Tuesday morning, a dentist in Sherbrooke calls me, panicked. A patient claims they were misquoted on the cost of a crown during a call handled by his AI voice agent three weeks earlier. The patient is threatening to file a complaint with Quebec's Commission d'accès à l'information. The dentist opens his dashboard. The call did happen. But the transcript? Doesn't exist. The summary generated after the fact by the LLM? Vague, readable, but missing the exact words. Evidence: gone.
Meanwhile in San Francisco, OpenAI launched on May 7, 2026 a model that would have prevented this nightmare: GPT-Realtime-Whisper, live word-by-word voice transcription priced at $0.017 per minute. One dollar per hour of call. At that price, choosing not to transcribe every conversation handled by your AI voice agent becomes an active decision. And in Quebec, where Law 25 demands clear traceability of any processing of personal information, that's a hard decision to defend.
Yet half the AI voice agent vendors I see in demos this week are still selling "post-call summaries" at $4 or $5 per call as if it were a premium service. It isn't anymore. Here's why.
What OpenAI actually launched (and what changes)
On May 7, 2026, OpenAI released three new realtime voice models through its Realtime API. The trio is easy to confuse, so let's break it down:
- GPT-Realtime-2 — the conversational brain. GPT-5-class reasoning, 128,000-token context window, lower latency. Pricing: $32 per million input audio tokens, $64 per million output tokens. This is the engine that talks with your customer.
- GPT-Realtime-Translate — simultaneous translation. 70 input languages, 13 output languages. Priced at $0.034 per minute. Useful for Montreal hotels and restaurants — I covered this recently in this analysis on deploying multilingual agents in 8 languages.
- GPT-Realtime-Whisper — streaming transcription. This is the one that changes the compliance equation in Quebec. Official documentation: developers.openai.com.
What GPT-Realtime-Whisper does technically: it listens to your phone conversation in parallel with the agent that's speaking, and produces a word-by-word transcript with sub-second delay. Not a summary. Not a paraphrase. The exact words your customer said, at the precise timestamp they were said, in the language they were spoken — Quebec French included, with accuracy that took a noticeable jump from Whisper v3.
Why $0.017 changes the Law 25 conversation
Let's do the math. An average AI voice agent for a Quebec SMB handles between 800 and 2,500 minutes of calls per month. At $0.017 per minute, full continuous transcription of every call costs between $14 and $43 per month. Less than a daily coffee.
Before this launch, vendors had a defensible economic argument: "transcribing every call costs too much, we'll generate a summary via the LLM instead." At $0.40 or $0.60 per minute (the old real-time transcription rates from Deepgram or AssemblyAI on phone audio), I got it. At $0.017, that argument collapses.
And this is exactly where Quebec's Law 25 becomes a practical issue, not a theoretical one. Since September 2024, all provisions are in force. The Commission d'accès à l'information can require a business to demonstrate how it processes, retains and documents personal information collected during a call. Three key obligations apply directly to AI voice agents:
- Incident registry. If a customer complains that confidential information (health insurance number, medical history, payment data) was mishandled during a call, you have 72 hours to notify — and you need the exact record of the exchange.
- Privacy Impact Assessment (PIA). Any project involving automated processing must be documented. A timestamped textual transcript is the foundation of a solid PIA.
- Right of access and rectification. If a customer asks "what did you note during our call on May 3rd?", a vague summary is not an answer. A word-for-word transcript is.
I covered the full compliance audit in this 6-point audit of Law 25 and PIPEDA-2026-002. Real-time transcription is now a practical requirement, not a luxury.
5 scenarios where transcription at $0.017 actually changes the math
1. The client-vendor dispute (the Sherbrooke dentist scenario)
A customer disputes a price announced during the call. Without word-for-word transcription, it's their word against your LLM summary — which, by the way, hallucinates 3 to 5% of the time. With timestamped live transcription, you open the file, filter by phone number, find the exact passage: "porcelain crowns start at $1,250 before taxes." Dispute resolved in 90 seconds.
2. Banking and insurance compliance (voice KYC)
If your AI voice agent collects identification elements — date of birth, auto insurance number, address — you need to prove the collection happened with explicit consent. Live transcription captures the exact phrase "Yes, I agree to my information being used for..." at the precise timestamp. Far more defensible than a Boolean flag in a CRM.
3. Continuous training and agent improvement
Want to know why your AI voice agent is transferring too many calls to a human? With full transcripts, you can run 500 calls through an analytics tool in minutes, identify the trigger phrases ("I want to speak to someone", "You don't understand", "This is complicated"), and adjust the prompts. Without transcription, you're flying blind.
4. Time-based billing (law firms, accountants, consultants)
A Quebec City lawyer who bills by the minute can now prove, line by line, what was said during a 22-minute call billed to a client. No more debate over "you billed me 22 minutes for 12 minutes of useful conversation." The transcript shows the full 22 minutes, word for word.
5. Automatic CRM logging (the real post-call workflow)
Instead of asking the LLM to summarize after the fact (and risking hallucination), you take the exact textual transcript and run a structured extraction script over it. Phone number, full name, reason for call, requested action, urgency — all extracted from the verbatim and pushed to Zoho, HubSpot or Jobber. Accuracy: around 96-98% based on internal tests we ran at TECHMA this week. Pure LLM summarization was tracking at 88-91%.
The trap vendors won't mention
Not every AI voice agent vendor integrates GPT-Realtime-Whisper the same way. Three traps to know before signing:
Trap 1: post-call transcription is not real-time transcription. Some vendors will tell you "yes, we transcribe calls" — but in reality, they send the audio to a batch service after the call ends, adding 15-30 seconds of delay and killing any real-time use case (live human agent assist, instant compliance alerts).
Trap 2: hidden markup. OpenAI raw cost: $0.017 per minute. Some vendors bill $0.15 or $0.25 per minute for the same thing, calling it "integration value-add". The margin is legitimate, but at some point it becomes opportunism. For 1,500 minutes per month, that's $26 vs $225 — the difference pays two months of subscription.
Trap 3: data retention. How long is the transcript retained? On which servers? Does the OpenAI Data Processing Addendum cover Canadian hosting? If the answer isn't clear and in writing, you have a Law 25 compliance gap that can cost you far more than you think. The average cost of a missed call — which I estimated at $126,000 per SMB per year in this analysis — is almost trivial next to a well-deserved CAI fine.
4 questions to ask your AI voice agent vendor this week
If you already have an AI voice agent in production, or you're in active negotiation, send these four questions by email (in writing, for the paper trail). The answers will tell you immediately whether your vendor is serious or selling you smoke:
- "Are you using GPT-Realtime-Whisper or a streaming equivalent for transcription? If so, since when?" A serious vendor has already switched or plans to switch before end of May 2026. If they're still pitching "LLM post-call summary" as the only product, red flag.
- "What is the exact cost you pay OpenAI for transcription, and what is your margin?" If they refuse to answer, something is hidden. A transparent partner will say "OpenAI bills us $0.017, we bill you $0.05 all-in with storage, search and export."
- "Where are the transcripts stored, for how long, and can I export them at any time?" Expected answer: North American servers with at-rest encryption, configurable retention (30, 90, 365 days, or longer), CSV or JSON export with no fees.
- "Can you provide a transcript sample of a test call in Quebec French with moderate background noise?" This is the transcription version of the 7-call test protocol I describe in this guide on testing an AI voice agent before signing. If the transcript confuses "Sainte-Foy" with "santé fois", you know what's waiting for you in production.
The angle nobody will tell you
The real change isn't the technology. Real-time transcription has existed since 2018 — Deepgram, AssemblyAI, Google Cloud Speech-to-Text have done it for years. The real change is the price. At $0.017 per minute, transcribing every call is no longer a technical or budgetary decision. It's become a legal and operational default.
A Quebec SMB that, in June 2026, deliberately chooses not to live-transcribe every call handled by its AI voice agent will have to justify that choice to an unhappy customer, an employee in training, or — eventually — the CAI. At $14 per month for 800 minutes, the "too expensive" argument is dead.
At TECHMA, we're configuring GPT-Realtime-Whisper on every new AI voice agent we deliver starting this week, at no extra cost to the client. The integration with Zoho, HubSpot, Jobber or Acuity happens in parallel, handled by our team, as part of the standard rollout. You don't configure anything.
If you want to see a demo of an AI voice agent with live transcription in Quebec French — a dental office, a mechanic shop, a restaurant, any use case — reply to this email or book 20 minutes at agentiavocal.ca. We'll show you the transcript appearing live during the call, and we'll let you walk out with the file. It's probably the most convincing demo we can run this month.