The experts are AI
The experts are AI — named roles, not people.
That sentence appears, once, wherever you first meet an expert: at the top of every consult room, on the free answer at /try, above every shared page, in the footer of the emails that check in on a position, and as a disclosure field in what an AI assistant receives through our MCP server and API. It is said once per surface, not under every reply.
Names like “The Lawyer” or “The CFO” describe a seat and a method, not a person or a credential. No expert carries a professional title, and none claims a licence or a school.
Which model sees what
Each expert seat runs on a model from one of the providers on the subprocessor list — 30 active models today. A seat receives what its answer needs: your message, the room so far, and the facts from your World and your connected tools that bear on it. Nothing else of yours is sent.
- Sensitive facts — anything marked personal or confidential — render into a prompt only inside the context they belong to, and only when every seat reading that prompt is verified zero-retention and not hosted in mainland China. The check fails closed: a model with no verified flag counts as retaining.
- As of today no model in the catalog is marked verified zero-retention, so a fact marked sensitive is not sent to any model at all. It stays in your World, where you can read and edit it.
- Models that keep data for a period never take a seat in a room where sensitive facts are in play: Claude Fable 5.1.
- Models served from mainland China: none.
- Training: we use every provider through its business API, under terms that do not allow training on what we send.
- Web research sends a research question derived from your request — not your file — to the search provider.
- Voice: your browser transcribes; we run no speech service and never receive the audio.
What each answer worked from
An expert’s position shows its basis: what it worked from (your words, a fact on file with its date, a number it computed from your figures, a source it read), what it did not know, and what it assumed. When the basis is thin, the stated confidence is capped by code, and the expert says what it would need before going higher.
A position on the record, not a licensed professional’s advice — and we come back on the date to see if it held.
How positions are later counted is on How the record is counted.
When a matter needs a licensed professional
Some answers are, by law and by our model providers’ own policies, for a licensed person: what to do in your specific legal matter, a diagnosis or a medication, a specific investment product to buy, sell or hold, or what to claim on your own taxes. In those matters the expert answers differently: it keeps its position and what would change it, and its next move is a named handoff — who to see where you live, the free route if cost is the barrier, and the three questions to take with you. It never gives the filing, the dose, the ticker or the deduction.
- Anthropic’s Usage Policy (effective 12 November 2026): legal, medical and finance recommendations to an individual need review by a qualified, licensed person before they are delivered as advice.
- OpenAI’s usage policies (29 October 2025): no tailored advice that requires a licence without a licensed professional’s involvement.
- Google’s Generative AI Prohibited Use Policy (17 December 2024): no automated decisions with a material effect on a person’s rights in high-risk areas without human supervision.
Whether a matter is regulated is the model’s reading of your words; code then holds the line — in Israel, the EU and the US (and wherever we don’t know your country), a legal, health or investment matter is always answered this way. As of the date above this ships behind a switch that is still off in production.
What you write here is not privileged — a court can ask for it.
Who else reads it
No one at SynthBoard reads your rooms unless you ask us to for support or the law requires it. Admin access exists for that and for incidents, behind an admin flag checked in the database. A provider may run its own abuse monitoring under its terms. Nothing about you is decided from your rooms or your record — by us, by a workspace, or by an AI assistant you connected.
Where your data is stored and how to delete it: How we hold your data. The contract terms: Data Processing Addendum.