A multilingual voice AI agent for ecommerce lets you take orders, answer product questions, and upsell shoppers entirely over the phone in any of your supported languages. In this guide you'll assemble a Vapi-driven sales bot, connect it to Shopify via API, and add OpenAI-powered conversation plus Google Cloud Text-to-Speech for truly multilingual responses.
What is a voice AI sales agent? It is an automated conversational system that interacts with customers through spoken language, answering queries, guiding purchases, and completing transactions without human intervention.
| Tool | Plan / Price | Role |
|---|---|---|
| Vapi | Pay-as-you-go (check the current pricing page) | Voice gateway, call handling, multilingual STT/TTS |
| Shopify | Standard plan (starts at $39 /mo) | Product catalogue, order management, GraphQL API |
| n8n | Cloud (free tier available for low-volume testing - verify on n8n.io) or self-hosted (Community Edition - free) | Orchestration of API calls and webhook routing |
| OpenAI GPT-4 | $0.03 / 1 K prompt tokens, $0.06 / 1 K completion tokens (check OpenAI pricing) | Conversational reasoning and dynamic response generation |
| Google Cloud Text-to-Speech | $4.00 / 1 M characters (standard voices) - see Google Cloud pricing | High-quality multilingual speech synthesis |
| Supabase (optional) | Free tier (up to 500 MB storage) - verify on supabase.com | Persistent storage for user sessions and analytics |
Estimated build time: 8-12 hours, depending on familiarity with each platform.
Below is a step-by-step walk-through that you can follow line-for-line. Every setting, field, and API endpoint is spelled out so you can copy-paste where possible.
Vapi supports over 30 languages with a single endpoint, removing the need for separate language-specific pipelines.
read_products
, read_inventory
, write_orders
. For the full spec see ** Shopify developer docs**.
We'll use n8n to glue Vapi, OpenAI, Google Cloud, and Shopify together.
POST
and copy the generated URL - this will be the callback URL you register in Vapi (see step 5). transcript
and language
fields from Vapi's request payload.
{
"nodes": [
{
"parameters": {
"value": "{{$json[\"speech\"][\"transcript\"]}}",
"type": "string"
},
"name": "ExtractIntent",
"type": "n8n-nodes-base.set"
}
]
}
What this does: isolates the raw spoken text and the detected language so downstream nodes can work with a clean payload.
prompt
to a template like:
You are a helpful voice sales assistant for a Shopify store. Answer the customer's question in {{ $json["language"] }} and keep the tone friendly and concise.
Set Model to gpt-4
.
https://{{store}}.myshopify.com/admin/api/2023-10/graphql.json
). Use the Authorization: Bearer
header. The GraphQL query can be:
{
products(first: 5, query: "{{ $json["userQuery"] }}") {
edges {
node {
title
variants(first: 1) {
edges {
node {
price
}
}
}
}
}
}
}
What this does: pulls the top-matching products for the user's request, which the OpenAI node can then reference when building its answer.
Connect the Webhook node to Vapi: In Vapi's dashboard, under Application → Webhooks add a new Call Ended webhook and paste the n8n webhook URL. Choose POST and set Content-type to application/json
.
Add a Google Cloud TTS HTTP Request node named SynthesizeSpeech. Use the languageCode
from Vapi's payload (e.g., es-ES
for Spanish) and the text from OpenAI's response. Example request body:
{
"input": { "text": "{{ $json[\"choices\"][0][\"message\"][\"content\"] }}" },
"voice": { "languageCode": "{{ $json[\"language\"] }}", "ssmlGender": "NEUTRAL" },
"audioConfig": { "audioEncoding": "MP3" }
}
Remember to set the Authorization header to Bearer {{YOUR_GOOGLE_CLOUD_ACCESS_TOKEN}}
.
Press Activate in n8n. The workflow will now listen for incoming calls, process the spoken request, query Shopify, generate a multilingual answer via OpenAI, synthesize it, and play it back.
If anything fails, the Execution Log in n8n will show which node errored and the exact response payloads.
Add a PostgreSQL node after ExtractIntent to store call_id
, language
, and last_intent
. This lets you implement multi-turn dialogues (e.g., "Add that to my cart") without losing context between webhook calls.
Set up alerts in n8n or your cloud provider to avoid surprise bills.
| Failure mode | Symptom | Fix |
|---|---|---|
| Vapi language detection mismatch | ||
| Agent replies in the wrong language | Verify the language codes sent by Vapi; override by forcing a known code in the Set node if needed. | |
| Shopify API rate limit (40 req/s per shop) | ||
| HTTP 429 Too Many Requests in FetchProduct node | ||
| Implement an n8n Delay node (e.g., 200 ms) before each Shopify request, or cache frequent queries in Supabase. | ||
| OpenAI token quota exceeded | ||
| 429 Too Many Requests from OpenAI node | ||
| Upgrade the OpenAI billing plan or add a Retry node with exponential back-off. | ||
| Google Cloud TTS auth expiration | ||
| 401 Unauthorized in SynthesizeSpeech | ||
Use a service account key and rotate the token every hour with a Cron workflow that refreshes YOUR_GOOGLE_CLOUD_ACCESS_TOKEN . |
||
| n8n webhook unreachable (e.g., public URL not reachable) | ||
| Vapi logs "Webhook delivery failed" | Ensure the n8n instance has a stable HTTPS endpoint (use n8n.cloud or expose self-hosted via ngrok for testing). | |
| Cost blow-up on high call volume | ||
| Unexpected spikes in Vapi or OpenAI bills | Set a daily spend limit in Vapi, and add a Function node that checks a Supabase-stored budget_remaining flag before proceeding with expensive calls. |
A single Vapi-Shopify integration can handle at least 5 concurrent calls on a modest cloud VM without degradation, provided you respect Shopify's 40 req/s limit.
Add the ISO language code (e.g., de-DE
for German) in Vapi's Multilingual settings, then ensure the Google Cloud TTS request uses the same languageCode
. No code changes are required unless you use language-specific prompts.
Yes. Use the n8n Community Edition Docker image, host your own PostgreSQL for Supabase-compatible storage, and run a small VM for the webhook. Vapi and Google Cloud remain SaaS components, but you can replace them with open-source alternatives if you need a fully on-prem solution.
Typical end-to-end latency is 1.2-1.8 seconds: ~300 ms for STT, ~500 ms for OpenAI processing, ~300 ms for Shopify GraphQL, and ~200 ms for TTS synthesis. Optimize by caching product data and re-using the OpenAI session token.
No. Vapi's voice gateway can detect the caller's language on a single inbound number, then route the request to the same workflow which dynamically selects the appropriate TTS voice.
Visit the Shopify GraphQL Explorer or the Shopify Developers page for sample queries. The snippet in step 4 works for most "search-by-keyword" use cases.
Store all secret keys in n8n's Credentials store, not in workflow JSON. Mark them as Encrypted and restrict access to the n8n UI via SSO or IP allow-listing.
If you're looking for more ready-made automation ideas, check out our guide on ** AI automations you can sell** and grab