GlassBox
Book a walkthrough
Built for EU AI Act Art. 12, MiFID II suitability, FINMA 2023/1

Make every AI answer defensible to a regulator — without slowing your advisors down.

GlassBox sits between your advisors and your AI. Every answer is cited to an approved document, every refusal is logged, every decision is replayable. The audit binder writes itself.

Open the workbenchBook a 20-min walkthrough
Citations on every claimReplayable decision tracesRefuses when evidence is missing
Conversation · C001 · Müller Family OfficeIPS v3.2 · Nov 2024
Advisor · 09:14 Can client move €2M into a single tech stock?
GroundedGlassBox · 09:14 · 312ms

No — IPS §4.1 caps single positions at 25% of AUM1. Current AUM €18M ⇒ max position €4.5M2. The trade is within the single-position limit, but combined with the existing NVDA position would lift tech exposure to ~38%, above the 25% sector cap3. Recommend: escalate.

Grounded · 3 sources · Audit #a91b…View replay →
IPS §4.1 · Concentration cap
IPS §4.3 · Sector concentration
Holdings snapshot · 2024-11-30
Used daily for
  • Mandate breaches
  • Suitability checks
  • Cross-border flags
  • Concentration caps
  • Exclusion lists
  • Sanctions screening
Four controls, always on

Not a smarter AI. A defensible one.

The trust layer that makes AI-assisted advice safe to actually deploy in a regulated firm. Four controls that work whether the underlying model is open-source, hosted, or hot-swapped tomorrow.

Grounded-only answers

Every claim cites a real document in your approved corpus. Unsupported sentences are stripped before they reach the advisor.

Refuses to guess

When the documents don't cover it, GlassBox refuses and routes to the right escalation desk — instead of inventing a rule.

Self-verifying

Chain-of-Verification re-checks each claim against its source, discards anything that fails, and only then returns an answer.

Replayable forever

Every decision stored with question, sources, claims kept and dropped, trust scores, model and seed. One-click replay for audit.

How it works

One pipeline. Four checkpoints. Zero unauditable answers.

  1. 01

    Advisor asks against the approved corpus

    Threads are scoped to a client. The IPS version, jurisdictions, and risk profile travel with every question.

  2. 02

    Every claim verified against its source

    Chain-of-Verification breaks the answer into claims and discards anything not supported by a retrieved chunk.

  3. 03

    Compliance reviews from a queue, not an inbox

    Flagged decisions land in a sortable grid: claim, decide, escalate. Reviewer outcomes train the grounding model.

  4. 04

    Audit binder writes itself

    Filter the audit log by date, client, outcome — export a regulator-ready PDF in one click.

Without GlassBox vs. with

The difference shows up at audit time.

Capability
Today (raw LLM)
With GlassBox
Source of every claim
Implicit, often invented
Cited inline to approved document
What happens out of scope
Confident hallucination
Explicit refusal with escalation target
Audit binder time
Days reconstructing decisions
Minutes — one filter, one PDF
Reviewer feedback loop
Lives in email threads
Becomes training labels automatically
Determinism check
Not measured
Nightly job, on the dashboard

Turn AI-assisted advice into a documented control.

Open the workbench with a demo dataset, or book a walkthrough with your own IPS.

Open the workbenchTalk to us