Every sales call for a tool like this arrives with the same stack of terms — NLU, latency, intent, resolution rate, LLM — most of it borrowed straight from an engineering team's internal vocabulary. Some of it matters to a buying decision. Most of it doesn't, once you know what it actually means. Here's a plain pass through the terms that come up most, from someone who works on this for a living.
The words about what the AI hears and says
ASR (automatic speech recognition). The part that turns a caller's spoken words into text the system can act on. If this is weak, everything downstream is answering the wrong question — worth testing with a real accent and a noisy background, not a quiet demo room.
TTS (text-to-speech). The part that turns the system's reply back into a voice. Ask how natural it sounds mid-conversation, not just in a single sentence — the tell is usually in how it handles pauses and interruptions, not the first "hello."
Latency / response time. The gap between a caller finishing a sentence and the reply starting. This is one of the few terms worth actually asking a number for — AIVA runs at roughly 198ms on voice, fast enough that it doesn't feel like a pause at all. A system that takes a second or two to respond will feel like it's "thinking," even when the eventual answer is right. We've written more on why this specific gap matters more than most owners initially assume.
Barge-in / interruption handling. The ability to keep up when a caller talks over the system mid-sentence, which people do constantly in real conversation — finishing your sentence for you, cutting in with a correction, or just starting to answer before you're done asking. A system without this either talks over the caller or gets confused when interrupted. It's worth testing directly: interrupt it mid-answer on a live call and see whether it adapts or restarts from wherever it left off.
The words about what the AI understands
NLU / NLP (natural language understanding / processing). The general term for "makes sense of ordinary sentences" rather than matching fixed keywords or menu options. You don't need to interrogate this one deeply — what it buys you, practically, is that a customer can phrase a question however they naturally would instead of hunting for the phrase the system expects.
Intent. What the customer is actually trying to do, underneath the specific words they used. "Do you have anything free tomorrow?" and "can I still get in this week?" are different sentences with the same booking intent. A system that only matches sentences, not intent, treats those as two unrelated questions.
LLM (large language model). The layer doing the actual reasoning — the reason a system can handle phrasing it's never seen verbatim, instead of failing outside a fixed script. Not something you need to evaluate as a spec sheet item; what matters is how it behaves on a real, slightly unusual question, not what it's called.
Multilingual / code-switching. Handling more than one language, including within the same sentence — a caller opening in Hindi and finishing a question in English, for instance. AIVA handles this natively across 12 Indian languages rather than translating into English and back, which is what typically adds delay and loses tone in systems that bolt on language support after the fact. See the full language list if this matters for your customers.
Bot vs. agent. Used loosely, almost interchangeably, in a lot of vendor marketing — but worth distinguishing if you're comparing options. A basic bot matches fixed keywords or menu choices to scripted replies and breaks the moment a question falls outside that script. An agent, used accurately, understands open-ended phrasing and can take real action — checking a calendar, applying a specific policy, completing a booking — rather than just retrieving a pre-written answer. We've written a fuller breakdown of this specific distinction, because the label a vendor uses matters less than what actually happens when you ask it something the demo script didn't cover.
Jargon describes how something is built. It doesn't tell you whether it works for your business — only trying it does.
The words about how well it's working
Resolution rate. The share of conversations where the customer's need was actually met without a human stepping in — not just "the system said something." We've written a full breakdown of what this number does and doesn't tell you, because it's more nuanced than a single percentage suggests.
Escalation. Handing a conversation to a human. Not a failure state — a well-run system escalates on purpose for specific situations, like an upset customer or anything involving account security, and we've written about how we think about that boundary in more detail.
Repeat contact. A customer reaching out again shortly after a conversation that was marked resolved, often about the same underlying issue. This is one of the better signals that a "resolved" conversation didn't actually fully answer what the customer needed, even though it technically closed out clean.
Analytics / dashboard. Where all of the above actually lives — a real-time view of conversation volume, resolution rate, and an auto-categorized breakdown of what customers are asking about, typically with a retention window (90 days, in AIVA's case) so you can look back at a specific stretch, and export the data if you want to review it elsewhere. This is usually the fastest way to make any of the terms above concrete — rather than asking a vendor to define "intent," ask to see an actual intent breakdown from a real account.
Feature phone / smartphone-agnostic. Working on a basic handset with no app, no data plan, and no smartphone required — relevant mainly to SMS, which sends and receives on essentially any phone. Worth asking about directly if a meaningful share of your customers are in areas where smartphone and data access aren't a given.
The words that describe what you're replacing
IVR (interactive voice response). The "press 1 for appointments, press 2 for billing" phone tree most callers already know on sight. It routes a caller somewhere based on a button press, but it doesn't understand anything the caller says — ask it an actual question and it has nowhere to send you except back to the menu, or eventually, a person. Worth naming plainly: an IVR and a conversational AI agent solve different problems, even though both often greet a caller first.
Auto-attendant. A step simpler than IVR — call routing based on extension or department, common in office phone systems, with no attempt at understanding language at all. If a vendor's "AI receptionist" turns out to be an auto-attendant with a friendlier voice recording, that's worth catching before you sign up, not after.
Virtual receptionist / answering service. A human-staffed remote reception service, not software — a real person, usually working across several client accounts at once, answers on your business's behalf from a call center rather than a desk in your building. It has the same one-agent-one-call concurrency limit as any human-staffed setup, which is the main practical difference worth understanding between that model and an AI agent: a person's judgment on every call, but no ability to answer call ten and call one at the same moment.
Call center outsourcing. Routing your calls to a third-party team, in-house or overseas, rather than handling them yourself. It solves the "someone answers" problem but usually not the consistency or concurrency problem, since it's still one human per call underneath the outsourcing arrangement — worth comparing directly against what an AI agent actually changes about that shape.
The words about setup and control
Configuration vs. training. What you do — adding your hours, prices, FAQs, and booking rules — is configuration, not training a model from scratch. You're not teaching it language; you're telling it about your business. This matters because configuration is something you can change yourself in minutes, not something that requires an engineering team.
Omnichannel. The same underlying system and the same answers across voice, web chat, and SMS, instead of three separate tools that might disagree with each other. Worth checking directly: ask the same question on two channels and see if the answers match.
Regional inference / data residency. Where the actual processing for a conversation happens, geographically — AIVA routes calls through regional infrastructure, Mumbai for Indian traffic, Frankfurt or Virginia for EU and US, specifically to keep response times low and consistent under real load. Worth asking any vendor where processing actually happens, not just where their office is.
Usage-based / pay-as-you-go pricing. Paying per minute, per conversation, or per message instead of a flat monthly seat fee — AIVA runs ₹4 a minute for voice, ₹2 per chat conversation, ₹1 per SMS, with no monthly charge. Worth comparing directly against what a fixed-fee alternative would actually cost at your real call volume, since usage-based pricing can be cheaper or more expensive depending on how much volume you actually have.
SLA / uptime. A vendor's formal commitment to how often the system is actually available, usually expressed as a percentage like 99.9%. Worth asking about directly if a phone line going down for even a few minutes would be costly for your business, rather than assuming any two vendors' uptime is roughly the same.
Fallback. What happens when something genuinely breaks — a connectivity issue, a system outage — rather than a normal escalation to a human. A well-designed fallback still gets the caller to a voicemail or an alternate number rather than a dead line, which is a different, rarer case than the everyday escalation rules covered above.
The words about what you connect it to
Webhook. A way for AIVA to notify another system the moment something happens — a new booking, a completed call — without anyone checking a dashboard for it. It's the plumbing behind a booking landing in a calendar or CRM automatically, instead of someone re-typing it there by hand.
API key (external-tool). Not the same thing as a public developer API — AIVA doesn't have one of those yet, and we've explained why separately. An external-tool API key is narrower: credentials that let AIVA connect out to a specific system you already use, like a calendar or CRM, to check availability or write a booking, without exposing a general-purpose API for building custom software on top of AIVA itself.
Integration. Connecting AIVA to a tool a business already runs — a calendar, a booking system, a CRM — rather than replacing that tool. The distinction matters because "integration" implies your existing systems stay the system of record; AIVA reads from and writes to them, it doesn't ask you to migrate your booking data somewhere new.
CRM (customer relationship management). The system that stores customer and contact history — Zoho, HubSpot, and Salesforce are common examples. AIVA can connect to one through an external-tool API key so a booking or call summary lands there automatically, but the CRM itself stays a separate tool, not something AIVA replaces.
Embeddable widget. The web chat piece specifically — a snippet added to an existing website, matching its look, rather than a separate page customers get redirected to. "Embeddable" is doing real work in that sentence: it means no separate login and no separate app, just a chat box that's part of the page a visitor is already on.
The one question worth asking regardless of the term
Whatever term comes up in a sales conversation, the question that actually matters is the same: does this specific vendor's version of it hold up on a real call, with your actual FAQs, not a rehearsed demo. That's also the fastest way to tell a genuine conversational AI agent from a scripted bot wearing the same label — the terminology sounds identical in a pitch; the behavior on a real, slightly unusual question doesn't. Start free and find out on your own questions, not ours.