How to Build a Real AI BDR With Gemini 3.5 Transcribe Live (2026 Guide)
Gemini 3.5 Transcribe Live enables a real AI BDR: stream calls via WebSocket at sub-second latency, qualify with fixed schema, gate every action with permission rules, and track 9 metrics over 30 days. Build in audit trails before launch.
Gemini 3.5 Makes the Real AI BDR Practical
Many "AI BDR" tools are cold-email machines with better branding.
An AI voice agent for inbound calls listens in real time. It qualifies the caller, routes the lead, books the meeting, and logs each action in the CRM.
Google's Gemini 3.5 Transcribe Live makes this setup more practical. The model streams speech through a WebSocket with sub-second latency and supports more than 85 languages.
Step 1: Treat the Phone Call as Revenue Data
A missed call can be a buyer raising their hand.
Email AI BDRs can produce results. Salesforce reports that SaaStr's Agentforce outreach generated $2.7 million in closed revenue and another $3.5 million in pipeline. The emails had a 72% open rate.
Batteries Plus launched an Agentforce SDR in under one month. It later generated 340 direct meeting requests and more than $15 million in pipeline.
Those are strong numbers. Email still creates a delay between intent and response.
An inbound caller is ready to talk.
Old phone systems treat callers as routing problems. "Press one for sales. Press two for support. Please listen carefully because our menu has changed."
An AI BDR treats the call as a live sales event.
Your first workflow should cover four situations:
1. A salesperson answers the call. 2. Nobody answers within a fixed window. 3. The caller reaches the agent after business hours. 4. The call disconnects before qualification.
For missed calls, trigger an approved callback or text workflow. The model shouldn't decide whether it can call someone.
Put that decision in a fixed policy layer.
Every inbound call should get a status within seconds. The CRM should show whether the call was answered, missed, qualified, booked, routed, or declined.
"AI handled it" isn't enough.
You need a record of what the agent heard, decided, and changed.
Step 2: Stream Audio Outside n8n
Gemini 3.5 Transcribe Live uses the `gemini-3.5-transcribe-live` model for bidirectional streaming.
According to Google's August 26, 2026 announcement, the Live API delivers continuous transcription with sub-second latency. It also supports custom vocabulary and more than 85 languages.
Artificial Analysis measured a 4% average word error rate for streaming audio. 9to5Google reported that time to final transcription improved by 70% over Google's Chirp 3 model.
That speed can change call routing before the caller finishes speaking.
Don't run the live audio session inside n8n.
n8n works well for workflow control. We use it instead of Zapier. A long-running WebSocket audio stream needs a small sidecar service built for live connections.
Use LiveKit, Pipecat, Agora, or a small asynchronous service for the audio stream. Google named LiveKit, Pipecat, Agora, Fishjam, LangChain, Vercel, and Vision Agents as supporting tools.
The data flow should look like this:
Phone audio → streaming sidecar → Gemini Transcribe Live → transcript events → rules engine → n8n → calendar and CRM
Send audio in short chunks. Give each call a unique session ID.
Store partial transcript events for live decisions. Use final transcript events for CRM notes and summaries.
Don't guess at audio formats, rate limits, or quotas from a blog post. The launch material doesn't include enough detail to make those production choices safely.
Google's announcement says the product is in public preview. Public preview differs from general availability, even if some coverage calls the announcement "GA."
Check Google's current API reference before shipping. Pin the model version when Google allows it.
The output should be a transcript event stream with timestamps, call ID, speaker, text, confidence data when available, and event status.
Step 3: Separate Transcription From Qualification
Speech-to-text doesn't qualify a lead.
Gemini 3.5 Transcribe Live should capture the words. A second model or rules engine should decide what those words mean.
That split matters.
If the transcript says, "We have 200 locations and need this next quarter," the qualification layer should extract five fields:
- Company name
- Caller name
- Need
- Timing
- Routing category
You may also need budget, location, product, account status, and preferred meeting time.
Make the qualification model return a fixed schema. Don't accept paragraphs.
A useful output might contain:
| Field | Allowed output |
|---|---|
| Intent | sales, support, billing, partner, unknown |
| Fit | qualified, unqualified, needs_review |
| Urgency | immediate, 30_days, 90_days, unknown |
| Route | sales_queue, support_queue, human_review |
| Next action | transfer, book, callback, collect_more |
| Evidence | Exact transcript quote and timestamp |
The evidence field matters.
The agent shouldn't mark a caller as qualified because the conversation "felt positive." It should cite the words that triggered the decision.
Use custom vocabulary for product names, employee names, SKUs, postal codes, and industry terms. Google says Gemini 3.5 can adapt to provided vocabulary and capture alphanumeric values such as order IDs.
Keep routing rules simple.
- If intent equals support, send it to support.
- If intent equals sales and fit equals qualified, offer approved calendar slots.
- If intent is unknown after two questions, route to a person.
- If confidence falls below your tested threshold, don't take action.
- If the caller asks for a person, stop the AI flow.
The model interprets language. Your rules control the business.
That keeps bad output out of your CRM.
Step 4: Put Permission Gates Before Every Action
A transcript contains sensitive data.
A recording may contain even more sensitive data. Consent rules vary by location, industry, and call type.
Get legal advice for your exact use case. Don't copy another company's disclosure and assume it covers you.
Build permission gates into the workflow:
Before transcription
Tell the caller when AI-assisted transcription or recording is in use, if required. Save the disclosure version, timestamp, and caller response.
If the caller declines, send the call to a no-record path or a person.
Before qualification
Limit the fields the model can extract. Redact payment data, medical data, passwords, and government IDs unless your approved process requires them.
Don't send every spoken word to every downstream tool.
Before routing
Give the agent a closed list of destinations. It shouldn't invent a department, Slack channel, or employee.
If no route fits, use `human_review`.
Before booking
Read back the date, time, time zone, and meeting type. Require clear confirmation before writing the event.
"Tuesday might work" isn't consent to book Tuesday.
Before CRM write-back
Check for duplicate contacts. Validate required fields. Record the proposed change and the final change.
Every action should create an audit event containing:
- Call ID
- Model name and version
- Prompt or policy version
- Transcript evidence
- Decision
- Tool called
- Data sent
- Tool response
- Human approval, if required
- Timestamp
- Error or retry status
Build compliance and auditability before launch. If you can't reconstruct why the agent booked a meeting, you can't safely run it.
You have a voice bot with CRM access and no audit trail.
Step 5: Launch With One Call Path and Hard Metrics
Don't start with every call type.
Start with one narrow path. After-hours inbound sales is a good first use case because the intent is clear and the cost of a missed call is visible.
A sensible first launch does this:
1. Answers the call. 2. Gives the approved disclosure. 3. Captures the caller's need. 4. Checks qualification rules. 5. Offers approved meeting slots. 6. Confirms the booking. 7. Writes the transcript, summary, and evidence to the CRM. 8. Alerts a person when any rule fails.
V1 won't be perfect.
SaaStr's lead agent reportedly reached nearly 1,000 code commits in four months. Jason Lemkin's team made seven or eight commits per day.
AI vendors rarely put that work in the sales deck.
Production agents improve through real conversations, failed tool calls, revised prompts, and better routing rules. A demo shows that a model can talk. Production data shows whether the system works.
Track these numbers for 30 days:
| Metric | Formula |
|---|---|
| Answer rate | Answered inbound calls ÷ total inbound calls |
| Rescue rate | Recovered missed calls ÷ total missed calls |
| Qualification rate | Qualified callers ÷ completed conversations |
| Booking rate | Confirmed meetings ÷ qualified callers |
| Human escalation rate | Calls sent to people ÷ AI-handled calls |
| CRM accuracy | Correct CRM writes ÷ audited writes |
| Permission failure rate | Calls missing valid consent ÷ audited calls |
| Cost per booked meeting | Total system cost ÷ confirmed meetings |
| Revenue per handled call | Closed revenue ÷ handled inbound calls |
Measure show rate and closed revenue separately. A booked meeting that never happens isn't a win.
StoryPros builds AI sales agents that prospect, qualify, and book meetings. Its strongest agent books more than 30 meetings per week.
Voice agents should meet the same standard. They need to create measurable pipeline without damaging trust.
That's the real AI BDR.
FAQ
How do you use Gemini AI to transcribe audio to text?
Use `gemini-3.5-transcribe-live` for streaming calls through a WebSocket. Use `gemini-3.5-transcribe` through the Interactions API for recorded calls, meetings, speaker labels, and word-level timestamps.
How do you convert speech to text in real time?
Stream short audio chunks from the call provider to Gemini 3.5 Transcribe Live. Send partial and final transcript events to a session store, then send final caller turns to your qualification workflow.
Which speech-to-text API is best for inbound sales calls?
Gemini 3.5 Transcribe Live is a strong option for sub-second streaming, custom vocabulary, and support for more than 85 languages. The best API depends on your call audio, accents, terminology, consent rules, and tested error rate.
Can Gemini transcribe Google Meet audio?
Gemini 3.5 Transcribe can process recorded meeting audio through the Interactions API. Google says the API supports speaker attribution and word-level timestamps for recorded files.
Should an AI voice agent write directly to the CRM?
No AI voice agent should have unrestricted CRM access. Put validation, duplicate checks, approved field lists, evidence capture, and human review between the model and systems such as Salesforce.
Related Reading
What metrics should I track for an AI voice agent handling inbound sales calls?
Track nine metrics for 30 days: answer rate, rescue rate, qualification rate, booking rate, human escalation rate, CRM accuracy, permission failure rate, cost per booked meeting, and revenue per handled call. Measure show rate and closed revenue separately from booking rate. A booked meeting that never happens is not a win.
How fast is Gemini 3.5 Transcribe Live for real-time call transcription?
Gemini 3.5 Transcribe Live delivers sub-second latency over a WebSocket connection. Artificial Analysis measured a 4% average word error rate for streaming audio. Time to final transcription improved 70% over Google's previous Chirp 3 model.
Can an AI voice agent book meetings and write to the CRM without human approval?
No AI voice agent should have unrestricted CRM access. Put validation, duplicate checks, approved field lists, and evidence capture between the model and your CRM. Require explicit caller confirmation before writing any calendar event, not just a positive-sounding response.