Conversational AI
Conversational AI designed as a dialogue, not a single prompt
Conversational AI is the design of a dialogue over several turns: what the system remembers, what it asks next, and when it should stop talking and act. We build that state on purpose, instead of sending the entire transcript to a model and hoping.
- Dialogue design
- State and validation
- Memory limits
- Exit paths
State is the product
A booking dialogue has to remember the date the person already gave. A support dialogue has to remember which order they named. That state should live in your application, where you can validate it, not only inside a model’s context window. When the conversation is over, you keep what you need for the record and drop the rest.
Voice and text
Most of the value we see in the UAE starts as text, especially WhatsApp and web chat. Voice is added when the setting truly needs it, such as a phone line, and it brings transcription errors that the dialogue must survive. We do not lead with voice for a problem that is really a missing FAQ.
A dialogue with a way out
Designed in
- A short memory of this conversation
- A state for the task in progress
- An exit to a person or a form
Left out
- A permanent memory of every customer
- Open chat with no task
- Personality that hides a missing answer
What changes the conversation
The job of the chat
Booking, support, and internal help are different state machines.
Memory
Remember this visit. Do not quietly build a profile you have not described.
Language
Arabic and English in one thread need a rule for which language the record stores.
What a dialogue brief needs
- 01
The task
Booking, support, or internal help.
- 02
The memory
This visit only, unless you have described a longer record.
- 03
The exit
A person or a form, and when the chat should offer it.
- 04
The language
Which language the stored record should keep.
Marks a conversation is one thread
- 01
The channel they use
The bot is on the desk, site, or WhatsApp people already write to.
- 02
Language is specified
Arabic, English, or both is a decision for the first release, not a surprise.
- 03
It knows the stop
Complaints, payments, and anything uncertain go to a person.
- 04
Staff see the same thread
The handover includes what was already said.
What we build
Dialogue design
Turns, confirmations, and the questions that are allowed to be asked only once.
State and validation
Slots filled by the conversation and checked against real availability or records.
Memory limits
A defined retention period for transcripts, aligned with your privacy decision.
Exit paths
Form, human agent, or a ticket, chosen per intent.
How an engagement runs
- 01
Script the golden dialogues
And the three ways each one usually goes wrong.
- 02
Implement state outside the model
The model proposes values. The application accepts them.
- 03
Listen to real sessions
The second week of transcripts matters more than the design workshop.
Questions we hear
How is this different from an AI chatbot page?
The chatbot page is the service: channels, handover, and business use. This page is the dialogue mechanics: state, turns, and memory. A project usually needs both.
Can the conversation switch language mid-way?
We can support that, and we test it. Mixed Arabic and English in one sentence is common and has to be in the sample, not treated as an edge case.
Do you store every message forever?
No. Retention is a product decision. We default to keeping what the business process needs and making deletion possible. Your counsel should set the period.