The current state
Most small and mid-sized businesses process supplier invoices like this. An invoice arrives by email. Someone opens the email, opens the PDF, opens the accounting system, and types the data in. Invoice number. Date. Supplier. Total. VAT. Line items, if they are detailed. They save. They move on. Forty invoices later, it is 3pm and they have done nothing else.
This is not a niche problem. It is the default state for most finance and operations teams. It is invisible because it is normal. It does not generate complaints because everyone assumes this is just how it works.
It is also expensive. Conservative estimates put the cost of manually processing a single supplier invoice at between £6 and £15 when you factor in staff time, error correction, and approval delays. For a business processing 200 invoices a month, that is between £1,200 and £3,000 per month — every month — for a process that produces no value beyond moving data from one place to another.
What Harold automates
Harold automates the extraction step. The supplier sends the document. Harold reads it, identifies every relevant field, and produces a structured data record. No human opens the PDF. No human types numbers into a system.
The specific tasks Harold eliminates: opening documents and reading them to find specific fields, manually keying data into a spreadsheet or accounting system, cross-referencing supplier names against internal codes, applying consistent formatting to dates, amounts, and references, and tracking which invoices have been processed and which have not.
For businesses using the Automations feature, Harold also handles the document intake — suppliers send to a Harold email address and the document is in the system before anyone has looked at it.
What Harold does not automate
Harold does not approve payments. It does not perform three-way matching against purchase orders automatically. It does not post journal entries. It does not replace your accounting system.
What Harold does is get structured data out of unstructured documents. Where that data goes next — into a spreadsheet, via Zapier into your ERP, into a CSV import — is your choice.
The review step is also intentionally not automated away. Harold produces high-confidence extractions and flags low-confidence ones. A human reviews the flags. This is the right design — the cost of an undetected error in financial data is higher than the cost of a two-minute review.
The realistic difference
For a business processing 50 invoices a week with Harold, a realistic workflow looks like this. Invoices arrive throughout the week via the Harold Automations inbox. Harold processes them automatically. On Monday morning, the finance person opens Harold, filters for In Review documents (typically five to ten percent of the total), checks the flags, and approves the batch. Total time: 15–20 minutes. Export to CSV or push to Zapier. Done.
Previously that same task took three to four hours spread across the week. The quality was also inconsistent — humans mistype, get distracted, skip fields when tired.
The error rate also drops substantially. Harold does not get tired. It does not accidentally transpose digits. It applies the same rules every time. The error rate on extraction is not zero — Harold will occasionally miss a field or misread a value — but the error rate is lower than human keying, and errors are surfaced with confidence flags rather than silently entering the system.
What this is not
This is not a pitch for Harold specifically. It is a description of what document automation actually does and does not do. If you are evaluating Harold or any similar tool, the honest question to ask is: what is the total staff time currently spent on manual document entry, and what is that time worth to the business?
If the answer is "not much", automation may not be worth the setup. If the answer is "a lot", it probably is.
Harold's approach is to make the setup low enough that the answer can be "not much" and it still makes sense. Train once. It takes an afternoon. After that, it runs.