A Quebec SMB owner recently told us about a conversation with his AI voice agent vendor. The vendor kept dropping barge-in, functions, on-device deployment, and deflection. The owner nodded along, scribbled three words at random, and signed anyway. Three months later, the agent was cutting customers off mid-sentence and flatly refusing to switch to French. The catch? The contract guaranteed things nobody had really understood.
The vocabulary around AI voice agents in 2026 has become a minefield of English jargon, telecom acronyms inherited from the 1970s, and marketing terms invented last month. For a Quebec SMB that has to juggle Loi 25, two official languages, and rising customer expectations, signing a contract without mastering that language is buying blindfolded.
This glossary covers 20 terms grouped into four families: product architecture, telephony plumbing, conversational intelligence, and operations-compliance. By the end, you'll be able to push back on any sales rep, read a spec sheet without tuning out, and ask the right questions before you sign. We calibrated the definitions for the Quebec context — because most glossaries online skip everything about Loi 25, French-English code-switching, and latency on Canadian SIP networks.
Product architecture: the 5 building blocks
1. AI Voice Agent
Software system that answers the phone, holds a conversation, and executes actions (booking an appointment, transferring, sending a text), all in a few hundred milliseconds. The important distinction: an AI voice agent is not a scripted phone tree (IVR). It interprets natural language, changes subject, handles ambiguity. In 2026, serious agents sit on top of a multimodal language model (see term #2) wired to a speech engine (#3).
2. Language Model (LLM)
The textual "brain" of the agent. April 2026 marked a turning point: OpenAI released gpt-realtime as generally available — a speech-to-speech model that processes audio directly without an intermediate transcription step. Concrete result: lower latency, better intonation handling, the ability to repeat a confirmation number letter by letter. Always ask your vendor which LLM underpins the agent, and whether it can switch between models depending on the conversation stage.
3. TTS and STT
TTS (Text-to-Speech) converts text into synthetic voice. STT (Speech-to-Text) transcribes the caller's speech. Until 2024, the two were separate steps each adding 300-500 ms of latency. Modern speech-to-speech models merge both, dramatically cutting delays. For Quebec, verify that the TTS correctly pronounces place names (Longueuil, Rimouski, Trois-Rivières) and that the STT distinguishes a Saguenay accent from a Parisian accent without choking.
4. End-to-End Latency
Time between the caller finishing a sentence and the agent starting its reply. Above 1,200 ms, callers cut off and hang up. Under 500 ms, the conversation feels natural. The current target for serious deployment sits around 800 ms or less, which we unpacked in detail in this analysis on latency and why 800 ms decides customer retention. Demand a number measured on your network, not in the vendor's datacenter.
5. Barge-in (Interruption)
Ability for the caller to cut the agent off. Without barge-in, the agent finishes its paragraph while the caller shouts "wait, no, I meant Saturday!". With well-tuned barge-in, the agent stops in 150-200 ms, listens again, and picks up. An agent without functional barge-in sounds robotic and grating. An agent with over-sensitive barge-in gets interrupted by a sigh. Tuning it right is a science.
Telephony plumbing: the 5 terms the telcos don't explain
6. SIP (Session Initiation Protocol)
Standard IP telephony protocol. Your current business line almost certainly runs on SIP under the hood, even if you see a classic phone number at the end. AI voice agents connect to your line via a SIP trunk. The TECHMA team handles the link between your telecom provider (Bell, Videotron, TELUS, etc.) and the agent platform. If a vendor tells you "configure your SIP yourself," run.
7. DTMF (Dual-Tone Multi-Frequency)
These are the "beep-boop" tones you hear when someone presses a phone keypad. In 2026, serious AI voice agents send and receive DTMF tones to navigate existing IVR menus (banks, utilities, call centers). Useful when your agent has to dial an external number and choose "press 2 for billing."
8. IVR (Interactive Voice Response)
The classic phone tree: "press 1 for service, 2 for billing." IVRs are on their way out but often coexist with modern AI voice agents in hybrid architectures. A SMB moving from a 12-branch IVR to a conversational agent typically cuts abandon rate by 30-50%. That remains a solid reference point for measuring your return on investment.
9. Warm vs Cold Transfer
Cold transfer: the agent puts the caller on hold and bounces them to a human with no context. The human says "hello, how can I help?" and the caller repeats everything. Warm transfer: the agent drafts a 3-line summary (intent, case number, detected emotion), pushes it into your CRM window, and passes the caller. The human opens the record already knowing everything. The difference in customer satisfaction: huge.
10. Twilio, Vonage, Telnyx
Communications infrastructure providers (CPaaS). They're the ones supplying numbers, minutes, and SIP connectivity under the hood. Your AI voice agent runs at a vendor (ElevenLabs, OpenAI, VAPI, etc.) but calls flow through Twilio or equivalent for the actual telephony. It doesn't show up on your main invoice, but it's a hidden cost that can eat 20-30% of the budget if poorly negotiated.
Conversational intelligence: how the agent really thinks
11. Function (Function Calling)
The agent's ability to call external code mid-conversation. Example: the caller asks "do you have availability Wednesday at 2 pm?", the agent calls check_calendar(date, time), receives the answer from the booking system, and replies to the caller. Without functions, an agent only talks. With functions, it acts.
12. Knowledge Base
Internal documents (pricing PDFs, procedures, FAQ) that the agent consults before answering. Technically, it's a vector index queried at every question. If your vendor asks you to put your prices in a shared Google Sheet rather than a proper knowledge base, that's a red flag: updates will be painful and Loi 25 security compromised.
13. Bilingual Code-switching
The agent's ability to change language mid-sentence when the caller does. A typical Quebecer might say "yeah but le cost c'est quoi?" and expect the agent to understand without breaking. Poorly configured agents lock into the first language detected and reply in English for the rest of the call. For technical depth, see our dedicated analysis on French-English code-switching in Quebec.
14. Hallucination
Moment when the agent invents a plausible but false piece of information: a price that doesn't exist, a phantom discount, wrong opening hours. In medical or legal contexts, a hallucination can cost dearly. Guardrails (#15) and a well-maintained knowledge base push the risk below 1%, but full elimination remains a moving target.
15. Guardrail
Explicit rule that keeps the agent from going off the rails. Examples: "never provide a medical diagnosis", "never confirm a price without checking the knowledge base", "never promise a refund." A good agent has between 15 and 30. A bad agent has three and relies on the model's "common sense" — which works 98% of the time, except when it matters.
Operations and compliance: what separates a prototype from production
16. Deflection Rate
Percentage of calls fully handled by the agent without being transferred to a human. The headline metric. A 70% deflection rate means that out of 100 calls, 70 are resolved without human intervention. Watch out for vendors who count a "transferred without handling" as a deflection: that's accounting fraud. We lay out the honest measurement protocol in our guide to measuring real ROI in 14 days.
17. First Call Resolution (FCR)
Percentage of problems resolved on the first call, no callback needed. Different from deflection rate. A caller can be "deflected" (handled without a human) but call back the next day because the problem wasn't actually solved. A good AI voice agent FCR hovers around 75-85%.
18. Recording & Retention
Virtually all agents record conversations for analysis and training. The question isn't "does it record?" but "how long, where, who can access it, and was the caller informed?" In Quebec, default retention should be as short as possible — 30 to 90 days maximum for most use cases.
19. Loi 25 Consent
Quebec's personal information protection law, fully in force since 2023-2024. An AI voice agent that collects a phone number, a name, or a reason for calling is processing personal information. The business must obtain valid consent (clear, free, informed) and appoint a privacy officer. We cover the full ground in our Loi 25 guide for Quebec SMBs.
20. On-Premise / On-Device Deployment
The term of the year in 2026. ElevenLabs announced on April 9, 2026 the ability to deploy their engine on-site or directly on a device, without sending audio through a third-party cloud. For medical offices, notaries, psychology clinics, or any business handling highly sensitive data, it's a tipping point: Loi 25 compliance becomes trivial when no data leaves the building. Per-call unit cost stays higher than cloud (+20-40% typically), but the risk-cost math changes completely.
Three vocabulary traps we see every single week
"Artificial intelligence" ≠ "AI Voice Agent." Plenty of vendors sell an "AI solution" that turns out to be a scripted chatbot with a synthetic voice on top. Ask for the underlying model and the ability to hold a five-turn conversation with subject changes.
"99% accuracy" is an empty number. Accuracy on what? STT? Intent classification? Correct answer? Without the metric specified, it's marketing fluff. An agent at 99% STT accuracy with 60% accuracy on business intents is… mediocre.
"Deployed in 24 hours." Possible for a demo. Impossible for a production deployment with CRM integration, bilingual tests, customized guardrails, and latency calibration on your real telecom network. A serious deployment takes 2 to 6 weeks, handled by a team that knows your systems — that's the work TECHMA performs for every client.
Frequently Asked Questions
Do I need to master these 20 terms to deploy an AI voice agent?
No. The TECHMA team handles implementation, integrations, compliance, and monitoring. The glossary mainly helps you follow scoping conversations, validate monthly reports, and ask sharper questions during quarterly reviews.
Which term will cost me the most if I ignore it?
"End-to-end latency" if you're comparing vendors, "Loi 25 Consent" if you're in a regulated sector, and "Deflection Rate" to measure return on investment honestly. The three form a triangle every leader should watch.
Does the vocabulary change fast?
Yes. Between 2024 and 2026, terms like "speech-to-speech", "on-device", and "multi-agent" went from academic jargon to industry standard. Expect another semantic layer by 2027. The good news: the base structure (LLM, latency, deflection, compliance) stays stable.
What do I do if a vendor uses a term I can't find here?
Ask them to define it in one sentence, then two paragraphs, then with a concrete example from an existing deployment. If the three levels of definition don't line up, the vendor is improvising. That's a useful signal.
The next step
This glossary isn't an end in itself. It's a lens that makes contracts, spec sheets, and vendor conversations much more readable. If you want to move from "I understand the vocabulary" to "I have a deployed agent that works," the next logical read is our post-mortem on why 70% of SMBs unplug their agent within 90 days — and how to dodge that trap. For the raw numbers on what the Quebec market is actually seeing in 2026, OpenAI's official gpt-realtime announcement and the ElevenLabs changelog are the sources of truth to follow monthly.