Trust, data, and compliance
Not account numbers. What a client is afraid of. What they haven't said out loud. What a marriage looks like from the inside when money is involved.
Wealth firms rightly ask hard questions before letting that kind of context near an AI system. This page is our answer, including the parts we haven't finished.
01 · Sanitized first, and tested
Identifying and sensitive details are removed before any analysis begins, in two passes, and we red-team our own sanitization rather than assuming it holds.
02 · Fully traceable
Every brief carries a complete internal record of what produced it, from what went in to what came out. Nothing is unexplainable after the fact.
03 · Advisor-reviewed by design
There is no client-facing mode. Nothing we produce reaches a client without an advisor deciding it should.
04 · Your clients stay yours
Your firm owns its client data. It never becomes part of the intelligence layer other firms benefit from.
It arrives
You share what you already have: meeting notes, CRM history, intake answers, an advisor's own write-up. Through a secure intake, and only what's needed. The system is built to work from limited context, so there's no reason to send more than the situation calls for.
It's stripped
Before anything is interpreted, the material passes through two sanitization passes: an automated pass over known patterns, then a second judgment pass that catches the names, institutions, and medical detail patterns alone would miss. A separate, tighter profile handles machine-generated material like CRM exports and meeting transcripts, where identifiers show up in shapes that ordinary notes don't, including numbers read aloud.
We test this adversarially rather than assuming it works. Our own red-team suite plants known identifiers in realistic client material, then checks every downstream artifact to see whether any survived. In the most recent full run, none did. We measure the opposite failure too: whether sanitization strips so much that the meaning is lost. Nothing that should have survived was missing.
It's interpreted
The sanitized signals are read against our understanding of how hard transitions unfold. This is where the brief is actually formed, and it happens without the system knowing who your client is.
It becomes a document
The brief is assembled, the client's name is put back so it's usable, and it's checked before it reaches the advisor. Runs happen in isolated working spaces rather than a permanent client database.
It stops
The brief goes to the advisor and nowhere else. No client-facing delivery, no third-party sharing, no quiet retention of what came in.
Most recent full red-team run
Tested adversarially, not assumed
An AI that confidently states a client is anxious, grieving, or hiding something is a liability in a regulated business, and it's wrong about people more often than its confidence suggests. So the constraint is built into how documents are made, not left to whoever is writing them.
Every brief separates three things, and cannot be produced without doing so:
Language rules sit on top of that separation, so the system doesn't produce diagnosed-sounding statements about a client's emotional state. And every brief opens by naming what it was built from and what it does not know, telling the advisor what to confirm rather than assume.
The standard we hold it to: better inquiry, not false certainty. The goal is a better question, never a verdict about a person.
Signals, questions, and cautions are possibilities to consider, not conclusions. You remain responsible for how you interpret this and guide the conversation.
Built from: two meeting notes, one intake. Unconfirmed: see "what to check first".
The boxed line is printed at the top of every brief and Deep Dive we produce.
Yours
Your client data. All of it. It stays under your control and is never pooled, shared, or sold. What makes your clients your clients does not become an asset of ours.
Ours
The understanding that reads it: years of work on how these transitions unfold. That's what you're licensing, and it's the same for every firm.
On whether the system learns: it does improve over time, and we'd rather explain how than hide behind a slogan. What improves is our understanding of hard transitions in general, drawn from de-identified patterns. What never happens is your clients' information becoming part of what another firm's briefs are built on. Nothing identifiable about your clients ever trains anything.
| Control | Status |
|---|---|
| Two sanitization passes before any interpretation, plus a tighter profile for CRM exports and transcripts | Running today |
| An adversarial test suite that plants identifiers and measures both leaks and over-scrubbing | Running today |
| Structural separation of known, signaled, and worth exploring, enforced before a document can be produced | Running today |
| Language rules preventing diagnosed-sounding claims about a client | Running today |
| Complete internal audit trail behind every document | Running today |
| Advisor review by design, with no client-facing delivery mode | Running today |
| Data minimization by architecture, and isolated per-run working spaces | Running today |
| Opaque client references, never derived from a client's name, so the system works without knowing who anyone is | Built, switching on |
| The name-to-reference mapping held outside the system entirely, and destroyed once a brief is delivered | Built, switching on |
| A record of what was retrieved and when, kept as a verifiable fingerprint rather than a stored copy of the content | Built, switching on |
| SOC 2 readiness | In progress |
| Role-based access controls | In progress |
| Firm-configurable retention and deletion settings | In progress |
| A formal vendor due-diligence packet, PIPEDA-aware and built for US enterprise review | In progress |
| A hosted client discovery experience, gated behind its own counsel review and privacy testing before any real client uses it | In progress |
"Built, switching on" means the control exists and turns on as each system connection goes live.
We'd rather show you this than let you discover it in a security review. If your firm needs something on these lists before you could proceed, tell us. Founding firms are shaping this roadmap, and several items are on it because a firm asked.
AskGrAIce prepares an advisor for a conversation. It does not make recommendations to clients, does not deliver anything to clients, and does not replace professional judgment at any point. Advisors remain fully responsible for the advice they give and how they use what we provide.
We're a Canadian company serving firms across North America, so we build to both Canadian privacy expectations and US enterprise requirements together rather than treating either as an afterthought.
To the advisor, and nowhere else
If you're the person at your firm responsible for saying no to things like this, we'd rather hear from you early than late. Send the question. We'll answer it plainly, including when the answer is "not yet."