What AI can — and can’t — access.
Plain English. Enforced automatically.
Procurement teams ask this question first. Here is the plain-English answer - with protections enforced automatically on every request, not left to chance.
Four quadrants. No ambiguity.
Read or write. Allowed or blocked. Every AI surface across the platform follows the same boundary contract.
AI can see
- Your organization's records (contacts, grants, documents, notes)
- Published global content (TappHQ help articles, product documentation)
- Anonymized tool help text (how to use a feature, not who else uses it)
- External general knowledge from the underlying model (no private data)
AI cannot see
- Data from any other TappHQ customer's organization
- Raw connection credentials for your integrations (kept encrypted and never exposed to the AI)
- Platform admin internal dashboards or other customers' audit logs
- Billing/payment details (Stripe-hosted; tokenized references only)
- Personal information about employees that isn't already stored in your own records
AI can change - with approval
- Draft documents, emails, and social posts (you approve before send)
- Create records (contacts, grants, notes) - with inline confirmation
- Update fields on existing records - with diff view and undo
- Trigger integrations (HubSpot sync, Google Calendar) - with explicit approval
AI cannot change
- Billing and subscription state
- Permission and role assignments (admin-only UI, not AI-accessible)
- Platform-wide settings outside your organization
- Other customers' data (each organization is completely walled off from every other)
How the boundary holds up
Boundaries that rely on people remembering to apply them eventually fail. Each of the four protections below runs automatically on every request, everywhere in the platform.
Organization boundary check
Every AI action is checked against the organization it belongs to before any data is read. If a request ever tries to reach another organization’s data, it is refused outright - nothing is returned and no lookup happens.
Field allowlist
Each AI tool can return only a specific, pre-approved set of fields. Anything outside that list is removed before the AI ever sees it, so no extra data is exposed by accident.
Confirmed writes
When an action is triggered by incoming data - like an email or an automated message - it is automatically held for review. Nothing is created or changed until a person explicitly approves it.
Scoped admin tokens
On the rare occasion a TappHQ administrator needs cross-organization access for support, it requires a short-lived, single-purpose approval that expires within minutes and records the reason. Every such access is written to a tamper-evident audit log.
Additional protections - including defenses against manipulated inputs, automatic screening of AI responses for sensitive data, and safeguards that stay in place across an entire conversation - are described on our AI Safety page.
Need the implementation reference?
We publish boundary-test scripts, reproducible failure cases, and architecture diagrams under NDA. Email the compliance team and we’ll route you to the engineering owner.