Trust & security

What we do
with your data.

Truacta answers questions about your revenue, so the first thing a security team asks is what leaves your boundary. The short answer is: your rows don't. The long answer is this page — including a list of what we don't have yet.

The short version

Truacta connects to your warehouse with a read-only role and computes in place. It returns aggregates and the SQL that produced them. Your rows never leave your warehouse, and the AI never sees one.

It also declines questions it cannot answer correctly, and we publish how often that happens — including the question sets it failed. The refusal contract →

You hold two independent kill switches, and either one ends our access immediately: revoke the warehouse role, or revoke the encryption key that unseals our stored credential.

The question everyone asks first

What the AI is actually sent.

The language model receives your question and the catalog — the names and descriptions of your metrics and dimensions. That is all. It is never sent a row, a value, a customer name, or a result.

It also never writes SQL. It emits a structured intent — a metric, a grouping, filters — which is re-validated against the catalog and then compiled to parameterised SQL by ordinary code. A model that cannot emit SQL cannot invent a number. If the intent names a metric that doesn't exist, or the model isn't confident enough, the answer is a refusal rather than a guess.

Our AI sub-processor

ProviderAnthropic (Claude). It is the only AI sub-processor in the request path.
ModelClaude Haiku 4.5.
What is sentYour question, plus metric and dimension names and descriptions. Never rows, values or results.
TrainingNo. Anthropic's commercial API terms do not train on customer data.
RetentionAnthropic's standard commercial retention. Zero-data-retention is available under a commercial agreement and we have not yet completed one — see “what we don't have yet”.
RegionInference runs on Anthropic's US infrastructure. We do not currently offer region pinning.
Human reviewNo one at Truacta reads your questions or answers except when you ask us to investigate something, and we will say so.
Change notice30 days' written notice before we add or replace an AI sub-processor. Model versions from a provider already named here may change without notice — the company handling your questions will not.

What we store, and what we never store

We store: your warehouse connection credential, sealed with envelope encryption under a key you control; the computed aggregates and the exact SQL behind them; and an audit record of every query, per tenant.

We never store: raw rows from your warehouse, or personal data about your customers. The engine is built to return aggregates and lineage — asking it for a list of individual customers is refused by design, not by policy.

Kill switches

You can end our access without asking us.

Revoke the warehouse role. Access stops at your boundary, on your side, immediately — no ticket, no waiting on us.

Revoke the encryption key. Our stored credential becomes cryptographically unreadable. We fail closed with a clear error rather than falling back to anything.

Answers to the rest of the questionnaire

Can we set retention to zero?
On our side, yes — audit records and cached aggregates are configurable per tenant, including off. At the model provider, not yet; see below.
Do you emit audit events we can stream?
Every query is recorded per tenant with the SQL, timestamp and whether it was allowed or denied. Export on request, and on termination.
Do you sign a DPA?
Yes — a GDPR Article 28 data processing agreement. We don't publish a standard form yet; ask and we'll send terms.
How fast do you report an incident?
Within 72 hours of confirming a security incident that affects your data, in writing, with what we know at that point.
Is this an AI system, and do you say so?
Yes. Answers are produced with the help of a language model, disclosed in the product and here — the EU AI Act's Article 50 transparency duty, which has applied since 2 August 2026.
Can it make decisions about people?
No, and our acceptable-use policy prohibits using its output to make consequential decisions about individuals — including pay, quota and headcount.
Said plainly

What we don't have yet.

A product whose whole claim is that it tells you when it doesn't know should be equally direct about itself.

Last reviewed 7 September 2026. Questions a page can't answer: hello@truacta.ai.