On a Tuesday at 9:04 AM, a traveler calls a Canadian airline. The voice agent promises a bereavement fare refund after the trip. Months later, the small claims tribunal rules: the airline must honor the promise because "the AI agent is an extension of the company's voice". The Air Canada case isn't an outlier — it's the precedent that now hangs over every Quebec SMB deploying an AI voice agent in 2026.
And the problem is getting worse. According to Suprmind's May 2026 benchmark covering 37 models, hallucination rates range from 15% to 52% depending on the LLM. For voice assistants powered by LLMs in real-world conditions, the median sits at roughly 12% on general knowledge queries. Translate that for your SMB: at 1,000 calls per month, about 120 contain a fabricated answer — a made-up price, a refund policy that doesn't exist, a guarantee nobody at your shop ever authorized.
At TECHMA we've installed this 7-step protocol on every AI voice agent we've deployed in Quebec since March 2026. It draws on Gladia and PolyAI recommendations but adapts them to the legal realities of Quebec's Law 25 and the risk profile of a small business in Montreal, Quebec City, or Sherbrooke. Here's exactly how we harden a voice agent before it ever takes its first call.
Why 2026 changes the risk math
Three things happened in May 2026 that make hallucinations more dangerous, not less:
1. The voices are now indistinguishable. With GPT-Realtime-2 shipping on May 7, 2026 and ElevenLabs v2.47.0 dropping on May 12, the agent now speaks with a cadence, breath pattern, and warmth that customers don't question. A hallucination now reads as an authoritative statement — which is precisely what cost Air Canada the case.
2. The context window is 4x larger. GPT-Realtime-2 jumps to 128,000 tokens (up from 32,000). Great for conversational memory, but it also expands the surface area where the model can interpolate content it never actually saw. More context = more opportunities to fill gaps with fabrication.
3. Courts hold you liable. Quebec civil law tracks the Air Canada precedent. If your AI voice agent promises a price, a delivery date, or a return policy, you pay. The emerging case law is clear: the "it was the AI" defense doesn't hold up.
The TECHMA 7-step protocol
Before diving in: every step must be implemented. Not optional, not "we'll add it later." At our shop, an agent that hasn't cleared all 7 gates doesn't get its Twilio number. Period.
Step 1 — Force the knowledge base as single source of truth
The most common mistake we see at Quebec SMBs: someone wires an LLM to Twilio, points it at "our website, figure it out," and prays. It doesn't hold.
What you actually do: stand up a RAG (Retrieval-Augmented Generation) base that contains only authorized facts — prices, hours, policies, FAQ vetted by management. The system stops consulting its training memory to answer customers. If the fact isn't in the base, the agent says "let me check on that and get back to you." Sounds harsh — it's what saves your case in court.
Concrete tool: we use the ElevenLabs Knowledge Base feature with 512-token chunking, weekly re-indexing, and a Slack channel that notifies us when a source document changes.
Step 2 — Refuse on weak evidence ("Show Sources or Say Sorry")
This principle comes directly from PolyAI's playbook. When the RAG system returns a similarity score below a defined threshold (we calibrate at 0.72 at TECHMA), the agent doesn't fabricate. It pivots to a transfer script: "I'd rather not give you an uncertain figure — let me grab your number and a colleague will call you back within the hour."
Measured result across our May 2026 deployments: -71% fabricated answers on out-of-scope questions, with only a -3% dip in first-call resolution. The trade-off is near-zero. The legal upside is massive.
Step 3 — Bridle the LLM's creativity
Pure mechanics. Three parameters in your agent config:
- Temperature at 0.2 — lower than the 0.7 default. The agent becomes more predictable, less "creative." For a service agent, that's exactly what you want.
- Max tokens per turn at 180 — kills speculative digressions. If the agent wants to say more, it asks a clarifying question.
- Top_p at 0.9 — cuts off the long tail of probabilities where confabulations live.
On an inn deployment in Charlevoix last week, this trio dropped fabricated room rates by 88%.
Step 4 — Second-pass guardrail
Before the synthesized voice goes back to the customer, a second (smaller, faster) model reads the answer and cross-checks it against the RAG base. If it detects a factual claim that isn't supported by a source — it blocks and forces a regenerate. Added latency: 90-120 ms. That's within the 700ms latency threshold we've been holding since May 2026.
Per Gladia's analysis, this defense in depth is the only approach that covers the full failure surface — no single technique is enough.
Step 5 — Freeze high-risk domains
Some topics should never route through the LLM. Here's our blacklist at TECHMA:
- Promotional pricing (always via real-time API to the pricing system)
- Delivery or appointment dates (always via API to the calendar)
- Detailed refund policy (pre-written response, word for word, signed off by management)
- Contractual commitments (mandatory human transfer)
The agent can orient the customer on these topics — it can't commit. That's the difference between "our return policy is generally 30 days; would you like to speak to an advisor to confirm the specifics of your situation?" and "yes, you're entitled to a full refund."
Step 6 — Record, transcribe, audit weekly
All our deployments turn on full-call recording with consent (Law 25 obligation). But recording alone is useless — you need an audit loop.
The TECHMA Friday ritual: we sample 30 calls at random, tag the hallucinations (light, moderate, severe), and feed the detected patterns back into either the RAG base or the guardrail rules. Over 12 weeks, our hallucination rates typically fall from 9% to under 2%. It looks like real hidden cost at first — but it's what turns an AI voice agent from a legal liability into a reliable asset.
Step 7 — The kill switch and the rollback plan
If a serious hallucination surfaces (a promised price you can't honor, a service commitment you don't actually make), you need to be able to cut the agent in under 5 minutes. Every TECHMA deployment ships with:
- A kill-switch on the ElevenLabs dashboard (the agent flips to "we're experiencing high call volume")
- A previous prompt version saved and deployable in one click
- A customer comms plan if an error touched multiple customers (apology template + pre-approved goodwill gesture)
We've only used this kit once in six months of deployments — at a Sherbrooke retailer where a poorly documented new promo generated 8 erroneous promises in 90 minutes. Rollback took 4 minutes. Total cost: $340 in goodwill gestures. Without the plan, the bill would've easily hit five figures.
What Vapi, ElevenLabs and Retell won't tell you
The platforms sell the promise of a "turnkey" agent. None of them will tell you that without these 7 steps, you're buying a legal time bomb. They'll counter with "our model is reliable" — and they're partially right: IrisAgent's benchmarks show 90-95% accuracy on well-defined cases with good knowledge base coverage. But that knowledge base coverage is exactly what doesn't exist at 80% of the SMBs we meet.
The real deployment work isn't the voice — it's the RAG, the guardrails, the thresholds, the audit. That's 60% of the effort, and it's why we charge for a full setup instead of "turnkey $499/month."
What it costs NOT to do this
Let's bring the Air Canada example down to SMB scale. A retailer with 800 inbound calls per month, 9% hallucination rate, 1.5% rate of fabricated commitment-grade promises (most hallucinations are context errors, not promises): 12 fabricated commitments per month. If 30% become customer complaints and the average disputed value is $180, you pay $648/month in losses. Annualized: $7,776. Plus the cost of a complaint at Office de la protection du consommateur or a formal demand letter.
The TECHMA 7-step protocol runs about $4,200 in initial setup and $600/month in maintenance and audit. ROI lands in under 12 months — before you even factor in avoided reputational damage.
FAQ
Can an AI voice agent be 100% hallucination-free?
No. Even with all 7 steps, we target under 2% — not zero. Which is why the rollback plan (step 7) is non-negotiable. Anyone promising zero hallucination is either lying or not really running in production.
Does Quebec Law 25 require disclosing the hallucination rate?
Not directly. But Law 25 requires that you inform the customer they're talking to an automated system and that the information provided is accurate. A hallucination that distorts a customer decision = potential breach of the accuracy obligation. Law 25 compliance and hallucination prevention are legally entangled.
How long does the 7-step protocol take to implement?
On a typical SMB deployment (1 agent, 1 use case), we plan 3 weeks: 1 week for RAG base + policies, 1 week for guardrails + tests, 1 week for progressive rollout (10% of traffic, then 50%, then 100%).
Can we do this without TECHMA?
Technically yes. Practically, we rarely see an SMB without a dedicated AI team get to the result alone. The know-how lives in the precise thresholds (0.72 for similarity, 0.2 for temperature, 180 tokens per turn) — values we calibrated across 40+ deployments since March 2026.
The right time to harden your agent is before its first call
If you already have an AI voice agent in production without these 7 steps — call us before the next incident, not after. If you're still shopping, ask every vendor the question: "how do you handle hallucinations?" If the answer is vague — "our model is state-of-the-art," "we've never had a problem" — keep walking. Legal precision is built, not promised.
At TECHMA we configure this infrastructure for you — RAG, thresholds, audit loop, rollback — so you can focus on what you do best: serving your customers. Book a 20-minute call and we'll review your use case together. May 2026 is also the right window to catch up before summer, when call volumes climb 30-40% at most of our customers.