AI Call Intake and Intelligent Routing with Pulastya
SDLC Corp replaced manual phone triage with an AI voice agent that understands every inbound caller in natural speech, classifies the call into a defined intent, answers from the client's approved knowledge base, and transfers the rest to the right department with transcript and summary already attached.
- Keeps your existing Twilio number
- Answers only from approved documents
- Every transfer carries context
Project Snapshot
The product, focus, and scope behind this AI call intake and routing deployment.
Service-Based Business
Customer Service & Call-Heavy Operations
AI Call Intake, Classification & Routing
Pulastya AI
The Challenge
Where Manual Call Triage Slowed Everything Down
Before this deployment, five conditions shaped how the client's team handled inbound calls. Each one is a constraint the build had to remove, and together they set its scope.
Callers Needed Interpreting First
People do not describe their problem in system terms. Staff had to listen, interpret, and only then decide where the call belonged.
Menus Pushed Work to Callers
Callers had to self-classify against a fixed list, guess the closest match, and often land in the wrong queue, repeating the whole story after a transfer.
Documented Answers Ate Staff Time
A large share of calls had answers already written down in service documents, policy files and FAQs, yet those calls still occupied a person for several minutes each.
Transfers Arrived Without Context
When a call reached the right team, the receiving agent started from nothing, and the caller repeated their name, reference and issue all over again.
Five Requirements Set
Before the Build Began.
Five requirements defined how Pulastya AI should understand callers, use approved knowledge, capture requests, and route every call with context.
Understand Naturally
Understand the caller in natural conversation, without menu prompts or keyword commands.
Classify Every Call
Classify each call into a defined intent before deciding what happens next.
Answer From Approved Knowledge
Answer documented questions directly from the client's approved knowledge base, not general model knowledge.
Capture Structured Requests
Capture structured service request details when the caller needs something logged.
Route With Context
Transfer any call the AI should not handle to the correct department, with context already prepared.
Planning the Same Shift?
Planning a similar shift from manual call triage to AI-driven intake? These requirements shaped everything that followed. Talk to our AI voice team.
The Pulastya AI Solution
One Intake Layer Built In Three Connected Pieces
Natural conversation, approved-knowledge answers, and a routing decision, applied to every inbound call.
Conversation Understanding & Intent Classification
Pulastya AI sits in front of the client's existing phone numbers. It handles the conversation itself, on the Pulastya AI platform. Barge-in and turn-taking let callers speak naturally. Every call is classified into a defined intent before anything else happens.
- Barge-in and end-of-speech detection
- Intent mapped to the client's real call types and queues
- No menu prompts or keyword commands
Approved-Knowledge Answers Or Request Capture
Answers come only from documents the client uploaded and approved. That means service details, pricing, policies and FAQs. When the agent lacks enough grounding, it says so rather than guessing. Some calls are requests rather than questions. For those, it captures the fields the downstream team needs.
- Retrieval limited to approved documents only
- Structured service-request field capture
- No answer given without sufficient grounding
- Caller identity and request details recorded
Deterministic Routing & Contextual Handoff
Once intent is known, the call goes to the department or queue that owns it. The client configures those rules. Any call passed to a human carries the transcript and an AI-generated summary. The caller does not have to restart.
- Same intent always routes the same way
- Transcript and summary attached to every transfer
- Receiving agent has full context immediately
From Inbound Call To Resolved Outcome
Five steps, running the same way for every call. Knowledge retrieval sits inside the workflow at step three, which is what stops the agent from answering beyond what it actually knows.
Inbound Call Received
A caller reaches the business through its existing phone number.
Conversation Understood & Intent Classified
The agent identifies what the caller wants from natural speech.
Knowledge Retrieval Or Request Capture
The agent answers from approved documents or collects structured request details.
Routing Decision Applied
The call is directed to the department or queue that owns that intent.
Resolved By AI Or Transferred With Context
The call closes, or a human agent continues with the transcript and summary already attached.
Grounding stays inside the workflow. Nothing is answered until it is checked against approved knowledge, and nothing is transferred without context.
Implementation Approach
From Setup to Live Call Handling
The setup prepared Pulastya AI to answer routine calls, use approved business information, capture caller details, and route the rest with full context.
- Step 01
Workspace And Number Setup
A Pulastya AI workspace was created and connected to the client's existing Twilio phone numbers, so no number changes were required.
- Step 02
Knowledge Base Preparation
Approved documents were uploaded and indexed. Content that was out of date or contradictory was resolved before indexing.
- Step 03
Intent And Routing Design
Call types were defined from the client's real call mix, then mapped to departments and queues.
- Step 04
Browser Test Calls
The team spoke to the agent from the browser before any live traffic, checking classification accuracy and answer grounding.
- Step 05
Live Call Handling
The number was pointed at Pulastya AI and the agent began taking real inbound calls.
- Step 06
Review And Tuning
Transcripts and summaries were reviewed to find intents that needed sharper definition or documents that needed updating.
Technology Stack
Technology Powering AI Call Handling
Voice, conversation, platform and data layers used to build the call intake and routing capability described in this case study.
Primary references: Twilio Programmable Voice, OpenAI Realtime API, ElevenLabs, pgvector.
Twilio Twilio Polly 
ElevenLabs OpenAI TTS OpenAI Realtime API 
React 18 
Node.js 
Socket.IO JWT HTTPS Express.jsREST APIWebhooks
PostgreSQL pgvector pgcrypto PostgreSQL JSONB SQL Data LayerSemantic SearchCall MetadataEncrypted FieldsWant to know which stack fits your existing systems?
Discuss Your Integration ArchitectureArchitecture Decisions
Why This Stack Was Chosen
The architecture separates telephony, live conversation, application logic, retrieval, storage, and security so each layer has a clear responsibility in the call path.
| Component | Role | Why Chosen |
|---|---|---|
| Twilio Programmable Voice | Telephony and call connection | Connects Pulastya AI to the client’s existing phone numbers so inbound calls can enter the AI workflow without changing the caller-facing number. |
| OpenAI Realtime API | Live conversation and intent understanding | Supports low-latency conversational turns so the agent can understand natural caller language and decide what should happen next during the call. |
| ElevenLabs + OpenAI TTS | Voice generation | Turns generated responses into spoken output, giving the voice layer flexible synthesis options for natural call interactions. |
| Node.js + Express.js | Backend orchestration and APIs | Provides the event-driven application layer that coordinates telephony events, AI requests, routing logic, integrations, and internal API endpoints. |
| Socket.IO | Real-time application events | Keeps browser and server state synchronized during testing, monitoring, and live call activity where updates need to appear immediately. |
| React 18 | Browser interface | Supports the component-based interface used for workspace configuration, test calls, operational views, and other interactive workflows. |
| PostgreSQL + JSONB | Structured and semi-structured data storage | Stores call records, configuration, routing data, and flexible metadata in one durable relational data layer. |
| pgvector | Semantic retrieval | Stores and searches embeddings so Pulastya AI can retrieve relevant approved knowledge by meaning rather than relying only on exact keyword matches. |
| pgcrypto + JWT + HTTPS | Data and session security | Protects sensitive fields, authenticated application sessions, and data moving between browser, application, and integration endpoints. |
Outcomes
What Changed in Day-to-Day Call Handling
No measured figures are published for this deployment yet. Here is what changed structurally in how calls are handled.
Triage Moved off the Front Desk
The decision about what a call is and where it belongs is made during the conversation, not by a person listening and then transferring.
Callers Stopped Navigating Menus
Intent is identified from natural speech, which removes the guesswork that sends callers into the wrong queue.
Documented Questions Resolve Inside the Call
Calls with an answer already in the knowledge base close without occupying a staff member.
Transfers Arrive Prepared
Agents receive the transcript and summary, so the conversation continues instead of restarting.
Every Call Leaves a Structured Record
Intent, routing decision, transcript, summary and outcome are captured per call, giving the client a view of their call mix that manual handling never produced.
Common Questions
AI Call Intake and Routing with Pulastya AI
The questions teams ask most often when evaluating AI call intake and routing.
An IVR asks the caller to classify themselves against a fixed list. Pulastya AI classifies the call from what the caller actually says, so there is no menu to navigate and no wrong button to press.
It says it does not have the information and offers a transfer. The call moves to a human with the transcript and summary already attached. It does not guess.
Yes. Pulastya AI connects through Twilio, so an existing Twilio number can be pointed at the platform without changing what customers dial.
Answers come from the documents you upload and approve. Updating pricing, policies or service details means updating those documents and reindexing the knowledge base from the dashboard.
Intents are mapped to departments and queues during setup, based on your real call types. The mapping is deterministic, so the same intent always routes the same way.
It is built for administrative and informational call handling. It should not be used for diagnosis, emergency support, financial advice, legal advice or sensitive account verification.
See Pulastya AI Handle Your Call Types
Bring your real call mix and your existing documents. We will show how intent classification and routing would work against them, and where the AI would hand off to your team.
