What can an AI employee see by default?

It can only see what you connect. If you give it WhatsApp messages, it sees those messages. If you connect Gmail or Google Drive, it sees the content covered by the scope you approved. It does not magically get the rest of your business. That is why the first security question is always scope, not model quality.

Do I need to give Gmail, Drive, or Calendar access on day one?

No. Start with the smallest scope that lets you test one job. Many teams begin in WhatsApp or Slack, then add Gmail, then Google Drive, then Calendar only if the use case truly needs it. That lets you see whether the output is useful before you widen the blast radius. It also makes the review step easier because there is less surface to inspect.

Is WhatsApp safer than email?

Usually yes, because channel-only work is narrower than full mailbox access. A WhatsApp-first setup can answer, route, and summarize without touching your inbox. But safety depends on the rule set, not the channel name alone. A locked-down Gmail integration can be safer than a sloppy WhatsApp setup, so the real issue is permission design and review discipline.

What data should stay human-only?

Regulated sign-off, bank transfers, legal approval, payroll changes, and anything where a mistake cannot be rolled back easily should stay human-only. If a step changes money, rights, or customer records, keep a person in the loop. That includes refunds above a threshold, contract edits, account closures, and admin changes that affect real people. The clean rule is simple: if the action creates legal or financial impact, do not let software be the final signer.

How do OAuth scopes affect risk?

Scopes are the permission list. Read-only is safer than send, and send is safer than full write access. If a product asks for broad access up front and cannot explain why, that is a signal to pause. The difference between read, send, and write is not semantic; it changes what the system can do if it misfires. A good pilot starts with the smallest scope that still lets the team verify real value.

Does AGI employee mean it should get more access?

No. AGI employee is a capability label, not a permission model. The access question is the same: give the narrowest scope that still lets it finish the job, then widen only when the output is stable. If anything, a stronger system deserves stricter guardrails because it can do more once it is inside the account. Capability and permission are separate decisions.

What should I grant first?

Start with one written task, one channel, and one reviewer. For example: let it summarize a Slack thread, draft a WhatsApp reply, or clean a Google Sheets report before you let it send anything on its own. Once that is stable, expand one permission at a time. The point is to build a clear evidence trail: what it saw, what it produced, and what a human approved. That makes the next step easier to trust.

How do I know if it is storing too much?

Ask where memory lives, who can read it, and how to delete it. A good setup keeps business memory tied to the organization and makes retention explicit. If the vendor cannot answer that plainly, do not move past a pilot. You want a setup where the history of decisions is available for the team, not hidden in a black box. If you cannot explain the retention policy to a teammate in one sentence, the policy is not ready.

What is DoYourWork.AI?

DoYourWork.AI is a platform that provides Perla, an autonomous AI employee designed to handle everyday business tasks across Google Workspace, WhatsApp, Slack, email, and more — so you can focus on what actually matters.

Perla can manage your Google Workspace (Docs, Sheets, Calendar), send and reply to WhatsApp messages, automate Slack workflows, draft and send emails, schedule meetings, and operate autonomously 24/7 without supervision.

How do I get started?

Sign up for an account, add Perla to WhatsApp, and start giving her tasks. No apps to install — she works through the channels you already use.

Absolutely. Your data never leaves your organization — it stays entirely within your own infrastructure. Perla only accesses the tools and information you explicitly authorize, everything is encrypted in transit and at rest, and you can revoke access at any time.

How much does it cost?

DoYourWork.AI is currently in private beta. Request access to try Perla and see what your first AI employee can do.

Which tools does Perla work with?

Perla integrates with Google Workspace (Docs, Sheets, Slides, Calendar, Gmail), WhatsApp, Slack, and more tools coming soon.

Who built DoYourWork.AI?

DoYourWork.AI was founded by Kelvin Tang, who built Perla to handle the overwhelming day-to-day work so founders and teams can reclaim their time and mental health.

Is an AI employee the same as AI staff or an AI workforce?

Yes — "AI employee", "AI staff", "AI workforce", "AGI employee", and "AGI staff" all refer to the same category: autonomous software that owns end-to-end work in the channels your team already uses, instead of waiting for prompts like an AI assistant. Perla is the same product whether you call it your AI employee or your AI workforce. For the full definition, see https://doyourwork.ai/blog/what-is-an-ai-employee.

How is an AI employee different from work automation tools like Zapier?

Work automation tools like Zapier, Make, and n8n run fixed if-this-then-that workflows that you wire up step by step. An AI employee like Perla decides what to do from a goal and a channel — she reads context, drafts replies, books meetings, and routes work end-to-end without you defining every branch. Use Zapier for predictable triggers; use an AI employee for judgment-bearing work.

faq

AI Employee Data Security: What It Sees

What an AI employee can see, what it cannot, and how to grant access safely across Gmail, Slack, WhatsApp, and Google Drive.

By Kelvin TangAugust 26, 20261 min read

Data security is the first real question when you put an AI employee inside Gmail, Google Drive, Slack, or WhatsApp. The issue is not whether it can write a good reply. It is what it can read, what it can send, and how much of your system you gave it. This FAQ is for founders and operators who want the speed without handing over the keys.

ChatGPT and Claude are useful assistants, and Microsoft Copilot is fine for draft work. The risk picture changes when software can send mail, edit files, or post into a live channel. That is the line this page draws.

A clean setup is not about giving the system nothing. It is about giving it one job, one path, and one reviewer, then proving it can stay inside that lane. If the lane works, widen it. If the lane gets messy, tighten the scope before you add more surface area.

For the broader model, read What is an AI Employee? and AI Employee vs. AI Assistant. For a live channel example, see Perla for Customer Support. The access model on the landing page matches the same rule: start small, then widen only when the job proves it needs more.

Frequently asked questions

What can an AI employee see by default?: It can only see what you connect. If you give it WhatsApp messages, it sees those messages. If you connect Gmail or Google Drive, it sees the content covered by the scope you approved. It does not magically get the rest of your business. That is why the first security question is always scope, not model quality.
Do I need to give Gmail, Drive, or Calendar access on day one?: No. Start with the smallest scope that lets you test one job. Many teams begin in WhatsApp or Slack, then add Gmail, then Google Drive, then Calendar only if the use case truly needs it. That lets you see whether the output is useful before you widen the blast radius. It also makes the review step easier because there is less surface to inspect.
Is WhatsApp safer than email?: Usually yes, because channel-only work is narrower than full mailbox access. A WhatsApp-first setup can answer, route, and summarize without touching your inbox. But safety depends on the rule set, not the channel name alone. A locked-down Gmail integration can be safer than a sloppy WhatsApp setup, so the real issue is permission design and review discipline.
What data should stay human-only?: Regulated sign-off, bank transfers, legal approval, payroll changes, and anything where a mistake cannot be rolled back easily should stay human-only. If a step changes money, rights, or customer records, keep a person in the loop. That includes refunds above a threshold, contract edits, account closures, and admin changes that affect real people. The clean rule is simple: if the action creates legal or financial impact, do not let software be the final signer.
How do OAuth scopes affect risk?: Scopes are the permission list. Read-only is safer than send, and send is safer than full write access. If a product asks for broad access up front and cannot explain why, that is a signal to pause. The difference between read, send, and write is not semantic; it changes what the system can do if it misfires. A good pilot starts with the smallest scope that still lets the team verify real value.
Does AGI employee mean it should get more access?: No. AGI employee is a capability label, not a permission model. The access question is the same: give the narrowest scope that still lets it finish the job, then widen only when the output is stable. If anything, a stronger system deserves stricter guardrails because it can do more once it is inside the account. Capability and permission are separate decisions.
What should I grant first?: Start with one written task, one channel, and one reviewer. For example: let it summarize a Slack thread, draft a WhatsApp reply, or clean a Google Sheets report before you let it send anything on its own. Once that is stable, expand one permission at a time. The point is to build a clear evidence trail: what it saw, what it produced, and what a human approved. That makes the next step easier to trust.
How do I know if it is storing too much?: Ask where memory lives, who can read it, and how to delete it. A good setup keeps business memory tied to the organization and makes retention explicit. If the vendor cannot answer that plainly, do not move past a pilot. You want a setup where the history of decisions is available for the team, not hidden in a black box. If you cannot explain the retention policy to a teammate in one sentence, the policy is not ready.

Hire your first AI employee

Perla handles your Google Workspace, WhatsApp, Slack, email, and more — so you don't have to.

See what Perla does