How do I stop my staff retyping emails and PDFs into a spreadsheet?
Short answer: You can automate most of it, but not all of it, and the part that makes it safe is not the AI. The usual setup is: an email arrives, a tool reads the body and the PDF, it writes the details into a spreadsheet row, and anything it is unsure about goes to a short review list for a person. Test it on 30 of your own real emails before you trust it, because accuracy on a vendor's demo file tells you very little about your suppliers' layouts.
Your team opens an email, squints at a PDF, and types the same six details into a sheet all day, and you want that to stop without the sheet filling up with wrong numbers.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
What can be automated, and what cannot
Most of the retyping in a small business falls into three kinds of email. Each one behaves differently when you try to automate it.
- Emails where the details are in the message itself. A web form notification, an enquiry, a booking request. These are the easiest. The layout barely changes, so even simple rules work.
- Emails with a digital PDF attached. An invoice or purchase order made by software. The text inside the PDF is real text, so a tool can read it directly. This is the sweet spot for AI extraction.
- Emails with a scan or a phone photo attached. The file is a picture of paper. A tool has to turn the picture into text first. This step is called OCR (optical character recognition). It works, but it makes more mistakes, and it makes the most on crooked, faded or handwritten pages.
So the realistic scope for most owners is this: automate the first two kinds fully, with a review list, and treat scans as a smaller, slower lane. If you want the same idea applied to orders, my guide to email order processing goes deeper, and purchase order processing covers the supplier side.
Three ways to build it
| Approach | Good for | Weak spot |
|---|---|---|
| Mail rules plus a manual paste | Under 10 emails a day, one fixed layout | Still a person doing the typing |
| A document-reading service connected to your mailbox and sheet | Invoices, receipts, standard forms, 10 to 200 a day | Odd layouts need fixing, and each service has its own limits |
| A custom build around your own forms and suppliers | Many layouts, scans, rules about what counts as valid | Costs more up front |
For the middle row, here are two real examples so you can see what these services look like. Microsoft's Power Automate, a tool that connects apps and runs steps automatically, has an AI Builder invoice model that pulls out common invoice fields such as invoice ID, invoice date and amount due. Microsoft's documentation says it accepts JPEG, PNG and PDF files up to 20 MB, wants one clear photo or scan per invoice, and allows 360 calls per 60 seconds per environment across its document models. Google Document AI, a Google Cloud service, turns documents into structured fields and can digitize scans with OCR. Check each vendor's pages for current limits and pricing before you commit, since both change.
Step by step: build it for one email type first
- Pick one email type. Choose the one that costs the most time, such as supplier invoices. Do not start with "all email".
- Write down the fields you actually need. For an invoice that is usually supplier, invoice number, date, total and due date. Six fields is plenty to start. Every extra field adds a chance of error.
- Collect the emails in one place. In Gmail, you can create a filter that applies a label to matching mail. Gmail's help page says that when you set a filter to forward messages, only new messages are affected, and that a reply is only filtered if it also matches the same search criteria. In Outlook, a rule needs a name, a condition and an action. Either way, you end up with one label or folder that holds only the emails you want read.
- Connect the extraction step. The tool opens each new email in that label, reads the attachment and returns your fields. This is where Power Automate, a similar connector or a custom script sits.
- Write a row to a sheet, not straight to your accounts. A staging sheet is the safety net. Nothing reaches your accounting or order system until it has passed the checks below.
- Add the checks. See the next section.
- Run it side by side. For two weeks, let staff keep typing as before while the tool fills its own sheet. Compare the two. This is the 30-email test, described below.
The part vendors skip: checks and a review list
An extraction tool does not know when it is wrong. It returns a number with the same confidence whether it read it perfectly or guessed from a smudge. So you build the doubt in yourself, with plain rules:
- Required fields. If invoice number or total is blank, the row goes to review.
- Format checks. A date must be a real date. A total must be a number. A phone number must have the right length. Google Sheets lets you set data validation on a column, which can reject or warn on entries that do not match a rule.
- Sense checks. Line items must add up to the total. A total ten times your usual size gets flagged.
- Duplicate check. Same supplier plus same invoice number means the second one is held, not added.
- Known-name check. The supplier name must match your supplier list. A near miss goes to review instead of creating a new "Acme Ltd" next to your existing "ACME Limited".
Rows that fail any rule land in a review list: a filtered view of the same sheet with the original email linked. A person fixes the row in under a minute because the source is one click away. Rows that pass go on. That list is the difference between automation you can trust and a spreadsheet that quietly fills with errors.
The 30-email test
Before you rely on any tool, test it on 30 real emails from the last month. Include the ugly ones: the supplier who sends scans, the one who sends three invoices in one PDF, the one whose invoices are in another language. A vendor demo will not do this for you.
Score it like this. With 6 fields per email, 30 emails gives 180 fields. Count each as correct, blank or wrong. Wrong is the one that hurts, because blank gets caught by the required-field rule and wrong does not.
Here is a worked example with made-up figures. Say the tool gets 162 fields right, leaves 12 blank and gets 6 wrong. That is 90 percent right, 7 percent blank, 3 percent wrong. The 12 blanks go to the review list automatically. Now look at the 6 wrong ones. If 4 are the same supplier's odd layout, you fix that one layout and the wrong rate drops. If they are scattered, add a sense check or keep a person on those fields.
Now the time. Suppose staff handle 40 emails a day at 3 minutes each, which is 2 hours a day (made-up figures). If the tool passes 32 of the 40 cleanly and 8 go to review at 1 minute each, the daily work falls to about 8 minutes, plus a spot check. That is the shape of a good result. If your test shows half the emails going to review, the tool is the wrong fit or the setup needs work, and you learned that cheaply.
Looks the same, but is not
- A PDF you can select text from versus a PDF that is a scan. Try to highlight a word. If you can, it is digital and easy. If not, it is a picture and needs OCR.
- One invoice per file versus several. Microsoft's invoice model, for one, expects one invoice per photo or scan. Bundled files need splitting first.
- The invoice total versus the amount due. After credits and part payments they differ. Decide which one your sheet wants, and test for it.
- The sender versus the supplier. A bookkeeper may forward every invoice. The sender is then wrong for your supplier column.
- A resent email versus a new one. Suppliers resend. Without a duplicate check you will count it twice.
Where the data goes
Invoices and orders hold names, addresses and bank details. Before you connect a mailbox to any service, find out where the files are processed, how long they are kept and whether they are used to train models. Anthropic, the AI company behind Claude, states that it does not use inputs or outputs from its commercial products to train its models by default, and its consumer plans are covered by separate terms. Every vendor has its own rules, so read the one for the exact plan you would use. If you handle personal data about people in the UK, the guidance on AI and data protection from the ICO (the UK's data protection regulator) covers accuracy and data minimisation, which is a good checklist. If the data is sensitive, ask a lawyer.
When ready-made tools stop being enough
A ready-made reader is fine until one of these shows up:
- You have dozens of supplier layouts and the fixes never end.
- Most of your files are scans or phone photos.
- The result must land in a specific system, such as your quoting or stock tool, with your own rules about what is valid.
- You need an audit trail showing who approved each corrected row.
At that point the job is less "read a PDF" and more "build a small process", and a custom build starts to pay. If the spreadsheet itself is the next problem, see how to replace five spreadsheets with one customer database. For searching your files once they are in, see how to search 40 supplier PDFs and get an answer fast. If the typing is really a staffing question, compare it with data entry outsourcing. And if the mail you are saving is leads, how to add emails to a spreadsheet of leads covers that case. The wider pattern is in business process automation and AI document processing.
Red flags when buying a tool or hiring someone
- They quote accuracy on their own sample files and will not run yours.
- No review list. Every result goes straight into your system.
- No answer to "what happens when it is wrong?"
- No duplicate handling.
- They cannot tell you where your files are stored.
How I'd approach this
I would not start with a tool. I would start with the 30 emails. Pull them, list the six fields, and type the right answers into a sheet yourself. That sheet is your answer key, and it also shows you which suppliers are the hard ones. Then try the cheapest reader that fits, score it against the key, and add the review list before anything touches your accounts.
If the score is good, stop there. Do not pay for a custom build to save 10 minutes a day. If your test shows scans, many layouts or special rules, that is when I would build something around your own process. Fixes and finishes at Blur Studio start around $1,500, and I will tell you on a call if a simple setup will do instead.
Questions people ask
Can I automate data entry from email to Excel for free?
For simple cases, yes. Mail rules can sort the emails into one folder, and many tools can copy the body of an email into a sheet row. Reading PDFs and scans is where limits and costs appear, so check each tool's current plan before you build on it. My guide to business process automation shows where the free options run out.
How accurate is AI at reading invoices and PDFs?
It depends on your files, not on the tool's marketing. Clean digital PDFs do well. Scans, folded pages and unusual layouts do worse. Test 30 of your own emails and count correct, blank and wrong fields, as described above.
Do I still need a person checking?
Yes, for the rows that fail your checks. The aim is that people only look at the exceptions, not at every row. For anything that moves money, keep a spot check on the rows that passed as well.
What if the PDF is a scan or a phone photo?
It needs OCR first, which means turning the picture into text, and it will make more mistakes. Ask staff to send clear, flat, one-page-per-document photos, and keep a larger share of scans in the review list. Microsoft's own invoice model guidance asks for one clear photo or scan per invoice.
Is it safe to send supplier or customer emails through an AI tool?
It can be, if you check the vendor's data terms for your exact plan: where files are processed, how long they are kept, and whether they are used for training. If you handle personal data, ask a lawyer about your duties. See also AI document processing.
Is data entry outsourcing a better option?
Sometimes. It suits low volume or work that needs judgment, and it still has typing errors. Automation suits steady volume with repeating layouts. See data entry outsourcing to compare the two.
More in this topic: Problem solving: automating service business work (21 guides) →
Sources
- Invoice processing prebuilt AI model (Microsoft Learn)
- Document AI overview (Google Cloud)
- Create a filter in Gmail (Google Help)
- Manage email messages by using rules (Microsoft Support)
- Create an in-cell dropdown list and data validation (Google Docs Editors Help)
- Is my data used for model training? (Anthropic Privacy Center)
- Guidance on AI and data protection (UK ICO)