Financial agents need access to user context, banking data, and action-taking integrations because retrieval alone cannot resolve real customer problems.
2
A validation and reasoning step should check the agent's proposed tools, sources, and expected results before any action is performed.
3
Voice agents must balance response speed with accuracy, authentic speech, constrained actions, and strict transfers to human staff.
Summary
Vamsi Saladi explains how Contour builds AI customer support for banks, credit unions, and fintechs. Financial agents operate under tighter constraints because they handle money, complaints, and regulated customer communications. A useful system must authenticate users, apply procedures and guardrails, retrieve account context, inspect past interactions, and use banking or payments integrations to resolve issues. Vamsi adds a validation step before tool use, where another process checks the proposed action, its sources, and its expected result. Voice introduces extra limits because customers cannot wait through several model calls, and angry callers react badly to artificial-sounding speech. Contour therefore uses faster, more constrained validation for voice, along with sentiment-based bailouts and transfers. Vamsi recommends co-pilots as a way for institutions to inspect and approve agent work before adopting fully automated support. He is candid that the orchestration problem is still open.
Financial agents must handle real account problems rather than only answer questions
Vamsi introduces Contour as an AI customer support platform for financial institutions. It supports voice, chat, and email, and tracks performance for both human and AI support agents. He says finance is especially difficult because agents operate near money and regulated communications, leaving little room for mistakes. A surface-level support bot that only answers from a knowledge base does not solve enough of the institution's problem. A useful agent needs to resolve the underlying issue, which may require reading account information, checking transaction history, or taking an action through a payments or banking integration.
Guardrails must be built into the agent's workflows
Vamsi describes several ways to make financial agents compliant. Institutions can provide knowledge-base content, standard operating procedures, and documents that define how agents should respond. Workflow rules can also specify canned or default responses for particular situations. He says a knowledge-base integration is now a baseline and does not distinguish a system by itself. The agent should cite its sources and explain why it took an action. That visibility gives the financial institution a way to inspect the agent's behavior instead of accepting an unexplained answer.
A production workflow combines identity, policy, user context, and actions
The simple pattern of sending a ticket to retrieval and returning an answer is insufficient for banks, credit unions, and fintechs. Vamsi's minimum workflow first authenticates the user, which establishes who they are, what information can be shared, and which rules apply. The system then considers the user's profile alongside procedures and knowledge-base content. A fuller design runs two related lines of work. One reasons about the complaint and applicable procedures. The other checks the account, banking relationships, available services, transaction history, and previous interactions. Combining those results lets the agent respond with relevant context and take a useful action.
Tool calls need an independent validation step before execution
Contour does not let the language model perform actions or use tools directly from the initial call. Vamsi says that approach creates too much error risk for production finance systems. Instead, the proposed action goes through a validation and reasoning step. That check asks whether the action follows the guardrails, compliance information, and operating procedures. It also asks the agent to explain why it chose the action, which tools it will use, and what result it expects. In his transaction example, the system compares transaction locations and history before submitting a dispute. This process reduces errors and gives the institution an explanation for the tool call.
Voice support requires authentic speech as well as accurate answers
Chat and email can tolerate several seconds of processing because the interaction is asynchronous. Phone calls cannot. Vamsi says callers to financial institutions are often already angry or worried about their money, so an artificial voice can make the situation worse. He advises against relying on a default text-to-speech voice that only sounds good in curated demonstrations. Contour fine-tunes its voice around support conversations, including their tone, inflections, pauses, and ordinary phrases. Vamsi says voice quality can be almost as important as language-model accuracy when the goal is to keep a caller engaged long enough to resolve the problem.
Voice systems should spend speed on safe, constrained decisions
Vamsi calls latency the showstopper for voice agents, while incorrect information is a deal breaker. Contour accepts trade-offs between model cost, context size, and response speed, but he says small latency gains are not worth losing accuracy. The validation stage remains necessary for voice, although it must run faster than it does for chat or email. Contour uses smaller models, narrower validation questions, and often a single validation layer. It can also play templated placeholder speech while checking a tool result, so the caller does not experience a silent pause. Limiting the agent's action space is another way to improve speed without giving up all safeguards.
Co-pilots give financial institutions a controlled path to automation
Vamsi presents a co-pilot as an intermediate option between manual support and fully automated agents. A human can approve or edit the proposed response while seeing how the system reached it. This lets the institution test workflows, inspect explanations, and add guardrails before allowing the agent to act alone. He also recommends exposing more than the initial model response. The system should show the original answer, the relevant reasoning trace when available, the counter-check, the sources, and the reason a tool call was necessary. For voice, Contour adds strict bailouts for strong anger, sensitive subjects, and transactions above an institution's chosen value threshold.
Access to real money is usually mediated by third-party providers
In the discussion, Vamsi says voice agents almost never receive direct write access to the banking core, and the same is generally true for other agents. Institutions often give access to a third-party banking or payments provider instead. That provider has controls over what can affect the actual account balance. An agent may submit a dispute or refund request to the third party, while the provider manages the operation involving real money. This design separates the agent's ability to initiate a support workflow from unrestricted control over the underlying account.
"We don't want the LLM to perform actions or use tools straight from the call because there you are introducing kind of a margin for error that is too large to be usable in a production environment."08:32
Who should watch
You are building customer support agents for a bank, credit union, or fintech and need to connect answers to authenticated account data and real actions.
Your voice agent works in demonstrations but struggles with angry callers, delays, artificial speech, or unsafe tool calls.
Your institution wants to introduce agent automation while keeping human approval, explanations, and strict transfer rules in the workflow.
