How We Scaled Inbound Sales

Every week, hundreds of enterprise sales forms come in, but in reality most of them are actually better suited to the self-serve plan. Previously, the sales team manually reviewed submissions to sort out leads that needed an enterprise conversation from those that should be directed to self-serve, and when information was insufficient, additional follow-up was required, which slowed down response times.

To close this gap, we built an inbound AI SDR on top of the Agents Platform, named Jon. The goal is to let leads book a meeting immediately after submitting the form, with a focus on eliminating the problem where manual review meant a lead coming in on Friday evening had to wait until Monday.

The build was organized around three pillars.

• Behavior: The system prompt defines Jon's personality and role. It's set up to maintain an approachable tone, like a "warm, consultative sales representative," and also specifies behavioral standards such as booking meetings or sending messages to teams.
• Capabilities: We fed in a knowledge base built from over 100 FAQs and product documentation. Initially we started with a low-latency model, but due to insufficient reliability and tool calling, we switched to a more powerful LLM. As a result, hallucinations like misidentifying the year decreased, and it became capable of reasoning, such as estimating monthly usage by asking about daily call volume and average call duration.
• Data: After a call ends, the lead's use case and qualification decision are logged directly into the CRM, along with structured data such as customer satisfaction. At the end of each call, we ask for a CSAT score from 1 to 10, and conversations with the highest or lowest scores are reviewed daily.

Clear lessons emerged from operating this system. We consider a wrong qualification better than a wrong disqualification, so we tuned it to favor false positives over false negatives. Also, trying to account for every exception bloats the system prompt, increasing the risk of errors, and entrenches issues further, such as the habit of asking multiple questions at once. We used guardrails to block attempts by users to bypass the flow mid-conversation by asking to be connected directly to HR or a specific contact, and mitigated the problem of long silences during tool execution with a force message feature that communicates progress.

Results came quickly. It currently operates 24/7 across 38 countries, handling 50+ calls per week, effectively performing the role of two SDRs. 78% of qualification decisions are completed without human intervention, and the remaining 22% are mostly false positives, where leads that should have actually been excluded were incorrectly passed through. Conversely, false negatives are rare, and mostly occur from ending calls too early, so going forward we plan to handle these incomplete cases as N/A instead of disqualifying them.

The average CSAT was 8.7, with leads giving high scores for the experience of having a natural conversation in multiple languages and then immediately moving on to schedule a meeting. Next steps include using the workflows feature to handle more complex, dynamic conversations, and expanding into outbound sales and re-engaging unresponsive leads. Ultimately, we're aiming to automatically route each meeting to the calendar of the appropriate AE, moving toward a future where voice becomes the default interface for technology.