Your authorized task data
Only selected information that you are permitted to access may be supplied for the task. Access must be checked before reading and again before returning a result.
AI assistance and your data
Your work. Your permissions.
Your choice.
AI assistance is optional. Using it does not give the assistant access to someone else’s private information, permission to change your accounts, or permission to train a shared model on your work.
Policy version: 15 September 2026. These requirements apply to current and future GridOnBoard AI features. Availability is described below.
Strict access boundaries
Asking is not permission.
The assistant must use your authenticated account and current permissions for each task. A name, job title, prompt or quoted instruction cannot grant access.
Other people’s privacy
Coworkers’ private mail, direct messages and personal files stay private. Shared company information still requires explicit current access. Other businesses’ records must not appear in your answers, summaries or search results.
No permission shortcuts
The assistant cannot grant access, bypass MFA, reveal credentials or change recovery destinations. Instructions found inside a message, document or website are treated as content, not authorization.
Separate choices
Assistance is optional.
Shared training is off.
Choosing to use AI assistance is separate from allowing information to be used for model training.
Complete AI opt-out
The account-wide “Allow AI assistance” control is being qualified and is not yet available in the live console. Turning off will stop new AI requests, including background AI jobs. Work already accepted will finish first, and the control will show “Turning off — finishing current requests” until it completes. Current account and resource permissions still apply throughout. Ordinary Mail, Arbam chat, Calendar and Storage Vault remain usable. A prompt cannot switch AI back on.
No automatic shared training
Emails, chats, files, ERP records and conversations are not automatically added to shared training data or another business’s memory. Using a feature or sending feedback is not training consent. Any future optional use needs a separate, clear choice and a review of the rights to that information.
Current availability
Requested drafting and meeting assistance in a private pilot. Wider summaries, website generation and SEO are planned; account-wide opt-out is being qualified. They will not be enabled on the strength of this policy page alone. You can choose not to invoke the current assistant.
Drafts and actions
You remain in control of the outcome.
An AI response is not proof that a message was sent, a payment was approved or a website was published.
Email and summaries
Review draft wording, recipients and attachments before sending. Summaries should distinguish recorded facts, proposals and completed actions. Verify important dates, amounts and decisions against their source.
Websites and SEO
Generated pages require a validated preview and authorized publication. We do not permit generated scripts to run as infrastructure commands. SEO advice must not invent traffic, scores, testimonials or statistics, copy protected content, or guarantee search rankings.
Changes need authority
Authorized workflows may carry out a requested task, such as scheduling a meeting. AI text by itself cannot authorize extra recipients, payments, permission changes, publication or account recovery.
Improvement and accountability
Reviewed improvements, not uncontrolled learning.
Model improvements must use traceable, permitted material and pass privacy, access and quality checks before release.
Limited feedback
The new feedback pipeline is designed to record an issue category and request reference without copying message contents into training examples. A rating never automatically changes the model or its rules.
Information handling
Selected task context is processed to produce the requested answer. It is not permission for permanent AI memory or shared training. Turning assistance off does not delete the underlying business records or necessary security audit records; those follow their own access and retention rules.
Known limitations
AI can misunderstand content or produce incorrect answers. Access controls, isolation and review reduce risk, but no model or policy can guarantee perfect accuracy or eliminate every attack. Do not enter passwords, secret keys, verification codes or payment-card details into prompts.
For privacy or access concerns, use Help & support. Do not include secrets in a report.
Further reading: OWASP on prompt injection · Google Search guidance on AI content.