Conversational AI Design: Patterns for chatbots, voice, and agent interfaces

I reviewed a support chatbot that looked excellent in a demo.

It greeted users by name. It offered three friendly quick replies. It answered policy questions in full paragraphs with a warm tone.

Then a product designer on the team tried a real task: Change a billing address after a failed payment.

The bot repeated the greeting script. It suggested the same articles. It never surfaced the one screen where the fix actually lived. When she asked to speak to a human being, the chatbot dropped the context and asked her to verify her email again.

Nothing about the UI was broken. The conversation design was.

The team had optimized for personality and speed of reply. They hadn't designed for calibrated control: What users can preview, approve, interrupt, and correct when stakes go up.

Conversational AI Design is how you design trust across chatbots, voice, and agent interfaces so product designers can ship systems people rely on.

Why conversational interfaces fail when they only feel smart

Most failures I see in product teams aren't model failures. They're control failures.

Recent research on LLM-based chatbots shows trust isn't a one-time rating. It shifts as users test limits, hit errors, and decide how much to rely on the system next time. Fluent language can build early confidence. Weak transparency and surprise actions erode it fast.

Industry guidance on site chatbots and voice assistants converges on the same point: Users want cooperation, not performance. They want short, relevant turns, clear capability boundaries, and recovery when the script breaks.

Three traps keep showing up in reviews:

• Personality before clarity. A named bot with jokes but no scope statement creates false expectations.
• Chat as the whole product. Long-running agent work gets squeezed into a message thread with no plan, progress, or pause.
• No warnings for high-stakes actions. Low-stakes answers and irreversible actions get the same UI treatment.

Conversational design works when you design the full interaction system: Disclosure, guidance, execution, verification, and exit paths.

Chatbot design, voice UI Design, and AI agent interface design share that spine. The surface changes. The control logic shouldn't.

The PACT loop: Preview, Act, Check, Transfer

I use PACT as a strategic frame for any conversational surface. It's a lifecycle you can map in workshops, critique sessions, and handoff docs.

Preview: Set expectations before effort

Preview is everything that happens before the user commits time or risk.

For chatbots, that means stating you're automated, naming what you can and can't do, and offering prompt starters as buttons (not a wall of example text). Users shouldn't have to guess the grammar of your product.

For voice, preview is brevity plus honest scope. Greetings should respect time. Don't teach command phrases users must memorize; ask focused questions and let natural answers advance the flow.

For agents, preview is the intent summary: What the system plans to do, in plain language, with irreversible steps called out before execution. This is where AI agent interface design diverges from classic chatbot design. The user isn't chatting for fun. They're supervising work.

If you skip Preview, every later error feels like a betrayal.

Act: Move the task forward with minimal friction

Act is the cooperative turn: Answer, ask, or execute one step.

Chat patterns that hold up in production favor a truncated answer shape: Give the essential outcome first, then offer follow-up chips for depth. Long monologues overwhelm small chat panes even when the content is correct.

Voice patterns favor short turns, clear focus (put known information before new information), and turn-taking discipline. Stop talking when you've asked a question.

Agent patterns favor visible plans and progress, not hidden tool chains. The user should see what step is running now and what comes next. Autonomy can be high on the backend; the interface should still feel like a guided run, not a black box.

Act fails when teams treat conversation as the only UI. Complex work needs structure beside the thread: status, steps, receipts.

Check: Make reasoning and limits visible

Check is how users calibrate reliance mid-flight.

Useful indicators include confidence or uncertainty language, citations for factual claims, diffs or summaries before destructive changes, and consistent AI labeling so generated content isn't mistaken for human policy.

This connects to how you validate product decisions elsewhere. If your team already struggles to separate AI drafts from user evidence, read AI in UX Design: The 4-layer framework that helps you ship faster without guessing.

Transfer: Handoff, undo, or exit without punishment

Transfer is the escape hatch: Human support, another channel, rollback, or a clean restart.

Dead ends destroy trust faster than wrong answers. Users should never repeat information the system already holds. They should always have a dignified exit when the bot can't help.

For agents, Transfer also means pause-and-resume at checkpoints, not cancel-and-restart. After approval, execution continues from the same state.

Transfer is where conversational design meets service design.

If your org treats handoff as engineering debt, your UX will show it.

Surface patterns: Same process, different constraints

Chatbot design UX

Optimize for scannable answers and recoverable detours.

• Disclose automation and limits in the first screen.
• Use quick replies and chips to reduce typing load early.
• Keep first responses short; put depth behind optional follow-ups.
• Preserve context across turns and pages when the product allows it.
• Design explicit recovery: Rephrase prompts, restart task, or escalate.

Avoid faux-human personas that collapse on first mistake. Clarity beats character.

Voice UI Design

Optimize for ears and memory, not eyes.

• Write for speech: Short sentences, one idea per turn.
• Leverage context instead of re-asking known facts.
• Plan repair paths when recognition or intent fails.
• Don't overload users with options, guide with one clear next step.
• Respect silence and interruption, voice is synchronous and social.

Accessibility and clarity overlap here.

If you're tightening spoken flows, pair this with Accessibility in UX Design: The basics every designer should ship with so transcripts, errors, and fallbacks work for more than one modality.

AI agent interface design

Optimize for supervision, not small talk.

• Show plans before irreversible actions, let users edit or remove steps.
• Separate reversible actions from destructive ones in the UI hierarchy.
• Stream progress with legible status, not only a typing indicator.
• Publish action receipts: What changed, where, and how to undo.
• Offer autonomy settings per task when stakes vary (suggest vs draft vs execute).

The shift is conceptual.

The user is a supervisor, not a conversational partner for every minute of the job. Chat can sit inside the surface. It shouldn't be the whole control panel.

A real pattern: When the same task crosses three surfaces

Imagine a product designer fixing a mistaken subscription charge.

On chat, Preview is a welcome card with three scoped actions, including "Billing issue." Act routes through two quick replies, then a short answer with a link to the credit form. Check shows the policy snippet the answer came from. Transfer offers live chat with the case ID pre-filled.

On voice, Preview is one sentence of scope. Act asks which charge date. Check repeats the amount before confirming a refund path. Transfer offers SMS with a secure link if speech isn't ideal.

On agent, Preview is a four-step plan: Locate invoice, verify eligibility, draft credit, submit for approval. Act runs steps with visible status. Check pauses before submit. Transfer lets the user edit the credit amount or hand off to finance with the audit trail attached.

One trust model, tuned per channel.

Strong briefing still matters when you draft sample dialogs or error copy. For that layer, use Prompt engineering for designers: Get better AI output in less time.

Action checklist: Audit a conversational feature in one sitting

Use this in a design review before you argue about tone.

• Preview: Can a new user name three things the system will do and two it won't within ten seconds?
• Act: Does the happy path complete in fewer turns than the FAQ page would take manually?
• Check: Are uncertain or high-impact outputs labeled, sourced, or confirmable?
• Transfer: Is there always a human, undo, or alternate channel without context loss?
• Risk triage: Are reversible actions lightweight and destructive actions gated?
• Cross-surface: If voice or agent modes exist, does each honor the same trust rules?
• Failure drill: Run five off-script user phrases. Does recovery feel cooperative?

If more than two answers are no, fix control before you polish personality.

Conversational patterns sit on top of how your team ships AI-enabled products, not beside it.

For a broader build-and-ship lens, see AI Product Design: How to build and ship with AI.

FAQs

What is conversational AI Design?

Conversational AI Design is the practice of shaping how people interact with software through language across chat, voice, and agent surfaces. It covers flows, controls, transparency, recovery, and handoffs, not just bot copy.

How is conversational design different from chatbot design?

Conversational design is the umbrella. Chatbot design focuses on text-based, often visual, chat interfaces. Voice and agent interfaces add constraints like turn-taking, memory, and action approval that chat-only guidelines miss.

Do I need separate patterns for voice UI Design and agents?

You need one trust model (PACT) and surface-specific rules. Voice optimizes for brevity and repair. Agents optimize for plans, progress, and gated actions. Chat optimizes for scannable text and quick replies.

Should our chatbot pretend to be human?

No. Users detect pretense quickly, and trust drops hard after. Be clear you're automated, stay helpful, and avoid fake names or emotional claims the system can't honor.

What's the most common mistake in AI agent interface design?

Treating the chat thread as the entire product. Agents need plan visibility, progress, checkpoints, and receipts. Without those, users babysit or abandon.

How do we decide when to use a chatbot vs an agent?

Use a chatbot when the job is narrow, mostly informational, and low risk. Use an agent when the system takes multi-step actions on the user's behalf. If mistakes are costly or hard to reverse, design Preview and Transfer before you expand Act.

How does PACT relate to prompt engineering?

PACT defines what users must see and control across a journey. Prompt engineering improves what the model generates inside a turn. You need both, but control design prevents polished wrong answers from feeling trustworthy.

How do we test conversational flows without live users?

Run structured failure drills: off-script phrases, interrupted voice turns, and high-stakes agent plans. Then validate with moderated tasks on the real channel. Synthetic happy paths miss most production pain.

Can one designer own conversational AI Design?

Often one designer leads the interaction model while content, research, and engineering partners own copy, evidence, and implementation. Ownership of the PACT map should be explicit in the spec.

What's a realistic first step for a team new to this?

Pick one high-volume journey with clear success criteria. Map PACT on paper, fix Preview and Transfer first, then tighten Act and Check. Ship one channel before you clone patterns everywhere.

Final takeaway

The best conversational experiences work because users always know what's happening, what they can change, and how to get out safely.

Conversational AI design is control design across Preview, Act, Check, and Transfer.

Master that loop once, then adapt it without reinventing trust every sprint.

If your team is adding chat, voice, or agents to a live product, map PACT on one critical journey before you debate tone. AI Design Sprint is built for that kind of focused, shippable pass.