How we cut document turnaround from four hours to five seconds

A UK cloud accounting platform came to us with a familiar problem. Its customers send in financial documents every day, and before anything could be booked, each one had to be read and turned into structured data. On average that took four hours.

Today the same platform processes around 100,000 documents a month at 99.5% accuracy, and average turnaround is about five seconds. This article explains what changed, and what we would tell anyone facing the same problem.

Where the four hours went

Four hours is not the time it takes to read a receipt. It is the time a document spends waiting: in an inbox, in a queue, for someone to open it, check it and key it in. Most of that time is idle.

That matters, because it changes what you fix. Faster typing, or more people, shortens the queue a little. Removing the queue removes the four hours.

What we changed

We put AI extraction at the front of the process. We deployed Caelum, the financial document-extraction engine we developed, to read each document as it arrives and return structured data straight away. There is no queue to wait in.

We took documents in the way customers already send them. The platform’s customers were sending documents by email and by WhatsApp. Instead of asking them to change, we built ingestion for both, so every document lands in the same pipeline whichever route it takes.

We made the data useful straight away. Extracted data feeds analytics and data-health modules, so the platform and its customers can see the state of each set of books, not just a pile of processed files.

We cleared the technical debt that was in the way. An ageing database was still running on AWS extended support at £500 a month. We migrated it as part of the same work, which removed a £6,000-a-year cost.

Lessons worth taking away

Measure accuracy on your own documents. Accuracy figures quoted in general mean little. The 99.5% here is measured on this platform’s real document mix. Any project like this should start by testing on a sample of your own documents.

Do not make customers change their habits. Customers will keep sending documents by email, WhatsApp or whatever is easiest for them. A system that accepts all of those routes gets used; one that demands a portal upload gets worked around.

Speed comes from removing waiting, not from working faster. Look at where documents sit idle before deciding what to automate.

Fix the foundations at the same time. Automation on top of a fragile, expensive legacy system just moves the problem. Retiring the old database paid for itself in a year.

Who this applies to

This applies to any organisation that receives a steady flow of documents and re-keys data from them: accountancy practices, software platforms for accountants, lenders, insurers and back-office teams in the public sector. If documents wait hours or days before they are useful, the same approach applies.

Read the full case study: From four hours to five seconds: AI document extraction for a UK cloud accounting platform.

If your team still reads documents by hand, book a 30-minute discovery call. We will look at your volumes and tell you honestly whether automation is worth it.

Let’s talk about your documents

A 30-minute discovery call: your documents, systems and volumes, and what could be automated. No obligation.

Download our capability statement (PDF)