Custom AI Voice Agent & Workflow Automation

Designed and built a custom AI voice-agent solution from the ground up for a client in the service sector, together with the connected business automation needed to manage customer inquiries from CRM intake through outbound conversations, appointment scheduling, call outcomes, and operational reporting.

The engagement included original agent construction, conversation controls, tool integrations, custom integration code, workflow architecture, debugging, testing, and a detailed operating handover. It was not limited to installing or configuring an existing voice agent.

Designed the AI caller’s role, persona, voice delivery, speaking pace, conversational structure, qualification boundaries, and customer-facing interaction behavior in Vapi.

Authored and refined the agent’s system instructions and conversation logic from the ground up: approved openings, qualification sequence, objection responses, appointment transitions, escalation paths, closing behavior, and rules for what the agent must never say.

Engineered source-grounded response controls: connected the approved business script and knowledge resource, required the agent to use authorized answers, and restricted unsupported claims about pricing, savings, financing, incentives, product terms, or availability. Unanswered questions were constrained to approved responses or appropriate human follow-up rather than invented explanations.

Designed tool-use rules in Vapi: the approved knowledge source answered product questions; calendar availability was checked through its dedicated tool; appointments could be created only after customer confirmation and required details; termination tools delivered the authorized closing and ended calls without duplicated messages.

Prompt-engineered conversation-state handling to recognize previously supplied customer details, skip answered questions, resume after interruptions or tool calls, avoid repetitive acknowledgments, and maintain a natural speaking pace. Tuned pronunciation, response timing, turn-taking, and call endings through controlled voice testing.

Defined separate live-answer, silence, no-answer, and voicemail behaviors. Voicemail-specific instructions prevented the agent from treating a recorded greeting as a person, repeating its introduction, continuing qualification, or speaking an unapproved message; approved callback language and termination controls were tested.

Designed and built Zapier as the central workflow engine that initiates and coordinates the automated customer-contact process.

When Zapier detects a new eligible Lead in Pipedrive, it retrieves the linked Lead and Person information, checks consent and permitted calling hours, prepares the customer context, and sends the outbound call request to the AI voice agent built in Vapi.

At the end of the conversation, Vapi sends its completed-call event back to a separate Zapier workflow. Zapier then interprets the call outcome and distributes the results automatically to the appropriate Pipedrive record and the client-requested Google Sheets operational log.

Designed the systems to work as one coordinated automation, with Zapier directing the event flow rather than treating each application as a separate task. Pipedrive is the source CRM for customer and Lead records and the destination for call-result updates; Vapi hosts the custom conversational agent and produces call-event data; custom Google Apps Script connects the agent’s booking tools to Google Calendar; and Google Sheets is the client-specified reporting destination, populated automatically by Zapier.

Defined how the originating CRM customer and transaction references travel through call initiation, conversation processing, appointment actions, event callbacks, and writeback.

Engineered two complementary Zapier workflows: an intake and outbound-calling workflow that retrieves CRM context and invokes Vapi, and a post-call workflow that receives Vapi events, processes call results, and updates the CRM and operational report.

Replaced fragile phone-number matching with direct CRM record-reference continuity so outcomes could be written to the correct customer record.

Reengineered the original CRM transaction workflow into a lead-first design after the client refined the required business process, including new triggers, lead-specific fields, consent evaluation, downstream mapping, and conversion handling.

Established a custom CRM consent field and an automated eligibility gate so an explicit affirmative consent value proceeds while declined or missing consent stops the outbound-call path.

Created custom JavaScript calling-hours logic with explicit timezone handling; integrated delay controls to hold eligible inquiries until an authorized calling window.

Implemented Zapier-controlled Pipedrive Lead/Person retrieval, contact-data preparation, consent and calling-hours decisions, outbound-request construction, and the Vapi call initiation after eligibility checks.

Tested positive, negative, and blank consent cases; in-hours continuation; and before/after-hours scheduling decisions using controlled records.

Protected CRM record continuity so the originating lead/contact context remains associated with the voice call and downstream operational outcome.

Built and connected dedicated AI scheduling actions to check live calendar availability and create appointments after a customer confirms an acceptable time.

Developed a Google Apps Script integration bridge to address limitations in the initial direct calendar/webhook path, including response-time and integration constraints.

Handled calendar availability requests, appointment creation, timezone-sensitive scheduling behavior, and booking checks against existing calendar commitments.

Validated the calendar integration with controlled appointment and availability tests and connected successful bookings back into the CRM workflow.

Connected completed-call event webhooks and developed the downstream processing path for call recordings, transcripts, termination reasons, call metadata, and operational results.

Created structured outcome handling for cases such as successful appointments, no answer, voicemail, and other relevant follow-up states; used normalization and fallback handling to keep reporting usable when conversations lacked intelligible responses.

Built the automatic data pipeline expressly required by the client. Following each completed AI call, Vapi sends its completed-call report to Zapier; Zapier associates that event with the originating Pipedrive Lead and Person, normalizes the transcript-derived outcome, and writes structured call information directly to the client’s Google Sheets operational log.

The sheet receives call references, relevant customer context, disposition, booking result, and human-follow-up indicators without staff manually copying information between systems. Zapier also writes the result back to the correct Pipedrive customer record.

Built Pipedrive note/writeback actions so the call outcome is recorded against the originating customer/lead record rather than a newly created or incorrectly matched contact.

Engineered the successful-appointment branch that converts an eligible originating Lead into a Deal and places it into the appropriate next-stage business process while preserving associated customer information.

Preserved non-booked contacts as Leads and exposed operational follow-up needs without automatically converting every contact into a sales opportunity.

Executed controlled real-world voice-call tests for answered conversations, silence/no-answer, voicemail handling, approved closing behavior, call recording/transcript generation, and termination.

Verified the Zapier-directed sequence with controlled evidence: the Pipedrive Lead and Person were retrieved; Vapi received the correct customer context; Vapi returned completed-call events to Zapier; and Zapier automatically wrote the processed result to both the original Pipedrive record and the client-requested Google Sheets log.

Investigated integration problems involving CRM API responses, custom-field retrieval, outdated workflow assumptions, event payload visibility, and mapping mismatches; corrected the relevant data paths instead of relying on superficial success indicators.

Reworked and regression-tested CRM orchestration when the intake model changed from deal-first to lead-first, including consent values, direct record association, outcome handling, and appointment-driven conversion.

Troubleshot voice and telephony behavior with recorded calls and technical support evidence; refined voicemail handling, call termination, and integration safeguards.

Validated individual components and connected behaviors using explicit test cases and evidence across the CRM, AI assistant, orchestration platform, calendar, and reporting outputs.

An answered outbound test call produced a live AI conversation, a recording, a transcript, call metadata, and a completed-call record for downstream inspection.

A controlled no-answer call triggered automated structured processing: the outcome and follow-up indicator were recorded against the originating Pipedrive record and in the client’s Google Sheets operational log.

Voicemail testing verified the approved voicemail message, callback details, and termination behavior without continuing the ordinary qualification conversation.

Calendar testing confirmed the ability to check appointment availability and create booked events using the custom scheduling integration.

CRM tests confirmed consent-based eligibility, preservation of customer record context, and the appointment-triggered Lead-to-Deal transition.

Original AI voice agent built from scratch, including prompt architecture, scripted and knowledge-grounded responses, hallucination-reduction guardrails, conversation-state logic, controlled tool use, qualification, booking, voicemail, and call termination.

Delivered Zapier-centered automations covering the complete operational flow: Zapier detects eligible new Pipedrive Leads and retrieves linked customer details; Zapier applies call controls and initiates Vapi; Vapi conducts the approved agent conversation and returns call results to Zapier; and Zapier processes those results and populates the correct Pipedrive record and the client-requested Google Sheets report.

Custom Apps Script scheduling bridge and JavaScript-based workflow controls.

CRM record continuity, automated result writeback, appointment-driven lead conversion, and operational follow-up visibility.

Detailed Operational Readiness & Governance Manual covering system roles, inputs and outputs, business controls, testing and inspection procedures, operational dependencies, change management, and support/escalation guidance.

Vapi (custom AI voice agent) | Pipedrive (CRM) | Zapier (workflow and data orchestration) | JavaScript | Google Apps Script (calendar bridge) | Google Calendar (booking) | Webhooks and APIs | Google Sheets (client-required reporting destination)