Trust and control

Nothing goes out without approval

Your assistant and your AI employees can read from the tools you connect while they are working for you. Anything that changes something outside BGP AI — an email, an invoice, a post, a booking, a payment — is prepared as a task and waits for approval before it happens.

• You choose who approves. By default a person approves, and the owner decides which team members may approve which kinds of work. An administrator can instead hand routine approvals to Board AI, which checks each action against your business and your standing rules and records a written reason.
• Some actions always wait for a person. Payments, refunds, ads, deletions, permission changes and anything rated high-impact go to a person even when Board AI handles approvals. An action the system does not recognise is treated as high-impact.
• Approved means what you saw. When work is planned in steps, what runs is checked against the version that was approved: the plan, the account and the details. If anything changed after approval, it does not run.
• Extra caution when you ask for it. When you hand over a piece of work or set up a schedule, you can ask to be consulted before anything that reaches a customer or spends money, even where it would otherwise be approved automatically.

AI employees stay inside their job

Each AI employee, whether for social media, operations, sales or one of the personal assistants below, comes with a written job: what it owns, what it must never do, which tools it may use and how far it may go on its own. We set those limits, and they cannot be widened from inside a workspace.

• Work past its limit waits for a person. If an AI employee asks for something beyond its risk limit, or outside the tools its job lists, that action goes to a person even when Board AI handles approvals.
• It never sees more than the person asking. An AI employee only uses the apps the person talking to it may use, so it is never a way around someone’s access.
• Your team sets the house rules. Your team can rename an AI employee, give it house rules to follow, choose the language it answers in and set a daily credit limit. When it reaches that limit it says so and stops answering until the next day, or until your team raises the limit.
• You can see what each one costs. The Usage page breaks down credits by AI employee.

Personal assistants are private

Two AI employees work for one person rather than for the business. The House Manager looks after the owner’s personal life, and the Personal Assistant helps one person with their own office work.

• Only that person sees the conversation. Not the owner, not an administrator, not the rest of the team. A personal conversation cannot be shared or joined.
• What it remembers stays with that person. Personal notes are never used in business conversations. That person can review and delete them at any time, and nobody else on the team can see or change them, including the owner.
• The House Manager is kept apart from the business. It is not given your business data or your business tools.
• Costs are shown without content. Personal assistants use the team’s credits. The Usage page shows their total as a single line, with no names and nothing that was asked.

You decide who can do what

• Roles you can shape. Every workspace has an Owner, Admins and Members. The owner can also build roles of their own on top of Admin or Member, choosing exactly which permissions, connected apps and AI employees come with each one.
• Exceptions for one person. The owner can give one person more or less than their role allows. A removal always wins over an addition, and the owner’s own access can never be narrowed.
• Only the owner changes access. Roles, individual exceptions and who may approve which work are the owner’s to set.
• Every change is on the record. Access history shows who changed what for whom, and exactly what was given or taken away.

Stop means stop

• Stop works mid-answer. Pressing Stop ends the answer in progress, even halfway through.
• It does not go in circles. If the assistant repeats the same step and gets the same result, it stops and tells you instead of spending more credits.
• Long work reports back. Work that runs longer than one reply continues in the background only a limited number of times. After that it stops and tells you, rather than running on.

Your data

• Encrypted at rest. Conversations, the inputs and results of every action, what the assistant remembers, the contents of your documents and the skills your AI employees follow are encrypted in our database.
• Credentials are kept from the AI. The tokens for your connected accounts are encrypted separately and never placed in the AI model’s context.
• Isolated to your workspace. Every record carries the workspace it belongs to, and queries are limited to that workspace.
• Retention. We keep your data while your account is active. Where a retention window applies to your workspace, anything older is deleted nightly, and a conversation that ages out takes the documents, voice notes and pictures made in it along with it. Tasks still waiting for approval are never deleted this way.
• AI providers. Which AI providers process your content, and on what terms, is set out in our Privacy Policy.

Honest limits

No AI system is perfect, and we would rather you heard it from us.

• AI can be wrong. Drafts and answers can contain mistakes. That is why actions wait for approval, so read what you are approving.
• What it reads can try to mislead it. An email, a web page or a document can contain instructions written to trick an AI. We do not rely on the AI to resist them. It never holds your credentials, it cannot reach tools you have not connected, and anything it tries to do still goes through the approval route you set.
• Board AI is an AI. If you hand approvals to Board AI, it judges each action against your rules, and it can misjudge. High-impact actions still come to a person.

Reporting a concern

If you think you have found a security problem, or something on this page does not match what you see in the product, email privacy@bgp.ai. A real person reads every message.