Build a Voice AI Agent to Answer Customer Invoice Questions
Voice AI agents use natural language processing to identify the customer and their specific question. They then query accounting systems via API to retrieve real-time invoice details or payment status.
The complexity depends on your existing systems. A business using Stripe Billing with a clean API is a straightforward build. One using a legacy on-premise ERP with no API requires a more involved database integration strategy. The agent's performance relies on direct, fast access to your financial data source.
We built a voice agent for a 15-person B2B SaaS company that was fielding over 300 billing calls per month. The system went live in 4 weeks and now handles 85% of these routine calls automatically. This freed up their two-person finance team by nearly 10 hours per week.
What Problem Does This Solve?
Most businesses start with a standard phone tree (IVR). These systems force callers into rigid menus like "Press 1 for billing," which cannot handle natural questions like "Was my last payment successful?" This poor experience leads to high abandonment rates, where frustrated customers press '0' repeatedly just to reach a human, defeating the purpose of the system.
A step up is a visual chatbot builder, but these tools are not designed for secure voice conversations involving financial data. Connecting a tool like Google's Dialogflow to your QuickBooks Online account requires custom middleware to manage API authentication, token refreshes, and secure data handling. Without an engineering team, building this secure bridge between the bot and your accounting system is a non-starter.
This leaves small teams stuck in a manual loop. For a subscription box company with 8 employees, this means one person spends half their day answering the same questions about invoices and payment dates. They know automation is the answer, but off-the-shelf tools either offer a poor customer experience or demand a technical integration they cannot build.
How Does It Work?
We begin by mapping the top 5-10 billing questions your team answers manually. We then build direct API integrations to your accounting system, whether it is Stripe, Chargebee, or QuickBooks Online. We use Python with the httpx library for asynchronous API calls, ensuring the agent retrieves customer data in under 300ms.
The agent's core uses the Claude API for natural language understanding. A custom prompt, engineered with examples of your specific customer questions, allows the model to interpret intent accurately. It understands that "How much do I owe?" and "What's my outstanding balance?" are the same query. The response is converted back to speech using a near-human quality text-to-speech service.
The entire system is a single Python application built with the FastAPI framework. We deploy it on AWS Lambda, which keeps hosting costs under $50/month for most clients, even with up to 10,000 calls. When a customer calls, the Lambda function executes the entire conversational turn, from speech-to-text to API lookup to audio response, in less than 2 seconds.
To protect customer information, the agent first authenticates the caller by asking for an account ID or the last four digits of their credit card. This is checked against your CRM or a Supabase database before any financial data is read. We use structlog for anonymized logging to monitor performance without ever storing sensitive information. A typical build is live in 3-4 weeks.
What Are the Key Benefits?
Answer 80% of Billing Calls Instantly
Deflect routine invoice and payment status questions 24/7. Your agent responds in under 2 seconds, eliminating customer wait times and freeing up your finance team.
Pay for a Build, Not Per Minute
A fixed-price build with minimal monthly hosting costs on AWS Lambda. Avoids the high per-minute or per-call fees charged by most contact center platforms.
You Get the Keys and the Code
We deliver the full Python source code to your company's GitHub repo. You have complete ownership and can modify the agent's logic in the future.
Know Immediately When a Call Fails
The system uses structlog for detailed performance logging and sends a Slack alert if API error rates exceed 5%, so we can fix integration issues proactively.
Connects Directly to Your Books
Direct API integrations with Stripe, QuickBooks, and other modern accounting platforms. The agent provides real-time data, not cached or delayed information.
What Does the Process Look Like?
Week 1: Scoping and API Access
You provide a list of common questions and grant read-only API access to your accounting system. We define the exact conversational flows and authentication methods.
Week 2: Core Agent Development
We build the FastAPI service that connects your APIs to the Claude API for language understanding. You receive a text-based version of the agent to test.
Week 3: Voice Integration and Deployment
We integrate text-to-speech and speech-to-text services and deploy the agent to AWS Lambda. You receive a dedicated phone number for live testing.
Weeks 4-6: Monitoring and Handoff
We monitor live call transcripts for 2 weeks to tune the agent's accuracy. You receive the final source code, full documentation, and a runbook for maintenance.
Frequently Asked Questions
- What factors determine the cost and timeline for a voice agent build?
- The main factors are the quality of your accounting system's API and the number of distinct conversational paths. A system with a modern REST API like Stripe is faster to build than one requiring a custom connector. A project to answer 5-10 common questions typically takes 3-4 weeks. The cost is a fixed-price engagement based on this scope.
- What happens if the agent doesn't understand the customer or an API is down?
- If the agent fails to understand a query after two attempts, it gracefully escalates the call to a human agent and provides a transcript of the conversation. If a backend API is down, the agent informs the caller it cannot retrieve their information at the moment and suggests calling back later. These failure events trigger a Slack alert for our monitoring.
- How is this different from a service like Twilio Flex or Amazon Connect?
- Twilio Flex and Amazon Connect are comprehensive Contact Center as a Service (CCaaS) platforms. They are powerful but complex and expensive, designed for large support teams. Our approach builds a lightweight, serverless agent focused only on automating specific tasks. You get a targeted solution without paying for a full-blown contact center platform you do not need, and you own all the code.
- How do you ensure customer financial data is handled securely?
- We use several security layers. The agent authenticates callers before accessing data. All API keys are stored securely using AWS Secrets Manager. Sensitive data like invoice amounts is never written to logs. The application logic runs in memory on AWS Lambda and is stateless, reducing the attack surface. We strictly follow the principle of least privilege for all API access.
- Can the agent handle different languages or accents?
- Yes. The underlying speech-to-text models we use are trained on vast datasets and perform well with a wide range of common accents. For multilingual support, we can deploy separate agent instances for each language. The scope for a multilingual build is larger, as each language requires its own set of conversational prompts and testing, but the core architecture remains the same.
- If we add a new service, can we update the agent's responses ourselves?
- Yes. The agent's core logic and conversational prompts are written in Python and are well-documented. To add a new query type, a developer would add a new function to call the relevant API endpoint and update the master prompt for the Claude API. The runbook we provide includes step-by-step instructions for making these common types of updates to the agent.
Related Solutions
Ready to Automate Your Small Business Operations?
Book a call to discuss how we can implement ai automation for your small business business.
Book a Call