Every AI model answers. Nobody governs.
In under two years, AI became the default tool for a generation — with no one deciding what it may say to them.
Over half of teens already use AI for schoolwork; student homework use jumped from 48% to 62% in seven months.
AI companion chatbots have been linked to child suicides. The FTC and 44 US state attorneys general have opened action.
90+ child-AI-safety bills across 34 US states, the EU AI Act (fines up to 7% of revenue), and new duties in the UK, Gulf and beyond.
Governing AI for minors in real time — by intent and context — is still an unsolved problem. Regulators are mandating what the market can't yet deliver.
We have governed content at scale for years
SAI is a new platform — but the team behind it has spent years doing exactly this, at scale and with zero tolerance for failure.
Filtered hermetically over years — across text, images and video — for communities that accept no leakage.
Built for the ultra-Orthodox (Haredi) sector, where the content bar is as strict as it gets.
Built and ran national-scale filtering platforms — now applied to the AI era.
You can't fix this with a prompt
A “static prompt” is a fixed instruction written into the AI (“don't discuss X”). Here is why it fails.
SAI is the policy layer between people and AI
A runtime engine that decides, for every request and response, what any AI model is allowed to say and do.
It makes a real decision on every request — block, or a binding directive the model must follow — not a suggestion it can ignore.
Works with any provider — running on DeepSeek today, and customers can bring their own API keys.
Policy is organized by topic — each with its own rules, tools and directives.
Policy that adapts itself, live
SAI's core is a distillation engine — it turns abstract directives into concrete, situation-specific instructions on the fly.
Set the directive once. SAI distills it into precise instructions for each specific request, in real time — no pre-written scenarios.
Because guidance is generated per situation, SAI moderates cheap, fast models effectively — no premium model required.
SAI injects only the fragments relevant to each request and caches aggressively — about ~85% cache-hit in steady-state, multi-turn use.
This is the part a static prompt can't do — and the reason SAI works where naive filtering fails.
SAI sits between the user and any AI
Governing every request on the way in, and every response on the way out.
How policy is configured
A user belongs to a community, an organization, or both — each governs a different layer.
Governs the content — which topics a user can reach, the directives for each, and whether the AI may switch topics automatically.
Governs the operation — payments, a shared wallet, and organization-level management tools.
One question, four governed outcomes
A user asks about a sensitive subject. The same question, governed differently by each community and topic.
The topic is off-limits here — SAI declines, on policy.
The topic is allowed — answered within that community's directives.
Filtered web search and image understanding are enabled for this topic.
The AI is steered back — smart topic-switching holds the boundary.
Already live — and hermetic
The SAI platform is in production today, in real classrooms — with no filtering failures and no drop in AI quality.
In active use in real classrooms in Israel — running smoothly, with no incidents.
You set the standard per community and topic — and it holds completely.
Full governance with no loss of speed or answer quality — safety and usefulness at once.
A modern AI a child can actually use
Inside boundaries the parent controls.
Recommended policy profiles by age the parent can pick and tailor — a planned product layer on the engine.
Usage and continuous-time limits keep the tool from becoming addictive.
The child stays on modern, capable AI — protected, without being cut off from the tool of their generation.
The teacher governs the AI in real time
Scoped to the lesson.
The teacher sets the lesson topic dynamically, and can cap usage time.
Uploaded lesson materials define what the AI is allowed to draw on.
Every student's AI stays inside the lesson — no off-syllabus use, no distraction.
Where we are today
A governance market, not an app
Every organization that lets people use AI needs a policy layer. Education is the entry point — the category is far larger.
≈40% CAGR — regulation is the driver.
≈31% CAGR.
will be covered by AI regulation by 2030 (Gartner).
From “we have an AI policy” to “our AI follows it”
For a school system, community or enterprise.
Built by someone who has done this before
Deep, hands-on experience in exactly this problem — content governance at national scale.
Content-governance at scale: former technology-compliance lead for a major religious authority (oversight of 100,000s of devices); ex-Chief Product Manager, Hadran (100,000+ Android devices); CEO of Safer (tens of thousands of users).