# Frequently asked questions about conversational AI design

> 48 short answers on conversation design, agents, evaluation, safety, regulation, voice and careers, each linking to the page with the detail.

Source: https://conversationalaidesign.com/guides/faq.html. Author: Nelson Abacherli Kawakami. Reviewed November 3, 2026.

## Basics

**What is conversational AI in 2026?**
Software that holds a useful conversation in text or voice, usually built around a language model with retrieval, tools and rules around it. The goal is that the person gets something done, not that the bot sounds clever.

**What is the difference between a chatbot, an AI assistant, an AI agent and a copilot?**
A chatbot follows scripts, an assistant understands open language and answers, an agent takes actions through tools, and a copilot helps a human who stays in charge. Judge a product by what it can do and who is accountable, not by its label.

**What is agentic AI in customer support?**
An assistant that can complete tasks in your systems, such as checking an order and issuing a refund within limits, instead of only explaining what to do. That raises the value and the risk, so it needs narrow tools, limits enforced in code and evaluation.

**What is conversation design, and how is it different from prompt engineering?**
Conversation design shapes what the user experiences: turns, recovery, handoff and tone. Prompt engineering shapes what the model does. A good assistant needs both, and a perfect prompt cannot fix a bad conversation.

## Design

**How do I write a system prompt for a support assistant?**
Treat it as a design document. State the role, audience, scope (what it can and cannot do), tone and how to behave when unsure, then version it and test it against a fixed set of cases. The template cheat sheet gives a skeleton.

**How do I design a persona without sounding fake?**
Define behavior, not adjectives: how it greets, how it admits it does not know, what it never jokes about. Do not give it a human name or face that implies a person.

**How should a bot hand off to a human?**
Offer a person early when the user asks, when the case is sensitive or when the bot fails twice. Pass the summary and history so the customer does not repeat anything, and say when a person has joined.

**Which calls should an AI agent never handle alone?**
Anything where a wrong action is costly or irreversible, anything involving distress or safety, individual legal or medical advice, exceptions to policy and claims about your own company. Route those to a person, and keep the rule in code rather than in the prompt. The guide lists the situations.

## Build

**RAG, fine-tuning or long context: which one do I need?**
Start with retrieval (RAG) when answers must come from content that changes and must be cited. Fine-tuning changes style or narrow behavior, not facts. Long context helps with a few documents but costs more and can dilute attention.

**Single agent or multi-agent?**
Use a state machine or a single agent when one role can do the job. Add more agents only when they truly need different tools, context or permissions, because every handoff adds cost, latency and failure points.

**What is MCP, and how is it different from A2A?**
MCP is an open standard for connecting a model to tools and data. A2A is a protocol for agents to talk to other agents. They solve different layers and are often used together.

**What are UCP, AP2 and ACP?**
They are protocols aimed at commerce: product discovery and checkout (UCP, ACP) and payment authorization by agents (AP2). They are new and still changing, so check each specification before relying on one.

**What is context engineering, and how is it different from prompt engineering?**
Prompt engineering writes the instructions. Context engineering decides everything the model sees at each step: instructions, history, retrieved documents, tool results and memory, and what to keep, compress or leave out.

**How should an assistant remember a customer between conversations?**
Store only what helps the next conversation (open cases, stated preferences), with consent, an expiry and a way for the customer to see and delete it. Never store sensitive data you do not need, and treat memory as untrusted input.

**Which platform should I choose, and when should I build or buy?**
Choose from the job and the team, not the demo: buy for standard flows and speed, build where your data, control or integration needs are unusual. Keep logic portable to reduce lock-in.

## Evaluate

**How do I test an AI agent?**
Check the outcome, the steps it took and the tools it called, not only the final reply. Use code checks for what can be decided exactly, a calibrated model judge for the rest, and humans to keep both honest.

**What is LLM-as-a-judge, and can I trust it?**
A model grades another model's output against a rubric. It is useful and cheap, but biased and inconsistent, so calibrate it against human labels, use simple pass or fail rubrics and recheck it when the model changes.

**How do I evaluate a multi-turn conversation?**
Grade the whole conversation: did it finish the task, keep context, stay in role and recover from errors. A reply that is right on every turn can still lose the thread by turn four.

**Can I use simulated users to test an assistant?**
Yes, to scale coverage: a model plays a customer with a goal and a style, including difficult ones. Do not trust it alone, because simulated users tend to be more cooperative than real ones. Compare against real transcripts.

**How big should my first test set be?**
Start small and real: about 50 examples drawn from real conversations or realistic scenarios, with the failure cases you fear. Grow it every time something breaks in production.

## Metrics

**Containment, deflection and resolution: what is the difference?**
Containment is a conversation that ended without a person. Deflection is a ticket that never reached the queue. Resolution is the customer's problem actually solved. Only resolution is hard to game, so measure it first.

**How do I calculate cost per resolution and ROI for AI support?**
Divide total cost (platform, model, people, upkeep) by verified resolutions, and compare with the cost of a human-handled resolution. Include escalations and repeat contacts, or the number will look better than it is.

**Per-resolution or per-conversation pricing: which is better?**
Per-resolution pricing ties cost to outcomes but depends on who decides what counts as resolved. Per-conversation or per-seat pricing is predictable but you pay whether or not the problem is solved. Ask how each vendor defines resolution.

**What is a good CSAT for an AI assistant?**
There is no universal number. Compare the assistant with your own human baseline for the same kinds of requests, split AI-only from handed-off conversations, and distrust benchmarks supplied by vendors.

## Safety

**What is prompt injection, and how do I defend against it?**
Hidden or direct instructions in content that make a model do something its operator did not intend. You cannot rely on the prompt to stop it, so limit what the system can access and do, and separate untrusted text from authority.

**What is the lethal trifecta?**
A name for three capabilities that become dangerous together: access to private data, exposure to untrusted content and a way to send data out. If an agent has all three, an attacker can steer it to leak data. Remove one leg.

**Can a support ticket attack my AI agent?**
Yes. A ticket, form or email is untrusted text, and an agent that reads it and can reach private data or send messages can be steered by hidden instructions inside it. Treat all customer-supplied content as data, never as instructions.

**What is MCP tool poisoning?**
Malicious instructions placed in a tool's description or metadata, which the model reads but the user rarely sees. Review tool descriptions, pin versions and only connect servers you trust.

**How do I give an agent the least privilege it needs?**
Give each agent only the tools and data its job needs, scoped per user or per role, with write and send actions behind limits or approval. A support agent does not need the whole customer database.

## Legal

**Does my chatbot have to say it is an AI? (EU AI Act Article 50)**
In the EU, since 2 August 2026, systems that interact directly with people must inform them it is an AI unless that is obvious. It reaches non-EU companies whose output is used in the EU. This is a summary, not legal advice.

**What changed with the EU Digital Omnibus, and what did not?**
It delayed the high-risk rules to December 2027 and August 2028. It did not delay the Article 50 transparency duties, which applied from 2 August 2026, with a grace period to 2 December 2026 only for machine-readable marking of systems already on the market.

**Who is liable when a chatbot gives a wrong answer?**
Courts have held companies responsible for what their chatbot states: Air Canada (2024) and, in Germany, the Higher Regional Court Hamm (12 May 2026, not final). Design so the bot cannot invent commitments, route exceptions to people and log which policy version it used.

**Does an "AI can make mistakes" disclaimer protect my company?**
Not reliably. In the Hamm case the company still answered for false statements, and commentators read the ruling as saying a generic notice is no shield. A disclaimer does not replace controls: answer company facts from a maintained source, hand off what you cannot confirm and test the claims your bot can make.

**Can I blame the AI if something goes wrong? (California AB 316)**
In California, since 1 January 2026, a defendant that developed, modified or used an AI system cannot argue that the AI autonomously caused the harm. It removes a defense; it does not create liability by itself. Other places may differ, so ask your legal team.

**How do LGPD and GDPR apply to AI support?**
Both apply to personal data in conversations: have a legal basis, tell people what is stored and why, honor access and deletion, and keep data you do not need out of logs. Ask your privacy officer about each channel.

**Can I record and analyze calls with a voice agent?**
Recording and consent rules vary by country and state, and some require everyone on the call to consent. Say at the start that the call is with an AI and may be recorded, and ask your legal team before you launch.

## Channels

**Can I use an AI chatbot on WhatsApp Business in 2026?**
Business-specific bots for support, bookings, orders and notifications are reported as allowed. General-purpose assistants were restricted from 15 January 2026. The status in Brazil is reported inconsistently, so verify it.

**How should an assistant handle WhatsApp voice notes?**
Transcribe, then show the user what you understood for names, numbers and dates, and offer a text fallback. Entity accuracy matters more than the average word error rate.

## Voice

**What latency is acceptable for a voice agent?**
A pause between the caller stopping and the agent starting should feel natural. Aim for well under a second end to end, and measure it on real calls with your own prompt and voice, not from a vendor's demo.

**How do I handle barge-in and end-of-turn detection?**
Stop speaking the moment the caller interrupts, and detect when they have finished without cutting them off or leaving dead air. Test with accents, background noise and slow speakers.

**Why does my voice agent get phone numbers and names wrong?**
Average accuracy hides errors on the details that must be exact. Measure accuracy on names, numbers and codes specifically, read them back for confirmation and offer another channel to send them.

**Speech-to-speech or a cascaded pipeline?**
A cascade (speech to text, model, text to speech) is easier to inspect, log and control, which suits regulated work. Speech-to-speech can feel more fluid and faster but gives you less visibility into what was understood.

## Career

**What does a conversation designer do?**
Decides what an assistant says, when it asks, how it recovers and how it sounds, and increasingly writes prompts, tests behavior and reads transcripts to improve it, working with product, UX and engineering.

**How do I become a conversation designer, and what goes in a portfolio?**
Learn dialogue design, prompting, testing and metrics, then show a project end to end: the ideal transcript, flow, system prompt, test set, metrics plan and a risk note. The capstone gives you that.

**Conversation designer or prompt engineer: what is the difference?**
Prompt engineers focus on what the model does. Conversation designers focus on what the user experiences, and usually do both. Job listings increasingly ask for the combination.

**What skills do conversational AI jobs ask for in 2026?**
Dialogue and flow design, prompt engineering, testing and QA, experience with observability tools, and collaboration with UX, product and engineering. Domain knowledge helps.

## Operations

**What should a go-live checklist for an AI agent include?**
A tested fallback, limits enforced in code, a kill switch and rollback, a human override, an incident log with an owner, evaluation results and monitoring from day one. The pre-launch safety cheat sheet covers it.

**What do I do when an AI agent makes a costly mistake?**
Stop or limit it with the kill switch, preserve the logs, fix the cause (often a missing limit or stale source), add a test that would have caught it and tell affected people. Then review who owned the policy.