survey, the number hits double digits, as 31% were using MCPs in limited production and 29% were piloting.i Our clients are asking increasingly about whether they need MCPs to optimize their private market workflows. Here are some reasons why we are championing MCP servers for private markets as essential components in making firms’ AI dreams come true.
Even before AI entered the picture, private credit managers struggled to integrate datasets from CRM, IBOR/ABOR accounting systems, Excel, and portfolio management systems. Each system has different application programming interfaces (APIs), data models, and permissions. Further, when it comes to private credit structures, data arrives in varying formats from originators, servicers, and lending platforms, making normalization and AI integration harder still. Now, throw AI agents and LLMs into the mix.
AI models don’t easily communicate or collaborate across firms’ data infrastructure. That interoperability gap stops AI from scaling. MCP creates an abstraction layer that connects disparate datasets, so agent-to-agent protocols can communicate directly. From there, analysts query the LLM once, through a single prompt, rather than logging into each system separately. MCP is a direct, high-speed express lane on a highway of autonomous agent vehicles.
“In LLM-based multi-agent systems, communication is not just the exchange of information — it is the medium through which collective reasoning emerges... Thus, communication in LLM-MAS presents a fundamental tension. It is simultaneously the key to emergent intelligence and a critical vulnerability that, if poorly designed, can undermine the entire system. Designing resilient, semantically meaningful communication architectures is therefore not optional — it is central to the success, trustworthiness, and safety of next-generation agentic AI.” — Survey of LLM Agent Communication with MCP: A Software Design Pattern Centric Reviewii
Traditional APIs require developers to write static code. With MCP, you don’t need to write code to discover tools, you just prompt it.
Let’s say a manager is looking at the aforementioned private credit infrastructure deal for a securitized pool of residential solar loans. They prompt the LLM: “What are the current deals that I have in the infrastructure sector and how have they been performing?” Without an MCP, the manager will need to log into five systems, export files, upload them to the AI, and repeat the process every time. The analyst could program APIs to execute these exports to pull, for example, NAV data, positions, and loan performance data. However, each API will need to be customized to integrate with the respective systems.
APIs are indispensable system connectors, but if a firm has 20 systems and an AI assistant, someone has to build and maintain 20 separate integrations. But an MCP is built for AI models to dynamically discover and consume tools in real time. MCPs have a more native communication style for LLMs. Permissions are enforced natively, error rates drop, and the model doesn’t require you to know which API to call or what parameters to pass.
Without MCP, the AI agent says:
“I found a new loan tape. You should create a schema and data pipeline.”
With MCP, the AI agent says:
“I found a new loan tape, created the schema, configured the pipeline, added validation rules, and scheduled monthly ingestion.”
MCP is a way for firms to use their own preferred enterprise LLMs such as Claude or ChatGPT, rather than learning new, potentially clunky application-specific chatbots. By using an enterprise LLM via MCP, firms can use their existing infrastructure, including Slack, Outlook, CRM data, and accounting data, to provide business context to the AI. You can instruct it to build a data model, configure a data pipeline, or write data quality rules. Assign it a task, and it executes autonomously.
For instance, you’ve inked a new deal, perhaps lent money to company X. That deal means that your system has to ingest and normalize five or six new files, which includes setting up reference data, a new pipeline based off those file types, and appropriate data quality rules. With an MCP server, you could just drop those files into your LLM and instruct it to normalize into an existing or new data model, run data quality rules checks on a recurring cadence, and create new queries, analytics reports, and data schemas.
The same Stacklok survey found that financial services tech leaders cited data quality and availability as the number two obstacle (behind number one: security) blocking or slowing MCP adoption.iii By the same token, multiple survey reports call out data quality as a top or second-leading blocker to AI adoption in general, including Deloitte’s findings that 81% of respondents citing data quality as a top challenge.iv Any hesitation investment management firms had about modernizing their data platforms in 2022 evaporated by 2026. AI made the business case that compliance and operations never could.
Just as MCP speaks the language of AI natively, the best data architecture for private markets speaks the language of investment management — deal structures, waterfall logic, credit ratings — rather than requiring a translation layer between the data team and the investment team. Poor data quality processed by an AI can produce reporting errors to regulators or investors, resulting in lawsuits, liquidity strain, or covenant breaches.
Firms’ data scientists and CTOs are recognizing that MCPs will be key to the agentic phase of AI. We champion MCP servers for their open architecture, native communication protocols, and agent-to-agent scale. The firms leading this transition are already using agents to manage dry powder, forecast cash flow, and distribute capital calls. The window for building that infrastructure before the volume arrives is narrowing.
Bernardo Cabada
As Vice President, Sales Operations & Enablement at Arcesium, Bernardo leads in showcasing the company's cutting-edge capabilities through executive-level engagements, delivering compelling presentations, technical demonstrations, and proof-of-concepts. His proficiency extends across crucial areas of middle- and back-office investment operations, with a particular emphasis on data governance and enterprise data management for investment managers and asset owners.
Sources:
[i] Stacklok, January 2026. https://stacklok.com/wp-content/uploads/2026/01/State-of-MCP-in-Financial-Services-2026_FINAL.pdf
[ii] Sarkar, Anjana & Sarkar, Soumyendu, 2025. https://arxiv.org/pdf/2506.05364
[iii] Stacklok, 2026. https://stacklok.com/wp-content/uploads/2026/01/State-of-MCP-in-Financial-Services-2026_FINAL.pdf
[iv] Deloitte, October 30, 2025. https://www.deloitte.com/us/en/insights/industry/financial-services/financial-services-industry-outlooks/banking-industry-outlook.html
No spam. Just the latest releases and tips, interesting articles, and exclusive interviews in your inbox every week.
