An AI agent for bookkeeping and inbox handles two separate jobs in my setup: Frieda sorts the inbox, manages appointments, and captures receipts, Anton turns that into quotes and writes invoices. Separate on purpose, because each job needs its own rules, its own approvals, and its own knowledge. Here's both of them on a real workday, with the limits I set for them deliberately.
What an agent is in the first place, and how it differs from a plain chatbot, is on the AI agents page. If you want to figure out which of your own tasks even fit an agent before you start with bookkeeping or the inbox, Which Tasks You Can Hand Off to AI helps you pick.
What actually lands in the inbox every morning
A normal inbox for me is a mix: inquiries from potential clients, follow-up questions on ongoing projects, invoices from suppliers, meeting proposals, newsletters, and occasionally something that needs my attention right away. The problem isn't the volume, it's the sorting: before I can answer anything, I first have to figure out what actually matters.
That's Frieda's first job. She reads every incoming email, assigns it a category, client inquiry, receipt, meeting request, follow-up question, or other, and files the right result for each category: an entry in the calendar, a captured receipt, a drafted reply, or a note that I need to handle it myself.
A workday: Frieda in the inbox, Anton on the quote
Here's what one pass actually looks like. An email comes in: a prospect describes what they need and asks for a quote. Frieda recognizes the inquiry, checks whether the necessary details are there, such as scope and a rough timeframe, and if something's missing, she drafts a short follow-up question that I only need to approve or adjust. Once the details are complete, she hands the case off to Anton.
Anton gets the inquiry with all the details, pulls from my price list and past quotes as a style template, and delivers a finished quote draft, structured like my examples, priced by my logic, no line items invented on the fly. I read the draft, correct whatever doesn't fit, and only then does the quote go out, sent by me, not by Anton.
If the client accepts later, the same process runs in reverse: Anton turns the accepted quote into an invoice draft, with the same line items and amounts that were in the quote. On a completely different day, in parallel, a supplier invoice arrives as an attachment. Frieda recognizes it as a receipt, reads out the amount, date, and sender, and files it in a structured way, with a short note if an amount looks unusually high or something is missing.
What the two decide on their own, what I approve
I draw the line clearly between preparing and sending. The table below shows exactly how that's split between the two of them.
| Step |
Who handles it |
Goes out without my approval? |
| Sort and file the inbox |
Frieda |
yes, internal |
| Capture a receipt from an attachment |
Frieda |
yes, internal |
| Draft a reply to a routine question |
Frieda |
no, I read and send it |
| Draft a quote |
Anton |
no, I read, correct, send it |
| Draft an invoice |
Anton |
no, I check line items and amount |
| Draft a payment reminder |
Anton |
no, I decide whether and when |
Internal steps like sorting and filing run without me, because a mistake there usually means a wrongly filed item, which I can fix. Anything that leaves the house, a quote, an invoice, a reply to a client, only goes out after I approve it. That is not caution for its own sake, it follows from how the risk is distributed: an agent that sends without asking saves the most time in exactly the moment a mistake becomes the most expensive.
Rules, limits, and the privacy angle
The same ground rule applies to both, as it does to every one of my AI employees: no invented prices, no invented services, ask instead of guessing when information is missing. For Anton, amounts on invoices have to come exactly from the accepted quote, not be recalculated from scratch. For Frieda, she only files receipts, she never books them herself, because the actual bookkeeping, with account assignment and tax logic, stays with my tax advisor and their software. Frieda and Anton prepare the ground, they don't replace bookkeeping software like DATEV or Lexware, and they don't replace a tax advisor either.
Because quotes, invoices, and the inbox inevitably contain names, addresses, and sometimes payment details of clients, privacy here isn't a side note. Which client data I actually put into an AI system and which I don't, with the three checks I run before every input, is covered in detail in AI and Privacy: What the AI Gets to See. That's not legal advice, it's my own practice; for the legal side, the regulation itself applies, available in the EU regulation text.
Two failure modes the rules guard against
Why the approval rule is this strict shows in two typical failure modes that come with this kind of work.
The first one is about scheduling. Let reply drafts to meeting requests go out directly because the topic seems harmless, and you risk a friendly but wrong meeting proposal the moment two calendar entries get mixed up. That is exactly why every email to a client crosses my desk here, no matter how harmless the topic looks.
The second one is about quotes. Copy a style template too literally and an add-on service from an old quote travels into a new one, even though it was never agreed for that case. So every line item belongs checked individually against the actual request before a quote goes out.
Frequently asked questions
Can an AI agent send invoices automatically?
Not in my setup. Anton drafts the invoice, I check the line items and the amount, and I send it myself. It costs a few minutes per case, but it's worth the control to me for anything involving money and client contact.
Can I submit receipts as photos?
Yes, I usually send receipts as a photo or a scan attached to an email, and Frieda reads out the amount, date, and sender. The actual booking with account assignment still goes through my tax advisor and their software, not through Frieda.
Does this work with Gmail or Outlook?
Yes, the principle works regardless of your inbox provider, as long as you give Frieda access to that inbox. The concrete technical steps depend on the provider, and I wouldn't describe them in general terms here without having checked them for your specific setup.
What happens if an email is misread?
That's exactly what the approval step is for. Frieda and Anton deliver drafts, not finished results, and I read every draft that leaves the house before it goes out. A misunderstanding costs me a correction, not a wrong send.
How many agents does an office need for this?
For me it's two, clearly split by job: one for the inbox, one for quotes and invoices. You can also start with just one and only set up the second once the first runs reliably.
AI agent for bookkeeping: where to go from here
Start with the part that costs you the most time, which is usually the inbox, not the invoice. Ten more examples of how agents actually work in practice are in AI Agents: 10 Real-World Examples. Frieda and Anton are available as ready-made packages, load them and start the onboarding conversation, in my community Claude Practitioners.
