In this continuation of our MCP blog series, our CPO Dana Ritter goes beyond the buzz to examine what it really takes to bring MCP to life in financial services. See more product-related content from Dana on his blog.
At Unique AI, we’re obsessed with building products that deliver real value to financial institutions, from helping relationship managers be more efficient to providing hedge funds with better alpha. As a product leader in this space, I’m constantly evaluating new technologies and protocols.
You can’t just dump a list of tools on them and say, "figure it out." The user experience needs to feel seamless and intuitive. As product builders, we must address this cognitive overload head-on. That means thinking beyond just the technology and focusing on the user. We're exploring solutions like:
Intelligent Curation: Designing the system to understand the user's intent and recommend the most relevant tools for a given query, essentially acting as an intelligent co-pilot.
Contextual Assistance: Providing "help" or "suggested action" buttons that only appear when the user is in a particular context (e.g., in an email draft, a button to find the right document).
Guided Workflows: Building pre-packaged workflows that combine multiple tools to solve a specific problem, such as "create a client summary" or "generate a research memo."
"Out of the box" MCP does not always solve the business problem. An example is an MCP Outlook connector. An MCP Outlook tool like "create draft email" works fine. But a prompt relying on a tool that leverages Outlook search capabilities for a query like "summarize the last 5 emails from John Smith" will not work. We need to find an approach for these type of queries where we ingest emails so we can do a RAG approach on the emails to get a good result.
This is non-negotiable in financial services. Connecting an LLM to external services via MCP, especially those that can take action, introduces significant risk. Imagine an LLM with the ability to auto-send an email to a client. What if it gets the details wrong, or sends something that violates a compliance rule? The reputational risk is immense.
Our approach is to build a layered system of both technical and human guardrails. This isn't just about a "do not send" button. We need to implement:
Granular Permissions: Allowing administrators to control which tools are available to which user groups.
Human-in-the-Loop: For sensitive actions, requiring a human review or approval before the action is executed.
Audit Trails: Logging every action taken by the LLM and the tools it uses to provide full transparency for compliance and auditing.
With great power comes great responsibility, and in a B2B platform, that responsibility falls on the administrator. We can't just throw all these tools at a client bank and expect them to sort it out. The bank needs to maintain control and curate the experience for its users.
This means we need to build robust product features for the MCP administrator. They need the ability to:
Curate Tools: Select which specific tools are available to different departments or user profiles.
Set Permissions: Define which users can access and use which tools, and what actions they are allowed to perform.
Add Custom Tools: The Holy Grail of customization. What if a bank wants to build its own proprietary tool and integrate it via MCP? We need to provide a framework that allows them to do this securely, giving them the flexibility to extend our platform to fit their unique needs.
MCP is powerful, but it's not magic. The real value for financial institutions will come not just from the connections themselves, but from the intelligent, secure, and user-friendly layers we build on top of them. That's our focus as product leaders.