Each one only takes a few minutes. The trouble is they never stop, and the person doing it usually has better things to do. Then come the typos: a transposed digit, a duplicate entered twice, a missed due date. None of it is dramatic on the day. It all shows up at month end, when someone has to find it.

The shape of it

When I build one of these, it nearly always looks the same:

1. Watch the place invoices arrive: a shared inbox, a Google Drive or OneDrive folder, or a supplier portal.
2. When a new invoice comes in, AI reads the PDF or photo and finds the fields you care about.
3. Run checks: does the total add up, is the supplier known, has this invoice number been seen before, is the tax sensible?
4. If everything passes, save the data to your sheet or accounting software and file the original.
5. If something fails, flag it for a person with a clear note on what looked wrong.

The AI part is reading the document. Everything around it is plain, predictable rules. That mix is what makes it reliable.

Why AI changed this

Older tools needed a template for every supplier layout and broke the moment a supplier redesigned their invoice. AI reads a document more like a person does. It understands that "Amount due", "Total payable" and "Balance" are probably the same thing, even in different corners of the page.

It isn't perfect. Blurry phone photos and handwriting are harder, which is exactly why the checks and the human review step exist.

What gets flagged

Anything unusual goes to a person instead of being guessed:

• a supplier you've never paid before
• line items that don't add up to the total
• an invoice number that looks like a duplicate
• a missing field, like no due date or no tax number
• an amount well above what that supplier normally charges

Someone approves or fixes those with a click, and the rest flows through. The automation does the typing. You still decide what gets paid.

A typical week

A small online shop gets around 40 supplier and order emails a day, most with a PDF attached. Today someone types each one into a Google Sheet and later re-types it into the accounting software.

With the automation, each PDF is read as it arrives, the row appears in the sheet, the bill is drafted in the accounting software, and the PDF is filed in a folder by supplier and month. On a typical day a handful get flagged, maybe a new supplier or a total that doesn't match the line items. Someone checks those in a few minutes instead of typing all 40.

The same pattern works for orders, delivery notes, receipts and purchase orders. If it's a document with numbers that someone types up, it's a candidate.

No new software needed

This connects what you already have: Gmail or Outlook, Google Sheets or Excel, cloud storage and your accounting software. If your accounting tool has no direct way in, the data goes to a spreadsheet you import from.

One tip before anything else: collect 20 or 30 real invoices from different suppliers, including the messy ones, and test on those. It tells you far more than any demo.

And if you get three invoices a month, don't bother yet. This pays off when someone is typing them up every week.

Originally published on dizilus.com.

Sales are in the shop. Leads are in the CRM. Costs are in the accounting software. Bookings are in a calendar. Somebody exports each one, pastes it into a sheet, fixes the formula that broke and tries to remember last week's number. By the time the report is ready, half of Monday morning is gone and nobody has the energy to think about what it says.

How the automated version works

Instead of a person collecting the numbers, a scheduled job does it:

1. Connect to each tool your numbers live in, read-only where possible.
2. On a schedule, say Monday at 7am, pull the past week's figures.
3. Put them into one consistent layout: a Google Sheet, an Excel file or a simple dashboard.
4. Compare with last week and the same week last month.
5. Have AI write a short plain-English note on what moved and what looks unusual.
6. Send it to your inbox, Slack or WhatsApp.

The numbers come straight from the source, so nobody types them. The AI only writes the commentary on top, and every figure it mentions can be traced back to the sheet.

Keep it to one screen

If a report needs scrolling, people stop reading it. A good weekly one covers:

• Money in: sales or revenue for the week
• New demand: leads, enquiries or bookings
• Work done: orders shipped, jobs completed, tickets closed
• One or two warning signs that matter to you, like unpaid invoices or low stock
• A few sentences on what changed and why it might have

A useful rule: pick the five numbers you'd want if you could only see five. Add more only when someone asks for them twice.

What it looks like

A cleaning company with twelve staff. Bookings come through a booking tool, invoices go out from the accounting software, enquiries arrive by email and WhatsApp.

Every Monday at 7am a report lands: jobs completed, revenue invoiced, invoices more than 30 days overdue, new enquiries and how many became bookings. Underneath, a note like: "Jobs were down slightly on last week, mostly Thursday, which had two cancellations. Overdue invoices went up because two larger clients haven't paid yet."

Two minutes to read, and you know where to look. A report you build is a chore. A report that arrives is something you actually read.

Can you trust the AI summary?

The AI never makes up numbers. It only reads figures the automation has already pulled and checked, and its job is to describe them and point at what's unusual. The raw numbers sit right next to the note, so if the summary ever disagrees with the table, you see it straight away.

What still needs you

Deciding what to do about it. The report can tell you enquiries dropped. It can't tell you whether that was the bank holiday, a broken form on your website or a competitor's sale.

Someone also has to own the definitions. What counts as a lead? When is a sale a sale? Agree that once, up front, and you avoid a lot of arguments later.

The easiest starting point is the report you already make by hand, however messy. Write down where each number comes from. That list is basically the blueprint.

Originally published on dizilus.com.