How an AI agent helps us stay on top of the paperwork

Every small business has paperwork that never feels urgent. For us, it’s capturing expenses: reading each invoice, typing it into the accounting system and filing the receipt. It matters, but it’s slow, repetitive work that takes time away from clients.

So we built an AI agent to help. It does the reading, checking and filing, and we stay in charge of the decisions. Every morning it captures new supplier invoices, flags anything unusual and leaves us a short summary to review.

Here’s how it works, how we keep it safe, and what it still can’t do. If the admin side of your business keeps getting in the way of the real work, this is for you.

How it works

An AI agent isn’t a chatbot. A chatbot answers questions. An agent does the work: it reads the invoice, checks the records and captures the expense, following rules we wrote for it.

Here’s its daily routine at a glance:

How an invoice gets captured

1. Invoice arrives By email, or a photo from a phone
2. Agent reads it Supplier, date and amount
3. Checks the rules Duplicates, categories and red flags
4. You approve Anything new or unusual waits for you
5. Filed Expense captured, receipt attached

In a bit more detail:

1. Invoices come in. Suppliers email them to a dedicated mailbox, or we snap a photo of a till slip on a phone.
2. It checks every morning. A scheduled job looks for anything new and wakes the agent up if there is.
3. It reads each one. It pulls out the supplier, date and amount, and checks it isn’t a duplicate of something already captured.
4. It sorts it. It matches the supplier (or creates a new one) and picks the right expense category, using written rules for what belongs where.
5. It asks before it saves. Normally it shows us a short table of what it plans to capture, and waits for an OK.
6. It files the paperwork. Each receipt is attached to its expense, renamed and stored by supplier and year, so nothing goes missing.

For routine invoices from suppliers we’ve already approved, like the monthly internet bill, it captures them on its own overnight. Anything new, unclear or unusual waits for us, and we get a message on the phone saying it needs a look.

Keeping it safe

Letting software near your finances sounds risky, and it would be without limits. These are the guardrails we put in place, and the ones we’d set up for any client:

• Its own login, with limited rights. The agent has its own user account, not ours. It can capture expenses and suppliers, but it can’t change settings, manage users, send emails to clients or touch billing.
• Approval before changes. Anything outside the routine needs a yes from a person first.
• A written rulebook. Its job, our business details, the expense categories and its limits live in one document it reads at the start of every session. Nothing depends on anyone’s memory.
• It treats documents as data, not instructions. If an email or invoice contains text trying to tell it to do something, it flags it instead of acting on it. That protects against a new kind of scam aimed at AI tools.
• It can’t delete things. The one clean-up task it does (archiving old receipts once they’re safely filed) runs as a single narrow step that we approve, and it keeps a record of exactly what it moved.
• Every change can be undone. Each change it makes is logged and backed up.

It also caught a mistake we didn’t know we had. While working through the old records, it found that some past entries had been captured the wrong way, which had been making our profit look higher than it was. We fixed them.

Giving it a filing cabinet

After the first week, we noticed a gap. The receipts were filed in folders, but only the agent really knew what was in each one. Other paperwork, like insurance letters, warranties and tax documents, had no proper home at all.

So we added a digital filing cabinet: a searchable document system that every scan, PDF and photo goes into. A separate, smaller AI tool running on our own server reads each document and labels it. The agent then only deals with the ones meant for the books. Each part has one job:

1. Filing cabinet Stores every document and makes it searchable
2. Reading tool Reads scans and photos, and labels them
3. AI agent Makes the bookkeeping decisions and captures the expense

Privacy was the main design rule. Documents are private by default, and the agent only sees one when it’s marked for the books at upload. Personal paperwork stays out of its reach entirely.

The results:

• 380 of 392 previously filed receipts were matched to their existing expense records automatically. The other 12 were zero-value invoices, one duplicate and one that needed a human to check.
• Everything is searchable, even handwritten till slips.
• A short summary arrives every morning of what was filed and what needs a look.

It isn’t perfect, and that’s the point of the checks. Early on, the reading tool labelled a new gas geyser as a utility bill, and read a handwritten R550 as R556. So every document now goes through simple sums (do the line items add up to the total?), and anything that fails, or any photo of a handwritten slip, waits for review instead of being filed automatically.

What it can’t do

An honest list, because AI gets oversold:

• It isn’t an accountant. It captures transactions accurately. Decisions about tax treatment, like how to handle a home office, a vehicle or a big equipment purchase, still belong with your accountant.
• It needs clear rules. It’s excellent at routine work it has been taught. A new supplier or an odd invoice still gets passed to a person.
• Handwriting is still hard. That’s why handwritten slips are always checked before they’re filed.
• It needs setting up properly. Connecting it to your systems, writing the rules and testing it on real documents is where most of the effort goes. Done carelessly, an agent can make a mess quickly.

Could this work for your business?

If a task in your business is repetitive, follows rules and lives in email or documents, an AI agent can probably take care of the repetitive part, so your team can focus on the work that needs a person. Bookkeeping is just one example. The same approach works for:

• answering and booking enquiries from your website or WhatsApp
• logging support requests and keeping clients updated
• reading quotes, statements and forms, and filing them where they belong

The lesson from running one ourselves is simple: the AI is the easy part. What makes it work is the setup around it, meaning limited access, approvals, clear rules and checks. That’s the part we build for our clients.

If there’s a job in your business that keeps getting pushed to tomorrow, tell us about it. The first conversation is free, and we’ll tell you honestly whether an agent is a good fit. You can also read more about our AI agent service.