Quick answer
An intelligent virtual assistant (IVA) is software that understands questions in everyday language and responds with answers or actions. In customer service, a good IVA is judged by whether it knows when not to answer and hands off to a human instead.
What is an intelligent virtual assistant?
An intelligent virtual assistant (IVA) is software that understands questions asked in everyday language and responds with answers or actions. Google Cloud's Dialogflow documentation defines a virtual agent as a conversational agent that "handles concurrent conversations with your end-users"; the "intelligent" part means it interprets intent with natural language processing instead of matching fixed keywords.
The term is deliberately broad. It covers consumer voice assistants like Siri and Alexa, chat widgets on support pages, and AI agents that execute multi-step tasks. That breadth is exactly why the label alone tells you little: in customer service, a good IVA is not the one that answers the most questions — it is the one that knows when it should not answer, and hands the conversation to a human instead. Everything below tests that idea.
IVA, chatbot, AI agent, voice assistant
These four words overlap, and vendors use them interchangeably. A useful split is by what the system does when it does not know, and whether it can act:
| Term | Understands | Answers | Acts on other systems | Typical failure mode |
|---|---|---|---|---|
| Rule-based chatbot | Fixed buttons, keywords, scripted flows | Only scripted replies | No | Dead end when the question is off-script |
| Intelligent virtual assistant (IVA) | Free-form intent via NLP | Scripted answers plus model-drafted answers grounded in your content | Usually no — replies only in support settings | Confident wrong answer if it lacks a "don't answer" rule |
| AI agent | Free-form intent, multi-step planning | Answers as a by-product | Yes — calls tools, APIs, other systems | Wrong action taken autonomously |
| Voice assistant (Siri, Alexa) | Speech plus intent | Spoken answers | Yes — device controls and app actions (Apple's Siri page lists taking action across apps) | Misheard command executed |
Two practical notes. First, "IVA vs chatbot" is mostly a marketing line: a modern support IVA is a chatbot with language understanding, so judge the behavior, not the label — we map the whole family in types of chatbots. Second, an AI support agent that can look up an order or issue a refund sits one row lower in this table; most support IVAs, including TapCub's, only generate replies.
Capability levels: from scripted answers to actions
Within support specifically, IVAs sit on a ladder. Each level answers a different share of questions and fails differently:
- Exact-match answers. A curated question in the knowledge base matches the visitor's question and the stored answer is sent verbatim. No model call, no improvisation; it either matches or it does not.
- Retrieval plus drafting. The question is matched against your documents, and a language model drafts an answer from the retrieved passages. Flexible phrasing, but the draft is only as good as the source material.
- Open generation. The model answers from its own training, with no grounding in your content. Fluent on everything, reliable on nothing you have not verified. Large language models are prone to hallucination — plausible-sounding but nonfactual output — which is well documented in the research literature (Huang et al., 2023).
- Actions. The system changes something outside the chat: refunds, order edits, emails. This is agent territory, not IVA territory, and it needs permissions and audit trails most support widgets do not have.
The design decision that matters is the rule between levels: if the exact answer exists, send it; if only documents match, draft from them; if neither, stop and hand off. An IVA without that last clause is level 3 wearing a level 2 costume.
Worked example: 500 questions, three answering strategies
Sample data: one support inbox receives 500 questions in a week. Of these, 180 match a curated Q&A entry exactly, 240 can be answered from help-center documents, and 80 have no usable material (account-specific issues, edge cases, complaints). Three configurations answer the same 500:
| Strategy | Exact match | Model draft | Handoff to human |
|---|---|---|---|
| A. Rule-based bot (scripted only) | 180 | 0 | 320 |
| B. Answer-always LLM (no grounding, no handoff) | 0 | 500 | 0 |
| C. Grounded IVA (exact → draft → handoff) | 180 | 240 | 80 |
Now count only answers that are correct on the first reply. Sample assumptions: exact-match answers are correct 98% of the time (a few entries go stale); drafts grounded in matching documents are correct 90% of the time; ungrounded drafts are correct 85% of the time on the 180 easy questions the model happens to know, 40% on the 240 document questions, and 5% on the 80 with no public answer; humans resolve 95% of what reaches them. Repeat contacts (visitors who return within 7 days after a wrong answer): 20% after a wrong bot answer, 25% after a wrong human answer.
- A: 180 × 0.98 + 0 + 320 × 0.95 ≈ 176 + 304 = 480 → 480 ÷ 500 = 96% correct outcomes. The cost is queueing: 320 of 500 visitors (64%) wait for a human, and 240 of those waits were avoidable. Repeat contacts: (180 − 176) × 0.20 + (320 − 304) × 0.25 = 0.8 + 4 = 4.8 ≈ 5.
- B: 180 × 0.85 + 240 × 0.40 + 80 × 0.05 = 153 + 96 + 4 = 253 → 253 ÷ 500 = 50.6% correct first replies, with zero waits. Repeat contacts: (500 − 253) × 0.20 = 49.4 ≈ 49.
- C: 180 × 0.98 + 240 × 0.90 + 80 × 0.95 ≈ 176 + 216 + 76 = 468 → 468 ÷ 500 = 93.6% correct outcomes, with only 80 handoffs. Repeat contacts: (180 − 176) × 0.20 + (240 − 216) × 0.20 + (80 − 76) × 0.25 = 0.8 + 4.8 + 1 = 6.6 ≈ 7.
Read the table again with the thesis in mind. B answers every question instantly and is the worst configuration: 247 wrong first replies and about 49 repeat contacts — roughly 10 times A's and 7 times C's — because it never says "I don't know." A is honest and has the highest raw correctness, but it buys that by making 64% of visitors queue, including 240 whose answers were sitting in the help center. C gives up 2.4 percentage points of correctness against A (96 − 93.6) and cuts the human queue from 320 to 80 — a 75% reduction — with repeat contacts still in single digits. C answers 420 of 500 questions itself, and its 80 handoffs are exactly the questions it should not answer. The difference between B and C is not model quality — same model — it is the handoff rule.
Common mistakes
- Buying the label, not the behavior. "Intelligent" in a product name proves nothing. Test: send five questions your docs cannot answer. If the system answers all five fluently, it is level 3, whatever the datasheet says.
- Treating deflection as success. "Bot handled 90% of chats" counts wrong answers as wins. In the example, B deflects all 500 and generates ~49 repeat contacts. Measure correct resolutions and repeat contacts, not deflection alone.
- Letting the model answer account-specific questions. "Where is my order?" has no answer in any document. If the IVA cannot look the order up (most cannot — they only generate replies), the right output is a handoff, not a guess.
- Expecting actions from a reply-only IVA. An IVA that drafts answers cannot issue the refund it just described. If your workflow needs the action executed, you need an agent integration or a human — see AI support agents for that architecture.
- Skipping the knowledge base. Drafting quality is capped by source quality. If the documents are stale, a better model produces more fluent wrong answers; fix the knowledge base first.
How to evaluate an IVA for customer service
A checklist you can run in an afternoon, phrased as decision rules:
- The "don't know" test. Ask 10 questions with no answer in your content. If more than 2 get confident answers instead of a handoff or an honest "I can't find this," reject the configuration.
- The exact-match test. Ask your top 10 FAQ questions verbatim. If the answers are paraphrased by a model instead of sent from a curated entry, you are paying model costs and hallucination risk for questions you already wrote answers to.
- The handoff test. Ask for a human mid-conversation. If the bot keeps answering, the escalation path is broken; a visitor who asks for a person should get one.
- The boundary test. Ask it to change an order. If it claims to have done so, either it has action integrations you must audit, or it is lying. A reply-only IVA should say it cannot.
- The measurement test. Check that the reporting separates exact matches, drafted answers and human replies, and shows resolution plus repeat-contact signals. If everything is one "AI handled" number, you cannot tell B from C.
Where these tests fit into real workflows — support, sales, onboarding — is covered in chatbot use cases.
The 500-question example in TapCub
Strategy C is how TapCub's AI assistant is built, so the example maps directly:
- Load the two entry types. In Live chat › Knowledge base, a Q&A entry is sent as written when the visitor's line matches its question or contains every keyword — that is the 180 exact-match column, with no AI call consumed. Documents are split into roughly 600-character passages, retrieved by database full-text search (not a vector store), and the model drafts from the highest-scoring passages — the 240 draft column. A score that is too low counts as no match, so the question goes to the handoff column instead of being guessed.
- Run the "don't know" test before going live. Playground in the knowledge base runs the same path, spends no quota and creates no conversation, so you can send the 10 unanswerable questions from the evaluation checklist and the top 10 FAQ questions and see which column each lands in.
- Pick the mode that matches your staffing. In Live chat › Chat settings › AI assistant there are three modes: Off, Only when all agents are offline, or AI answers first (an agent taking the chat or replying, or the visitor asking for a person, stops the AI). Before its first reply the visitor is told they are talking to an AI, and that notice cannot be turned off. When the model decides the material is not enough, the reply carries an internal handoff mark the visitor never sees and the chat goes to a person — the 80-question column is designed behavior, not a failure. Setup steps live in set up AI chat replies.
- Read where it stopped. The inbox labels every handoff with its reason: no knowledge match, the model cannot answer, the visitor asked, the turn limit (Max AI turns per conversation, default 6), or an agent took over. Live chat › Reports adds the knowledge-base hit rate (direct Q&A match and retrieved answers together, as one share), resolution rate, first response that excludes AI replies, and satisfaction, which is attributed to the AI rather than an agent when only the AI answered. The hit rate does not split exact matches from drafts, so the reports will not show the 180/240 split on their own; the handoff reasons do show you the 80.
- Know the boundary. TapCub's AI only generates replies: it does not change orders, send emails or call other systems, and it is not a voice assistant. The refund question from the mistakes section gets handed to your team, with the visitor's source and recent page history already in the agent's sidebar. When a drafted answer is wrong, Fix & save stores the corrected reply as a Q&A entry, moving that question from the draft column to the exact-match column.
That is the thesis, restated as a product fact: the value of the 80 handoffs is not a gap in the AI — it is the feature that keeps the other 420 answers trustworthy.
Sources
- Google Cloud, Dialogflow ES: Agents — official definition of a virtual agent. Checked 2026-11-15.
- Apple, Apple Intelligence and Siri — voice assistant capabilities including actions across apps. Checked 2026-11-15.
- Huang et al., A Survey on Hallucination in Large Language Models (arXiv:2311.05232, 2023) — LLMs generating plausible yet nonfactual content. Checked 2026-11-15.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Customer Experience Consultant
Writes about live chat, visitor conversations, and self-service, sharing practical ways to make customer communication clearer and more effective.