Exactly what happens on a call — start to finish. This describes the demo as it's built today: a conversational voice line where Amazon Connect and the AI meet live Salesforce data.
Architecture — how it's built
One inbound call flows left to right through the stack. Telephony and the AI live in AWS; the customer data lives in Salesforce.
An inbound phone call.
Contact flow + voice (Polly generative, Matthew). Answers, identifies the caller, plays the greeting, streams the conversation.
Self-service AI orchestrator. Understands natural speech and decides which tool to call — no menus.
Deterministic functions. Pull the data, compose the exact spoken answer, apply policy. The AI never invents a fact.
System of record — contacts, prescriptions, cases — plus the policy knowledge base and a shipment-tracking feed.
Spoken back word-for-word; the call routes to the right queue on fixed rules.
Inside Salesforce, agents see all of it in one screen: the Amazon Connect softphone is embedded in the Service Console via SCC-AC (Salesforce Contact Center for Amazon Connect).
How it's configured
| Telephony & voice | An Amazon Connect contact flow handles the call; the spoken voice is Amazon Polly generative (Matthew). |
| Conversation | An Amazon Q in Connect self-service AI agent (the orchestrator) drives the natural-language dialogue and calls tools — there is no press-1 menu. |
| Tools | AWS Lambda functions, exposed to the agent through an AgentCore gateway. They own every fact and judgment; the model only decides which tool to call. |
| Data | Salesforce over secure API (identity, prescriptions, orders, cases), a Q in Connect knowledge base (policy), and a stand-in carrier API for tracking. |
| Agent desktop | Salesforce Service Console with the Amazon Connect softphone embedded via SCC-AC — one screen for the human agent. |
What happens the moment a call connects
We take the calling number and search Salesforce for a Contact whose phone matches — so the line can greet a known customer by name.
If recognized, the greeting is personalized from their Salesforce record: "Thank you for calling Companion Pet Pharmacy. Welcome back Sarah. How can I help you today?"
No press-1 menu. The caller speaks naturally and an Amazon Q in Connect AI agent handles the conversation — understanding the request, looking up real data, and answering.
When the conversation is done or needs a human, the AI reports who the caller is and what they want, and fixed rules choose the queue. The AI never picks a queue itself — routing is deterministic.
Who the caller is (identification)
We match the calling number against the Salesforce Contact phone field:
| Result | What the IVR does |
|---|---|
| Exactly one match | Recognized. Greet by name and continue. |
| More than one match | Two records share the number — we never guess. The caller is not identified and goes to a person. |
| No Contact match | Fall back to an Account phone match; if still nothing, identify by voice (ask role, then an account number, read back and confirmed, up to three tries). |
What the caller can get answered
Every answer is built in code from real Salesforce data and spoken back word-for-word, so the AI can't rephrase a fact, get vague, or flip a judgment.
Real status, location, expected delivery, and — for a refrigerated medication — whether the cold chain held.
For an order still in processing, the system compares the order date to the standard window (~5 business days) in code. Within it, it reassures; past it, it hands the caller to a person. Never a guess.
If a refrigerated shipment's temperature indicator tripped in transit, the IVR leads with a safety warning and arranges a replacement through a person — it does not just read a status.
Refills: it knows how many are left and whether AutoRefill is on; zero refills needs vet authorization. Cases: it surfaces the caller's open Salesforce case by subject. Clinical/controlled: it never improvises — it consults policy and routes to the right team.
Where the call ends up (routing)
The AI establishes the facts; the phone system chooses the destination from fixed rules — predictable and auditable.
| If the caller is… | They go to… |
|---|---|
| Asking for a person | Straight to a human — immediately, no questions |
| A prescriber | Prescriber queue |
| A pet owner | Support queue |
| Unidentified, ambiguous, or unclear | Support queue — a person always gets it |
See it live on the Try it page, or the exact records behind it on Data & sources.