AI can now read the paperwork that arrives in a small business, such as delivery notes, timesheets, purchase orders and supplier invoices, and turn it into data a system can use. It does not read perfectly. The design that works is to let it read everything and have a person look only at the documents it is unsure about.
This post covers what the reading looks like now, where it goes wrong, how the checking works, how to match documents against purchase orders, and how to start with one document type.
What “reading a document” means now
Reading a document is more than capturing the text. It means knowing which number is the total, which is the purchase order reference, which lines are products and which is the delivery charge. Document-reading services return those as named fields: supplier, date, invoice number, line items, totals.
The tools come in a few shapes. Anthropic’s documentation says Claude can answer questions about text, pictures, charts and tables in PDFs and lists converting document information into structured formats as a use case. Microsoft’s Document Intelligence service has a prebuilt invoice model. And if you use Xero, its own capture tool, Hubdoc, is worth knowing about: Xero’s app listing says you can capture bills and receipts from a phone, email or scanner, that the key information is extracted automatically, and that it is published to Xero with the original attached. Check your own Xero plan to see whether it is included.
For plain supplier bills, it is sensible to try the tool you may already have. Whether it does the purchase order matching and the coding rules your business needs is a question to test on your own documents, not to assume.
Which documents to start with
Roughly from easiest to hardest, and this ordering is our judgement, not a measured result:
- Supplier invoices. Mostly printed, mostly one or two pages, and the fields are predictable.
- Purchase orders. You issue them, so you control the format.
- Delivery notes. Layouts vary by supplier, many arrive as phone photos, and quantities are often ticked or written in by hand.
- Timesheets. Regular in layout but often handwritten, and they hold personal data about your staff.
Start at the top of the list. The later ones are doable, but you want to see what corrections look like on an easy case first.
What it gets wrong
Knowing the failure modes is what lets you design around them.
- Poor images. Photos taken at an angle, in shadow or of a folded page. Microsoft’s documentation on confidence scores says that when the confidence of the text read is low, you should improve the quality of your input documents. A short rule for whoever photographs delivery notes (flat, in light, the whole page) helps.
- Awkward tables. The same documentation says merged cells should be expected to produce lower confidence, and that a table row split across two pages looks different to the model and is less certain. Long delivery notes and multi-page invoices are where this shows up.
- Handwriting and stamps. Harder than print for any reader, human or machine. A handwritten quantity and a signature are the fields most worth a second look.
- Invented values. A language model can state something that is not on the page. Anthropic’s guidance on reducing hallucinations suggests giving the model explicit permission to say it does not know and having it quote the text it relied on, and it adds that these techniques “don’t eliminate them entirely” and that you should always validate critical information.
- Plausible wrong digits. A 3 read as an 8, or a decimal point lost. The result looks reasonable, which is why checking by eye alone is a poor safeguard.
None of this is a reason to avoid the technology. It is a reason to build the checking in from the start.
Check the exceptions, not every document
Reading tools usually return a confidence score for each field. Microsoft’s documentation describes it as an estimated probability between 0 and 1 that the prediction is correct, and says that for scenarios where accuracy is critical, “confidence can be used to determine whether to automatically accept the prediction or flag it for human review.”
That gives you the shape of the process. Every document is read. Most pass straight through. A document goes to a person when any of these is true:
- A field has low confidence.
- The line items do not add up to the total, or the VAT looks wrong.
- The supplier is one the system has not seen before.
- The invoice number has already been used.
- It does not match its purchase order (more on that below).
The person sees a short queue, with the original document next to the extracted fields, and corrects or approves. Each correction is useful, because it tells you which rule needs tightening.
Do not set the thresholds from a guess. Run the first few weeks with a person checking everything, record every correction, and set the rules from what you see.
Match against purchase orders
For firms that buy materials or subcontract work, the matching is where the real time goes, not the typing. A three-way match compares three documents: the purchase order (what you asked for), the delivery note (what arrived) and the invoice (what the supplier is charging for).
If all three agree, the bill can be approved. If they do not, a person looks. For example, a purchase order for 40 sheets of plasterboard, a delivery note showing 36 received, and an invoice for 40: the system should flag a short delivery before anyone pays for sheets that never arrived.
The practical rules are simple. Match on supplier and purchase order number first. Then compare amounts within a tolerance you choose (a few pence on rounding is different from a few hundred pounds). Then compare quantities against the delivery note where you have one. Anything outside the rules goes to the exceptions queue.
Where the data lands
The extracted data has to end up somewhere useful, and for most owner-led firms that means two places.
Your accounts package. In Xero, that is a draft bill with the supplier, amounts, VAT and account code filled in, the original document attached, and a person approving it. The original matters. HMRC’s guidance says that, generally, you must keep VAT records for at least 6 years. A scanned image is acceptable if it holds all the detail required for VAT, but where data has only been typed into software, the invoice must still be kept in its original form. So the document should travel with the record, never be replaced by it.
Your job or project system. Delivery notes can update what a job has cost so far. Timesheets can put hours against jobs, which feeds both invoicing and payroll. That removes the weekly job of re-keying figures just to see whether a job is making money.
If you are pulling sheets together by hand today, our piece on signs your business has outgrown spreadsheets covers the options for the other end of this pipeline.
Mind where the documents go
Supplier paperwork contains prices and bank details. Timesheets contain names and pay information about your staff. Before sending either to an AI service, check which terms you are on, because the default for personal accounts is different from the default for business ones. We cover the difference in where your data goes when you use ChatGPT and Claude.
Start with one document type
A sequence that works:
- Pick the document that costs the most time and has the most regular layout. For most firms, supplier invoices.
- Collect 50 to 100 real examples, including the worst ones: the crumpled, the photographed, the multi-page.
- Run it for a few weeks with a person checking every document, recording each correction and what caused it.
- Set the exception rules from what you learned, and move to checking exceptions only.
- Add the next document type and repeat.
We built this for a construction firm in Hertfordshire: supplier bills are filed from the mailbox, matched to purchase orders and coded, instead of being keyed in by hand. Our AI and data engineering service covers this kind of build.
Where to start
If you want a view on which paperwork to start with, the Systems and AI Review is half a day with us, on site or on a call. You get a two page plan covering the three places your admin costs you most and what to fix first, plus a fixed quote for the next step, and every piece of work is agreed in writing first. You can book a 30 minute call to talk it through first.