On May 13, 2026, at 9:47 AM, a Sherbrooke accounting firm noticed something strange. Their AI voice agent, deployed six months earlier, was still picking up calls — but transfers to the accountants stopped happening. No alert. No 500 error. Just clients ending up in voicemail after two minutes of conversation.
The cause? One line in OpenAI's changelog: Realtime API Beta deprecated and removed — May 12, 2026. Their AI voice agent vendor had built the platform on the beta interface. Nobody warned them.
If you're a Quebec SMB with an AI voice agent in production, you now have a window of a few weeks before your clients start silently hanging up. This is not an exaggeration. This is what happens when an API nobody watches changes shape overnight.
What happened on May 12, 2026 (and why nobody is talking about it)
Two things happened on the same day, and it's their combination that creates the problem.
First, OpenAI permanently killed its Realtime API Beta. The beta interface is no longer wire-compatible with GA (General Availability). Events were renamed — response.text.delta became response.output_text.delta, conversation.item.created became conversation.item.added. Ephemeral key generation moved to POST /v1/realtime/client_secrets. If your vendor wrote their code in 2025 and didn't migrate, the agent doesn't crash with a big red error. It fails silently, losing its tools one by one.
Second, ElevenLabs released SDK v2.47.0. This update adds IP allowlisting for service account API keys, new mandatory webhook request headers, and voice metadata moderation. Good things — except that agents built before May 12 often use webhook integrations that no longer respect the new format.
Result: your AI voice agent in Quebec may be running code that shouldn't run anymore. And the worst part is, it won't tell you.
Why "silent failure" is the worst-case scenario
An agent that crashes is easy to detect. The phone rings, nobody answers, your clients text "your AI doesn't work," you call your vendor. Stress, yes — but fixable in a few hours.
Silent failure is something else. The agent picks up. It talks. It pretends everything is fine. But:
- Transfers to a human never happen
- Appointments are booked into the void (never written to the CRM)
- Confirmation SMS never go out
- Technical questions get made-up answers (because the RAG tool is broken)
You discover the problem three weeks later, when a furious client calls because their appointment was never logged. By then, you've probably lost 20 to 40 contracts. This is exactly the scenario the anti-hallucination protocol article hinted at — except this time, the cause isn't the LLM. It's the infrastructure.
The 7-point audit (to run this week)
Before I describe the seven points, one note: if you work with a serious vendor, they should have already run this audit for you, without you asking. At TECHMA, we spent the weekend of May 9-10 migrating all our clients' agents to OpenAI's GA interface and updating webhooks for the new ElevenLabs format. That's our job — not yours. But if you didn't receive a recap email from your provider, that's a signal.
Point 1 — Call your own agent and ask for a transfer
The simplest test. Dial the number, explicitly ask "I want to speak to a human." The agent must transfer within 5 seconds of your request. If the call loops, or you end up on a generic call-end message, your transfer function is broken.
This is the first symptom of the broken Realtime beta API: tools (function calls) no longer trigger correctly because events were renamed.
Point 2 — Check your webhooks
Log into your vendor's dashboard (or ask for a screenshot). Look for the "webhooks" or "outbound calls" section. Check the "last successful" column.
If you see a date before May 12, 2026 on a webhook that should run continuously (like CRM saves or confirmation SMS), it's broken. ElevenLabs' new mandatory headers block old integrations that haven't been updated.
Point 3 — Read the latency of the last 100 calls
Latency is the first indicator that an API isn't responding the way it used to. If your agent replied in 600 ms before May 12 and now replies in 1,200 ms, that's not a coincidence — it's probably because it's doing silent retries on API calls that fail on the first attempt.
The 700 ms threshold we documented in April still holds: above it, you lose on average 18% of calls to abandonment. If your agent crosses that threshold this week when it was under before, you have an infrastructure problem, not a prompt problem.
Point 4 — Audit conversation tags
ElevenLabs' conversation tags, also released May 12, let you see which calls triggered which tools. Filter on calls from the last 7 days. For each critical tool (transfer, booking, SMS), count the ratio of "calls that should have triggered this tool" vs "calls where it actually triggered."
If the ratio dropped below 80% this week, you have a problem.
Point 5 — Test conversational memory
Call your agent. Give it your name at the start. Ask three questions. Then ask it to repeat your name. If it forgets, the context system is probably affected by the conversation.item.added event change.
It's subtle but critical: an agent that forgets a client's name mid-conversation loses all credibility, and therefore loses the sale.
Point 6 — Verify Law 25 compliance
May 12 also changed how ElevenLabs handles voice metadata moderation. If your vendor hasn't activated this feature and you're in Quebec, you're unknowingly accumulating voice metadata (gender, accent, emotional state) without explicit consent — exactly the kind of thing that can trigger a CAI complaint.
Ask your vendor to show you the "voice metadata moderation" configuration enabled. If they don't know what you're talking about, you have your answer.
Point 7 — Prepare a rollback
If the audit reveals problems, don't panic. ElevenLabs versioning lets you revert to an earlier version of the agent in less than five minutes. But that assumes your vendor kept snapshots. Ask them: "can you restore the version from before May 12 if I ask?"
If the answer is no, you have a governance problem that goes beyond this update.
The trap few SMBs see coming
What makes this situation dangerous is that most Quebec SMBs will never know their agent is broken. Why? Because the symptoms — missed calls, failed transfers, lost appointments — look like normal operational noise. "Eh, we had a few no-shows this week." "The client probably didn't get the SMS." "Our AI is a bit slow today."
Meanwhile, your competitor with a serious vendor — one that did their migration the weekend before May 12 — is capturing your clients one by one. As this Medium article explains, May 12 collapsed the voice agent stack: vendors who migrated will soar, those who didn't will disappear in silence, and their clients will disappear with them.
What TECHMA does for its clients (and what you should demand)
When we deploy an AI voice agent for a client in Montreal, Quebec City, or the Eastern Townships, we take care of the entire technical update. It's in the contract. When OpenAI or ElevenLabs changes something at 3 AM on a Sunday, we handle it before the client even notices. That's exactly what we did for the May 12 migration.
If your current provider doesn't operate like that — if you're the one watching changelogs, approving technical updates you don't understand, "validating" infrastructure refactors — you're paying for a tool, not a service.
The 7-point audit above, you can run yourself. But in theory, you shouldn't have to. That's the difference between buying a platform and buying a complete solution.
The good news
If you run the audit and everything passes, enjoy it: your vendor did their job, and you're now ahead of 60% of Quebec SMBs who have a quietly broken AI voice agent without knowing it.
If the audit fails on one or more points, the next step isn't to panic. It's to ask your vendor for a written remediation plan with a deadline. If you don't get one within 48 hours, it's time to shop around.
May 12, 2026 is a tipping point for the AI voice agent industry in Quebec. The SMBs that run the audit this week will know if they're on the right side. The others will find out three months from now, when they read their revenue statements.
One last thing about timing
The Quebec tourist season opens in late May and runs through October. Restaurants, marinas, B&Bs, golf courses, and outfitters all see a 3x to 5x spike in inbound calls during this window. If your AI voice agent is silently broken when that wave hits, you won't just lose contracts — you'll lose the entire season. There is no "we'll catch up in Q3" when half your annual revenue happens in 90 days.
That's the real urgency. Not "did the vendor migrate on May 12." But "will you find out before or after July 1." Run the audit this week. Even if everything passes, you'll sleep better in June.