Would you give a new hire your Gmail password on their first morning? Or the login to your bank, your payroll and your Shopify store, all at once, before you'd seen a week of their work?
That's roughly the choice the new wave of AI employees puts in front of you. On October 21, Anthropic opened a waitlist for Claude Duet, an app where Claude "works as a proactive agent." It connects to tools like Gmail, Slack, QuickBooks, Stripe, Gusto and HubSpot, keeps working after you step away and asks before it acts for you. A couple of weeks earlier, Google introduced a Gemini agent that "receives its own Workspace account, including an email address, calendar, Drive, and presence in your company directory." Google's own words for how that identity is handled: "governed like an employee."
You're probably not starting from zero. The U.S. Chamber of Commerce's 2026 small business technology report says 66% of small businesses now use AI, nearly 3 times the share in 2023. What's new is AI that does things in your accounts instead of only answering questions.
This guide is designed to:
- Explain what an AI employee is and what the new tools can do.
- Show you how to treat one like a new hire, with a job, access and a manager.
- Walk you through a 30-day trial you can measure.
- Give you rules for approvals, money and offboarding.
After reading it, you will be able to:
- Write a 1-page job description for an AI employee.
- Decide which accounts it may touch, and which it may not.
- Set the approvals a person must give before it acts.
- Run a 30-day review and decide whether to keep it.
The basics: what AI employees are (and what they cost)
"AI employee" is a marketing phrase, but it points at something real. It's an AI agent that works across your apps on its own schedule, does multi-step tasks and reports back. A chatbot waits for you to type. An AI employee checks your inbox every few hours and tells you what needs you.
Here's how the 2 announcements compare, from the vendors' own pages:
| Claude Duet | Gemini agent | |
|---|---|---|
| What it does | Morning brief, check-ins, tasks across connected apps | Long-running work across Workspace and other systems |
| Its identity | Works through apps you connect | Its own Workspace account and email address |
| Who can get it | US Pro and Max subscribers, by waitlist (not Team or Enterprise yet) | Google's post doesn't say |
| Cost | Included in Pro and Max during the preview, uses your plan's limits | Google's post gives no price |
| Safety rails | Asks before sending, paying, signing or deleting | Role-based permissions set by your admins, audit trail per agent |
Sources: Anthropic's announcement, the Duet help page and Google's Gemini at Work post, all read in October 2026.
Both vendors are candid about limits. Anthropic says Duet "still might misread a thread or miscalibrate what's urgent," and the help page notes it checks accounts at set intervals, so an urgent item can arrive between checks. Microsoft 365 isn't connected yet either.
OK, so the tools exist and some are already in plans you may pay for. How do you bring one in without handing it the keys to everything?
Step 1: Write the job description before you sign up
Larger companies don't hire anyone without a role description and a manager. Give your AI employee the same, on 1 page.
Write down:
- The job. 1 or 2 tasks, in plain words. "Every morning, list unpaid invoices over 30 days old and draft a polite reminder for each." Not "help with admin."
- The manager. The 1 person who reviews its work and answers for it.
- The hours. Duet runs a morning brief and checks in every few hours. Is that what you want, or only a weekly report?
- The success test. What would make you keep it after 30 days?
Ask yourself:
- Which task do I do every week that follows the same steps each time?
- Would I trust a temp with this task on their first day?
- If it gets this task wrong, who notices, and how fast?
- Does the task touch money, contracts or customer personal details?
Pause and reflect: Look at the task you picked. If a new hire did it badly for a week, what would it cost you? If the answer scares you, pick a smaller task.
Step 2: Give AI employees the least access that does the job
Security teams in big companies work from a rule called least privilege. NIST, the US standards body, defines it as the principle that each entity "is granted the minimum system resources" it needs. For an AI employee, that means it can open the 2 apps its job needs and nothing else.
Google's post describes agents with "least-privilege permissions" and says a coworker agent "sees only what you share with it." Duet works through the connectors you set up, and you start every connection yourself. Both give you the controls. You still have to decide.
Make an access list, adapted from the NIST least privilege principle:
| App | Job needs it? | Access | Notes |
|---|---|---|---|
| Gmail (your inbox) | Yes | Read, draft | No sending without approval |
| QuickBooks | Yes | Read | Invoice list only |
| Bank | No | None | |
| Payroll (Gusto) | No | None | |
| Shared Drive | Only 1 folder | Read | Share the folder, not the whole drive |
If your plan supports a separate account for the agent (Google's version has its own email address), use it. Then its actions show up under its name in your logs, not yours. Google says every action "is written to an audit trail and attributed to the agent rather than to a person."
Step 3: Set the approvals a person must give
A new hire asks before spending money. Your AI employee should too.
Duet's design is a useful model. Anthropic's safety page says that before sensitive actions it "shows you the action, the destination, and the content, including every recipient." Some actions are off-limits even with approval, such as trading securities and moving money between accounts. Purchases use a one-time card tied to 1 merchant and amount.
Write your own approval rules, whatever tool you choose:
- A person approves every email to a customer or supplier.
- A person approves every payment, at any amount.
- Nothing is deleted during the trial.
- Any request to change bank details gets a phone call to a number you already had.
That last one matters more with an agent reading your mail. Anthropic is open that "these safeguards will occasionally miss something," and an approval screen only helps if the person reads it. Read the recipient list every time (yes, every time).
Pause and reflect: Who in your business will approve the agent's requests? If they're out for a week, who covers?
Step 4: Run a 30-day trial with a scorecard
Big organizations agree on success measures before a pilot starts. Do the same on 1 sheet, and fill it in every Friday.
| Measure | Week 1 | Week 2 | Week 3 | Week 4 |
|---|---|---|---|---|
| Tasks it finished that you'd have done yourself | ||||
| Approvals you rejected, and why | ||||
| Things it flagged as urgent that weren't | ||||
| Things it missed that were urgent | ||||
| Minutes you spent checking its work |
Watch the costs too. Duet's background checks count toward your plan's usage limits, including the 5-hour session and weekly limits, according to the help page. If your own Claude use hits a limit mid-week, the agent is part of the reason. Google's post describes spend caps that pause an agent when a project hits its limit.
Step 5: Plan the offboarding on day 1
Large IT teams have a leaver process: when someone goes, every account is closed the same day. AI employees need one too, and it's easier to write while you still remember what you connected.
Keep a short record:
- Every app you connected, and the date.
- Where to disconnect each one. Duet's help page says you can "disconnect any app at any time," and that stops access from that point on.
- What it remembers. The same page says Duet keeps its memory of past information after you disconnect, so check how to reset or delete it.
- Any separate agent account to close, if your tool created one.
Review the list once a month, even if the trial goes well. Vendors are changing these products quickly (Duet's approval settings and sign-in methods are "still changing," per Anthropic), so a feature you checked in October may work differently by December.
Questions owners ask
Will an AI employee replace a person on my team?
The tools today handle tasks, not whole roles, and both vendors describe a person approving or sharing what the agent can touch. The Chamber's report says 47% of small businesses say AI is adding jobs and 6% say it has enabled headcount cuts.
Is it safe to connect my business accounts?
It depends on which accounts and what the agent may do in them. Start with read-only access to 1 or 2 low-risk apps, keep a person approving every outgoing message and payment, and check with your accountant or lawyer before an agent touches payroll, tax or client records.
Can my whole team use Duet?
Not yet. The help page says the preview is for US Pro and Max users through a waitlist, and Team and Enterprise plans aren't supported.
Concluding thoughts: start with the job description
AI employees are worth trying once you treat them like any new hire, with 1 clear job, the least access that does it and a person who signs off on anything that leaves the building. Go back to Step 1 and write that 1-page job description this week.
If you'd like a hand, our audit can check which of your accounts are safe to connect, and our automation work can set up the agent, its access and its approvals so the trial starts on solid ground.