Five controls for sensitive data
01

Discover: field and data classification

02

Authorize: role, project, scope

03

Query: policy enforcement and thresholds

04

Output: detail, export, sharing controls

05

Audit: question, query, result trail

Enforce authorization outside the model

A model may interpret a request, but it must not decide whether a user can see salaries, contact details, or personal identifiers. Identity, data permissions, and policy must make that decision.

Even if a prompt is bypassed or a query is generated incorrectly, the underlying source should reject unauthorized reads.

Classify data and map identity first

Label fields public, internal, sensitive, or restricted, with an owner, purpose, and retention rule. Map enterprise identities to organization, project, role, row scope, and column access.

Anonymized data is not automatically safe. Small groups, unique combinations, and repeated questions can enable re-identification.

Control both query and output

At query time use read-only accounts, allowlists, row and column policies, aggregation thresholds, scan limits, and timeouts. At output time govern detail display, download, copy, external sharing, and notifications.

Charts and reports may outlive a conversation. They should inherit the generation scope and have a policy for access revocation.

Make audit answer four questions

Who asked, when, under which identity, and against which data? What query ran? What class of result returned? Who reviewed or exported it?

Avoid duplicating sensitive values in logs. Store identifiers, summaries, policy decisions, and versions with access, retention, and deletion rules.

Accept with adversarial cases

Test direct requests, synonyms, staged follow-ups, cross-project references, tiny groups, exports, and sharing. Expected behavior is denial, masking, or safe aggregation with a recorded reason.

AskTable.ai confirms organization, role, project, and data-scope controls. Field policy, audit retention, identity integration, and private deployment conditions require explicit confirmation.

Public references

Ready to help your team start?

Talk through a real scenario and see how AskTable.ai can fit your business.

Book a demo