An email assistant for bookkeepers with no admin help

An email assistant for bookkeepers does the reading and the first drafts an admin would do. It sorts the morning so a client’s question sits above forty bank and payroll notices. It boils a long thread down to what’s outstanding. It drafts the routine replies and the monthly chase, and it puts emailed dates on your calendar. You still decide what a charge is, what goes out about a client’s money, and anything that touches bank details.

That split is what makes it safe. Three controls hold it in place. Drafts wait for your review until you choose otherwise. You know what the tool reads and where that data goes. And there’s a record of what it did. This page covers what to hand over, what to keep, and how to set those controls before it reads a single client’s mail.

What an email assistant for bookkeepers can take off your desk

Without admin help, you’re the one who opens every notice, reads every thread and types every “just following up”. Most of that is reading and routine writing. Those are the parts an AI email assistant does well.

Each row has a guide of its own in this set. The point here is the pattern. The assistant does the first pass, and you do the last look.

What stays with you

Some of the work only you should do. An assistant can still help with the reading on these, but you make the call and you write or approve every word.

• What a transaction is. Whether a $640 charge is supplies or the owner’s lunch is your judgment. It goes in the books under your name.

• Anything that changes where money goes. A vendor’s “please update our bank details” is how business email compromise starts. The FBI’s Internet Crime Complaint Center is plain about it (IC3, September 11, 2024):

  Use secondary channels and/or two-factor authentication to verify requests for changes in account information.

  That means a call on a number you already had. No reply drafted from the email itself settles it.

• Files with bank or payroll data. How those travel is its own decision, covered in bank data in a thread.

• Bad news and fee talk. A late-fee warning, a scope change or “we found a problem in last year’s books” goes out in your words.

• A new client’s first message. You don’t know their tone yet, and neither does the assistant.

Microsoft says the same about its own tool. Its Copilot privacy page notes that AI responses “aren’t guaranteed to be 100% factual” and that users “should still use their judgment when reviewing the output before sending them to others” (Microsoft Learn, checked October 4, 2026). That’s true of every assistant, including the one you pick.

A Thursday, one desk, no admin

Here’s what a day looks like with the reading handed off. You keep books for fourteen small businesses. It’s the 9th, so most statements are in and a few aren’t.

• 7:50. The briefing is waiting. Three items need you. Fenwick Landscaping asks about a $1,260 charge. Calder Physical Therapy’s payroll provider says Friday’s run needs approval by noon tomorrow. Ridgeway Bakery’s bank feed stopped syncing on Tuesday. Below that sit 41 notices you won’t open.
• 8:05. You open Fenwick’s thread. It’s 19 messages long. The summary says the client sent August receipts on the 3rd, you asked about two charges on the 5th, and one is still open. The draft reply asks for the receipt. You change one line and send it.
• 8:20. Calder’s payroll notice has become a task due tomorrow at noon. You approve the run in the payroll software. That part was never email’s job.
• 8:30. Ridgeway’s feed needs the owner to sign in again at the bank. The assistant drafted a short note asking them to. You add “no rush before Monday” and send it.
• 10:15. Marsh Electric’s office manager asks to change the bank account for a regular vendor payment. The email reads fine. You don’t reply to it. You call the vendor on the number in last year’s file. The vendor didn’t send it.
• 2:00. Two statements are still missing. The assistant noticed and drafted the reminders. You send one and hold the other, because that client mentioned a family emergency last week.
• 4:30. A client asks for “a quick call about September”. The assistant offers two times from your calendar. You pick the one after your close work and approve it.

You wrote about four sentences of new text all day. Every message that left had your eyes on it. And the one email that needed a phone call got one.

Three controls to set before it reads client mail

An assistant that reads a dozen clients’ mail needs the same things you’d ask of a person you hired. You want to approve their work at first. You want to know what they see and where it goes. And you want a record.

Review: a setting for each kind of work

The safest setup decides per kind of work, not once for everything. A good assistant lets you start each kind on “suggest” or “wait for me” and move it up as it earns trust.

Some bookkeepers never move replies past “draft”. That’s a fine place to stay. More on how a dial like this works is in the dial, and approvals covers what waiting for you looks like.

Data handling: what it reads, and where that goes

Before you connect anything, get plain answers to a few questions. Each one is answerable from the vendor’s own documentation.

• What does it read? Your whole mailbox, or only what you point it at? Attachments too?
• Is your mail used to train its models? Get the answer in writing, from the vendor’s own page.
• Who else processes it? Most assistants send text to a model provider. The vendor should name them.
• What happens to mail from strangers? A message from an unknown sender can carry instructions aimed at the AI. Ask whether that mail is kept away from it.
• How do you disconnect, and what’s deleted when you do?
• What about the mailboxes clients gave you? A client’s mailbox is the client’s. Ask before connecting an AI tool to it. A mailbox per client covers those arrangements.

The two big suites publish their answers. Microsoft says “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs” (Microsoft Learn). Google says “Workspace does not use customer data for training models without customer’s prior permission or instruction” (Google Workspace Privacy Hub). Both were checked October 4, 2026. A smaller vendor should be able to point you to a page that says as much.

If you prepare returns, the FTC Safeguards Rule likely covers you, and an AI assistant is a service provider under it. Whether a books-only practice is covered is set out in bank data in a thread. For those it covers, the rule on vendors is short (16 CFR 314.4(f)):

Oversee service providers, by: (1) Taking reasonable steps to select and retain service providers that are capable of maintaining appropriate safeguards for the customer information at issue; (2) Requiring your service providers by contract to implement and maintain such safeguards; and (3) Periodically assessing your service providers based on the risk they present and the continued adequacy of their safeguards.

In practice that’s three things. Choose a tool whose security you’ve checked. Keep its terms on file. Look again once a year. The wider question of AI and client confidentiality is in is it safe to use AI with client financial data?, and what to check before connecting is a ten-minute list.

The record: what it did, and when

A year from now a client may ask, “Did you ever get our June statement?” You want the answer in two places: the thread and a log.

• Sent from your own address. Replies should go from your real mailbox, so they sit in Sent and thread under the client’s message like anything you wrote.
• A log of what the assistant did. Each label, draft, task and send, with a time. You should be able to read it without asking support.
• A way to reverse it. A misfiled thread or a wrong task date should take one click to undo. A sent email can’t be pulled back off someone else’s server, which is why sends start on review.
• The answer kept where it’s needed. When a client explains a charge, the explanation belongs on the transaction in your accounting software. Explaining a charge covers that step.

What logs and undo cover goes further on this.

Your first two weeks with an assistant

Starting slowly costs little and tells you a lot.

1. Days 1 to 3: reading only. Connect your own firm mailbox. Let it sort and summarize, and change nothing else. Each morning, check the briefing against the inbox. Did anything that mattered sink?
2. Days 4 to 7: drafts. Let it draft replies, and send none unedited. Note what you change. Wrong figures are a reason to stop. Wrong tone usually fixes itself as it learns your voice.
3. Week 2: tasks and chases. Let it suggest tasks and draft the month’s document reminders. Check every date it picks against the client’s close.
4. After the first close: decide, kind by kind, what moves up. Connect client-provided mailboxes only once those clients have said yes.

If the first week leaves you checking more than you used to read, it isn’t the right tool. Disconnect it and try another.

How Point works as your assistant

Point is an AI email client for Gmail and Microsoft 365. You connect the accounts you already have and keep your address, your history and your contacts. Your clients see no change.

On that Thursday, here’s what Point does.

• The morning arrives sorted. Point weighs how important and how urgent each message is. Fenwick’s question sits at the top. The notices sit below it.
• A briefing gathers what needs you, on the schedule you choose.
• Every thread carries a plain summary, so the 19-message thread reads in a line.
• A request becomes a task. Calder’s payroll deadline lands on your list with its date.
• Drafts wait for you by default. Point writes the reply in your voice. The decision to send stays yours until you raise the dial for that kind of work.
• The dial is per kind of task. It runs from suggest-only to fully handled, so sorting can run free while replies wait.
• A log shows every action Point took. Anything Point did can be approved, turned down or reversed.
• Mail from senders Point doesn’t recognize is held at the door. It isn’t read into anything, so a stranger can’t instruct your inbox.

Point works on the mail. Your ledger, bank feeds and payroll runs stay in QuickBooks, Xero and the payroll software. Point is in private beta with no published price, and access is by waitlist. The rest of what it does is on the benefits page.

When a person, or a smaller tool, fits better

• The work is on the phone or on paper. Calling clients, opening mail and scanning receipts need a person. A part-time virtual assistant fits that.
• The work is in the ledger. Matching receipts and categorizing transactions belong to your accounting software’s own automation, not an email tool.
• You have five clients and a quiet inbox. The AI already in your suite may be enough. Copilot in Outlook can summarize a thread from a “Summary by Copilot” button at the top of it (Microsoft Support), and Gmail’s Gemini panel offers the same kind of help. They work one thread at a time.
• Two or three people answer one address. A shared inbox gives each message an owner, and that’s a different purchase.

Common questions

Can an AI email assistant send emails for me?

Some can, once you allow it. Start every kind of reply on “draft, you send”. Move a kind up only after you’ve stopped editing those drafts. Anything about money, bank details or bad news should stay with you.

Is it safe to let an AI read my clients’ emails?

It can be, with the right tool and settings. Check that the vendor doesn’t train on your mail, names who else processes it, keeps unknown senders’ mail away from the AI, and logs what it does. Ask clients before connecting a mailbox they gave you.

Will an email assistant categorize transactions for me?

No. An email assistant works on the mail. It can find the client’s explanation of a charge and summarize it. The category decision stays with you, in your accounting software.

Do I need my clients’ permission to use an AI email assistant?

For your own firm mailbox, it depends on your engagement terms and any rules that cover your practice. Many bookkeepers mention it in their engagement letter. For a mailbox a client set up for you, ask first, because it’s their account.

Is a free AI tool like ChatGPT enough?

A general chatbot isn’t connected to your inbox. You’d paste client mail into it, which is a data decision you make for the client. An email assistant reads the mailbox in place, under terms you can check.

How is an email assistant different from a virtual assistant?

An email assistant is software that reads and drafts inside your mailbox. A virtual assistant is a person who can also call clients, chase paper and work in other systems. Many solo bookkeepers start with the software and add a person when phone and paper work grows.