Conversational UX Design
A conversation is written, not improvised
The model can talk. That isn't the same as being useful. We write the scenes an agent actually plays: how it opens, what it confirms, the voice it does it in, and above all what it says the moment it's lost. That last one is where the experience is decided.
SC. 01 — the booking
voice
Persona
(Warm, brisk, never chatty. Uses the caller's words back.)
I can book, move, or cancel an appointment. Which is it?
i need to move my thing
Happy to. I've got your cleaning Thursday at 2. Is that the one?
Off-script
(If the reply matches no intent, twice)
I didn't catch that. I can book, move, or cancel, or put you through to the desk.
The last box is the whole job. Anyone can write the greeting; the experience is decided by the line that offers a way forward instead of an apology.
Designed
Not improvised
No dead ends
Always a next step
A door to a human
Always reachable
On-brand
One consistent voice
The brief
Same moment. One of these loses the customer.
The model is fluent in both takes. Fluency was never the problem. What the agent says when it doesn't understand is.
Take 1 — improvised
Sorry, I didn't quite get that. Could you rephrase your question?
i said i want to move my appointment
Sorry, I didn't quite get that. Could you rephrase your question?
(The caller hangs up.)
Take 2 — designed
I can book, move, or cancel. Which is it?
i said i want to move my appointment
Got it: your cleaning Thursday at 2. What day suits you?
(The appointment moves. Nobody needed a human.)
What changes
SC. 01
The loop that made people hang up is gone
The agent that answered every confusion with the same apology is replaced by one that offers a way forward. The single biggest cause of abandoned conversations simply stops happening.
SC. 02
People reach a human when they need one
A clear, always-available exit means nobody feels trapped talking to a machine. Counter-intuitively, that's what makes them willing to try the machine at all.
SC. 03
It sounds like your team
A defined persona and tone, applied consistently (errors included), makes the agent read as part of your team. Nobody should feel a generic assistant has been bolted onto your brand.
SC. 04
Engineering builds from a tested script
Flows, prompts, and fallbacks arrive written and tested, so the build implements a validated conversation and nobody invents the wording along the way.
What we design
Every scene the agent has to play
Chatbot conversations
The scripts, flows, and prompts that make a chat agent genuinely helpful, written for a reader who can see options and scroll back.
Voice interfaces
Voice flows designed for the ear: how it opens, confirms, recovers, and closes a call, with no buttons and no scrollback to lean on.
Dialogue flows
Branching conversation maps that cover the real paths people take, well beyond the happy one that demos well.
Errors & fallbacks
What the agent says when it doesn't understand. It's the part that makes or breaks the experience, so we write it as carefully as the main scene.
Persona & tone
A defined personality and voice, so every message is consistent and on-brand, from the greeting to the apology.
Escalation & escape
Clear paths to a human and out of loops, so no one ever feels trapped talking to a machine.
The engagement
From what people actually want to a script that ships
Week 1
Map the intents
We chart what people actually come to the agent to do: the real jobs, in their own words, whatever the org chart suggests.
Week 2
Write the scenes
The flows, prompts, and fallbacks get written, with a persona and tone defined and applied to every line, including the errors.
Week 3
Read it with real users
The conversation is tested with real people against real tasks, and tuned wherever it confuses, loops, or frustrates.
Week 4
Hand it to build
Flows and scripts go to engineering build-ready, and we refine after launch against what real conversations actually do.
How we work
Four rules we hold while writing it
Design the failure first
Anyone can script the happy path. We write the off-script moment first (the misunderstanding, the change of mind, the request for a human) and polish the greeting after, because that's where the experience is decided.
Test the dialogue, not the diagram
A flow chart always looks sensible. We put the actual wording in front of real users and watch where they stall, because a conversation only fails out loud.
Write to the medium
Voice has no buttons and no scrollback; chat can show options and history. We write shorter turns and firmer confirmations for the ear, and never ship one script for both.
Hand off build-ready
Flows, prompts, fallbacks, and persona notes arrive specified enough to implement without a meeting, so the wording that shipped is the wording that was tested.
The notes
Six rules every line gets held to
None of them are about the model. All of them are the difference between an agent people finish a task with and one they escape from.
01Say less
Every extra word is friction in a conversation. A shorter turn is understood faster and misheard less. In voice especially, it's the difference between a caller following along and a caller giving up.
02Always offer a next step
Never a dead end. Every turn, including the ones where the agent is lost, ends with something the person can actually do. That's the whole difference between a helpful agent and a wall.
03Confirm before anything irreversible
Canceling, paying, sending. The agent reads it back before it commits, because the cost of a misheard word only ever shows up afterward.
04Handle 'I don't understand' as well as the happy path
The failure line gets the same care as the greeting. It names what the agent can do and skips the looping apology, which is what turns confusion into progress.
05Give a clear way to reach a human
Always. The exit is stated up front, and knowing it exists is what makes people willing to try the agent first before demanding a person.
06Keep the persona and tone consistent
One voice everywhere, errors included. Tone that drifts between charming and robotic is how an agent stops feeling like your brand and starts feeling like software.
Why it pays
Fewer conversations abandoned
The dead-end loop is the main reason people quit an agent. Removing it keeps the conversation alive long enough to actually finish the job.
Fewer escalations that didn't need a human
An agent that recovers gracefully resolves the things it can, so the handoffs that do reach your team are the ones that genuinely needed a person.
A brand that survives the error message
Consistent persona and tone across every line, the apology included, means the agent reinforces your brand. It never quietly undermines it.
A build that ships the tested wording
Because the script is specified and validated before handoff, engineering isn't inventing copy mid-sprint and nobody relitigates the words after launch.
Selected work
Two agents that knew what to say when it went wrong
What you get
A script your engineers can build from
- 01Mapped intents and conversation flows
- 02Written scripts, prompts, and fallback responses
- 03A defined persona, voice, and tone
- 04Error handling and escalation paths
- 05Conversations tested with real users
- 06Build-ready specs for your engineering team
Industry expertise
Where a bad line costs a customer
Healthcare & dental
Reception and booking conversations where a misheard date matters and the exit to a human is non-negotiable.
Home & field services
Missed-call and dispatch agents that have to qualify a job by voice, quickly, without sounding like a robot.
Restaurants & hospitality
Ordering conversations designed for the ear, where confirmation before payment is the difference between a meal and a complaint.
Real estate
Lead-qualifying chat that asks the few questions that matter and hands warm ones to an agent while they're still interested.
Professional services
Internal assistants over policy and process, where tone and precision both matter and a wrong confident answer is worse than none.
E-commerce & retail
Support conversations that deflect the routine and escalate the rest, with a voice that matches the storefront it sits behind.
SC. 07 — the read-through
Tell us the moment your agent loses people
Usually it's one line, repeated. Send us the transcript where the conversation dies and we'll show you the scene rewritten, with persona, fallback, and the door to a human included.
Why us for this
Note 01
We write the failure before the greeting
The off-script moment is where conversational UX is actually decided, so it's the first scene we write. We don't leave it to the model to improvise.
Note 02
We design to the medium
Voice and chat are genuinely different instruments. For the ear, we write shorter, confirm harder, and recover more carefully, and each medium gets its own script.
Note 03
We test the wording with real people
Flow diagrams always look reasonable. We put the actual lines in front of users and tune wherever they stall: the same test-and-learn we bring to any interface.
Working with Flaidex
Design and build under one roof
This is the design of the conversation; our AI Agents team can build it. Nothing gets thrown over a wall, and the people who wrote the script stay reachable through implementation.
A persona you can hold us to
The voice is written down and applied to every line, so consistency becomes a document your team can check against. It won't drift from release to release.
Honest about what the model won't fix
A fluent model still frustrates people if the conversation isn't designed. We'll tell you when the problem is the script, not the technology.
Questions
What people ask about conversation design
Isn't the AI model good enough that we don't need to design the conversation?
The model can talk, but talking and being useful are different things. Good conversational design decides what the agent is for, how it opens and confirms, what it says when it's unsure, when it hands off to a human, and what tone it uses. Skip that and you get a fluent bot that still frustrates people. Designing the conversation is what makes it genuinely help.
What matters most in conversational UX?
How it handles going wrong. Anyone can script the happy path; the experience is made or broken by what happens when the agent doesn't understand, when the user changes their mind, or when they need a human. We design those moments as carefully as the main flow.
Chat, voice, or both?
We design for either, and they're genuinely different. Voice has no buttons and no scrollback, so it needs shorter turns, clear confirmations, and careful recovery. Chat can show options and history. Each medium gets a script written for it.
How do you set the persona and tone?
We define a personality that fits your brand and audience (warm or brisk, playful or precise) and apply it consistently across every message, including the errors. It's what makes the agent feel like part of your team.
Do you test conversational designs?
Yes. We put the flows in front of real users, watch where they get confused or stuck, and refine the wording and structure. It's the same test-and-learn approach we bring to any interface, applied to dialogue.
Does this include building the agent?
This is the design of the conversation; our AI Agents team builds it. We hand off flows, scripts, and specs that are ready to implement, and we can stay involved through build and after launch to keep the experience sharp.
More in Product Design & UX
Let's talk
Running a large platform, shaping a first MVP, or getting a product ready for a funding round? Tell us where you are. We'll shape the process around it, and stay with you after launch.
