You want to add a conversational AI to an existing children's product
The product already works, and you want to know which role suits it: tutor, companion, or search assistant, before building the feature.
A Dawani service
Designing and reviewing AI products for children
We design the AI features a child meets, and review a feature you already have, for brands and bodies that want the child to know they are not talking to a person. We govern every decision with parental consent and real testing with children, and design around as little memory and data as possible kept by the system about the child.
What you get
- A scope document naming the AI's role in your product: tutor, story character, companion, coach, or search assistant
- An assessment of the design decisions that set the risk: memory across sessions, an apparent persona, attempts to persuade, epistemic authority, and how it handles a sensitive disclosure
- A plain gray prototype tested with real children with their parents' consent before any launch
- A success indicator and a guardrail indicator agreed before launch
How we work
Every AI feature passes through our design laws, and before any launch we test a plain gray prototype with real children. We classify the feature's role and its risk with the AI Child-Risk Classifier, and review it as a whole against the Dawani Child-Safe Design framework.
When you need this service
Three situations that need review before a child meets the AI
An existing AI feature you want reviewed from the child's angle
The feature already works, and you want an independent eye to review its memory, its persona, and how it handles a sensitive disclosure.
A product where you want the child to know they are not talking to a person
The goal is repeated clarity, not a mention made once at sign-up and then forgotten.
How the service runs
Four stages, a gate after each
1 Understanding the goal
We start from the AI role you want, and from the age of the child who will meet it.
Gate: The goal leaves this stage once we agree on a specific role for the feature, not a general description of its idea.
2 Written framing
We write the feature's limits: its memory, its persona, and what it does when a child discloses something sensitive.
Gate: The framing leaves this stage once these limits are written and agreed, instead of staying a verbal understanding.
3 Gray prototype and testing with children
We build a plain gray prototype of the feature, and real children try it with their parents' consent.
Gate: The prototype leaves this stage once the feature works with children as intended, not once the internal team is pleased with it.
4 Delivery and measurement
We deliver the feature, and measure it against the success indicator and the guardrail indicator agreed from the start.
Gate: This last stage is never left behind; the guardrail indicator stays with the feature for its whole life, not only at launch.
In scope and out of scope
| Aspect | In scope | Out of scope |
|---|---|---|
| The feature's role and persona | Classifying the role, decisions on memory, persona and persuasion, and the gray prototype tested with real children. | Building or training the language model itself from scratch, unless agreed within the proposal. |
| Content safety and security | Designing how the feature handles a sensitive disclosure from a child, and when it hands them to an adult. | A certified safety assessment of the model itself, or a legal opinion on its liability. |
| Timeline and price | A proposal built on your feature's own case once we understand it, naming what is out of scope as clearly as what is in it. | A time or price announced in advance, before the request is understood. |
| The outcome for the child | A success indicator and a guardrail indicator agreed before we start. | A guaranteed outcome for every child who uses the feature. |
Before you write to us
A tool, a post and a reference
AI Child-Risk Classifier
A tool that runs in your browser: classifies your feature's role and its risk, before you describe the product to us.
Open the toolAI tutor or AI friend?
A post explaining the design decisions that separate a safe tutor from a companion carrying a much larger risk.
Read the postThe Dawani Child-Safe Design framework
The reference every AI feature we design or review is measured against.
Open the frameworkBefore a first conversation
Questions often asked about this service
Do you build the AI model itself, or the feature around it?
The feature around it: the role, the memory, the persona, and how it handles a sensitive disclosure. Choosing the model and its provider is a separate technical decision.
How do you know a child understands they are not talking to a person?
We test that directly with real children during the gray-prototype trial, not by assuming a passing mention is enough.
What happens when a child discloses something sensitive to the feature?
We design a clear path that hands the child to their parents or a responsible adult, instead of leaving the reply to the model alone.
Start here
Have an AI feature a child will meet?
Describe your situation plainly, and we will get back to you within a few working days.
Your path: brand or company
Your idea arrived.
A real person from the Dawani team will read it and reply within a few working days. Thank you for thinking about kids.
Updated: