Live in Production
2024
SaaS support teams spend a disproportionate share of their time on repetitive, low-complexity queries — password resets, billing questions, feature clarifications — while high-value issues queue behind them.
The traditional solution is to hire more support agents or push users toward a knowledge base. Neither works well: hiring is expensive and slow to scale, and knowledge bases require users to search for their own answer rather than just asking.
The specific challenge for CreekSoftware was that embedding FAQs directly in the AI's system prompt would mean stale information the moment anything changed, and an AI trained to answer confidently would hallucinate when asked something outside its static prompt — giving wrong answers with full conviction.
I built Alex — a Vapi voice agent backed by a dynamic Airtable knowledge base rather than a static prompt. Every question triggers a real-time lookup, so the agent only answers from current information. When it can't find a match, it escalates rather than guessing.
Vapi receives every inbound support call and connects Alex. The agent greets the caller, confirms the product they're calling about, and asks them to describe their issue. GPT-4o drives the conversation, handling natural phrasing, interruptions, and the varying ways customers describe the same problem.
Rather than answering from a static system prompt, Alex triggers a real-time Airtable lookup via an n8n webhook every time a question is identified. The FAQ database contains answers, related topics, and a 'confidence required' flag. When a record is found, Alex answers from the current data. When no match exists, the agent acknowledges the gap and initiates escalation — it never invents an answer.
Identified issues are classified by type (billing, technical, account access, feature request) and severity. For issues within scope, Alex walks the caller through resolution steps, confirms whether the issue is resolved before ending the call, and collects a satisfaction rating. Resolution steps are pulled from the same Airtable knowledge base, keeping the agent's guidance current with product changes.
When Alex cannot resolve the issue, it prepares a structured handoff before transferring the call. n8n assembles a summary containing the caller's account details, the issue category and severity, a transcript of the conversation so far, and any resolution steps already attempted. The human agent receives this context before the call connects — the customer never has to repeat themselves.
Every call — regardless of whether it was resolved or escalated — generates a support ticket in HubSpot via n8n. The ticket includes the issue type, resolution status, call duration, satisfaction rating, and the full interaction log. This creates an auditable support history and surfaces recurring issues that indicate product or documentation gaps.