Scaling Compliant AI Agents in Financial Services: Watch Webinar Recap
- Symphony
- Tech4Fin
Find out more about Symphony
Financial firms are currently at a critical inflection point. Just a few short years ago, AI agents were largely a conceptual exercise, confined to the edges of innovation labs. Today, the conversation has shifted entirely. Institutions are no longer asking if they should deploy AI agents across their workflows, but how to do so without introducing compliance exposure, data risk or operational gaps.
The challenge for technology and risk officers is balancing rapid innovation with the stringent controls, auditability and governance required in regulated industries. It is a tension between the need for speed and the non-negotiable requirement for security.
In our recent webinar, “Moving from Proof of Concept to Production AI Agents in Financial Services,” we gathered industry experts to discuss how to navigate this shift. Audrey Costabile, Senior Analyst at Coalition Greenwich, joined Symphony’s Rob Friend and Richard Ring to outline the path from pilot programs to enterprise-wide deployment.
What you’ll learn in this webinar
This session is designed to provide a realistic roadmap for firms currently scaling their AI initiatives. We move beyond the hype to show exactly how Symphony’s AI Studio bridges the gap between powerful agentic capabilities and institutional-grade governance.
Key highlights from the discussion include:
- A live demo of Symphony AI Studio: See how to build, deploy and govern agents directly within the Symphony ecosystem, including the use of MCP infrastructure to connect safely to downstream systems.
- The governance blueprint: Learn how to apply the same administrative controls, permissioning and encryption standards used for human users to AI agents, ensuring sensitive financial data remains secure.
- Expert perspectives on scaling: Rob Friend shares insights from across the market on how firms are evolving from simple internal efficiency tools to more complex, client-facing AI workflows.
- Mitigating risk: Understand how to implement human-in-the-loop oversight and preventative controls that flag malicious prompts before they reach the language model.
Whether you are a developer looking for tools to build the next generation of agents, or a risk officer concerned about auditability and data leakage, this session offers a practical look at where the industry is heading.
Unlock the full strategy: Complete the form
The path from a proof of concept to a fully operational, compliance-enabling AI agent doesn’t have to be opaque. Symphony provides the infrastructure to safely build, manage and scale AI, drawing on our verified trust network and years of experience serving the financial industry.
Complete the form below to access the full demo and Q&A video featuring our experts.
Questions from the audience
Integrations & MCP connectors
What different integrations are required to provide this contextual knowledge?
The trade break example used the context of the conversation that the user was active in. The agent had the permissions to access that chat context to perform the analysis on behalf of the user. The agent also used a JIRA/Confluence MCP to compare this information to existing tickets and suggest ticket creation.
Are you working with any industry MCP servers?
We’re working with Tradefeeder and other firms like Carbon Arc and Trumid, and we’re in talks with many others. These are optional services to plug in as MCP, so you can have natural language chat with the data residing where it lives without needing to expose the entire data set. It’s something we’re really excited about because it helps bridge and reduce friction for a lot of data acquisition for our customers. It also helps with distribution for those data vendors themselves. This is not exclusive to data vendors – this can apply to any third-party MCP services. In the demo, we showed quite a few public ones like Jira and Salesforce. There are plenty of other ones there that will be added as we see demand.
Data privacy & DLP controls
Individual trade data is PI data; how can that be? Does single sign-on work with this Agent?
The trade break agent we showed acted on behalf of the firm’s user and analyzed the chat between them and the client. This means that no data was passed by the agent outside of the firm. Symphony already has strict controls that are enforced by the firm’s administrator to prevent data leakage. For example Admins can enforce chat rooms to only be with 2 parties (to prevent multi-party rooms). Identities of users are provided by the participant firms, so the directory entries are controlled and there is no chance of misrepresentation, spoofing or phishing.
Can requests to these custom agents be intercepted before execution to enforce preventive controls and block malicious prompts?
Both the human-to-agent and the agent-to-LLM are operated in Symphony rooms which adhere to preventative DLP controls. Symphony already supports existing human-to-human conversations. If a human or agent attempts to generate content in a room that breaches these controls their content will be refused and not sent by the system (this is a preventative measure to block vs. a reactive measure to alert).
For Agentic AI, how secure is PI data like trade information? How can anyone get the confidence?
The trade break agent we showed acted on behalf of the firm’s user and analyzed the chat between them and the client. This means that no data was passed by the agent outside of the firm. Symphony already has strict controls that are enforced by the firm’s administrator to prevent data leakage. For example Admins can enforce chat rooms to only be with 2 parties (to prevent multi-party rooms), Identities of users are provided by the participant firms, so the directory entries are controlled and there is no chance of misrepresentation, spoofing or phishing.
Human oversight
How should firms design human oversight mechanisms for AI-driven workflows?
The ‘human in the loop’ (HITL) design is a common way to provide human oversight in agentic processes. This means that an agent has to have human approval before taking any action. Symphony’s Agent Studio is designed to allow firms to design processes that range from simple question/answer interactions to human-in-the-loop processes all the way to more autonomous ‘act on behalf’ of processes. Depending on the use case being considered, any of these processes can be adopted by a firm.
Agents in chatrooms
Additional cost associated with this feature? Also, if there is an external agent chatroom existing--can you write an agent to 'face off' with this to reduce human engagement?
Yes, there is an additional cost associated with this feature – please contact your Symphony account manager for more details of this.
If there is an existing chatroom, an agent can interact with it in a few different ways.
- An agent can analyse the chatroom ‘on behalf of’ a user (assuming it is granted that permission). This means that the agent can ‘read’ the room as if it were the user and then perform tasks.
- An agent can send a message on behalf of that user (again if permissions are granted for this). If this is required, the agent can post into the room ‘on behalf’ of the user.
- An agent can explicitly join a room (similar to a bot today) and interact with users in that room
We have agents that have been built to summarize rooms, extract actions from rooms and propose ticket creation and/or resolution steps for issues.
Agent identity and verification
When we deploy an agent on Symphony, how can we be sure the identity and constraints we set for it are strictly enforced when it interacts with external counterparties on the network?
Agentic AI requires broad API access, sensitive data integration, and elevated permissions. To unlock productivity, enterprises must solve the high-stakes security and compliance surface created by autonomous actions.
Our platform is trusted by the financial services community. We provide communication and markets technology to 500k+ authenticated financial users and 1,400 institutions globally.
Our unique architecture provides end-to-end encryption for messaging, with data control and audit capabilities built for regulatory compliance.
Some of the key features include:
- Confidential Computing: Agents operate in secure enclaves; data is encrypted and inaccessible even during processing.
- End-to-End Encryption: A private, encrypted channel is maintained between the user and the agent.
- Identity-centric actions: All actions are performed on behalf of the user, strictly respecting existing information boundaries.
- Compliance-ready: Every interaction is captured and recorded in your existing archiving platforms for full regulatory compliance.
WhatsApp, Federation
How do these agents work with your WhatsApp and Federation connections?
Since agents can interact with any Symphony conversation (subject to the access they have been granted), they can interact with Federation conversations. Agents can access conversations on behalf of a user, post messages etc just as they can in a regular Symphony room. There are some limits to the types of rich content that the various consumer networks can handle – but these are the same limits that apply to human users when they post messages.
Wealth management
How can I utilize Symphony’s AI assistants to support my advisory team and me with research and client insights? We’re exploring agents along with improvements to our CRM. It’s important to us to maintain control and provide a personal touch to our client communications.
Symphony helps wealth advisors spend less time gathering information and more time engaging with clients. With AI Studio, you can build your own AI assistants, bring your own model (BYOM), or work with Symphony to develop assistants that surface relevant client history from your CRM, draw on approved research data, and identify potentially relevant products or services based on the context of each conversation.
Symphony can then help you connect with clients through their preferred channels – including WhatsApp, SMS, WeChat and LINE – while supporting your firm’s governance requirements. You remain firmly in control of suitability, professional judgement and the final client communication.
