Caelum AI document processing for a UK accountancy practice
Mindrops built Caelum AI for a large UK accountancy practice, which uses it to process tens of thousands of documents a month. It reads each receipt, invoice and statement, classifies it, applies finance-specific rules, and returns clean, accounting-ready data with a human-review step for control.
At a glance
Caelum AI is Mindrops' AI document-processing platform, built for a large UK accountancy practice. It reads receipts, invoices and bank statements with AI, OCR and human-in-the-loop review, applies VAT, DRC and CIS rules, and exports to Excel, CSV or Tally. It processes tens of thousands of documents a month at 99% extraction accuracy, cutting manual statement entry by about 95%.
The challenge
Every receipt, invoice and statement had to be read and typed into accounting software by hand — accurately, and against fixed reporting deadlines. Generic capture tools handled simple documents but missed the detail that matters in bookkeeping.
Generic OCR dropped or mis-read transactions on bank-specific layouts (NatWest, HSBC, Santander and similar). Multi-receipt PDFs and partial statements made results unreliable without human checking. VAT bands, Domestic Reverse Charge (DRC) and Construction Industry Scheme (CIS) amounts had to be worked out per document. Output still needed reshaping into Excel, CSV or Tally, and there was little visibility of volume or cost per client. The cost was a steady drain — hours, accuracy and headroom lost on data entry that should have been automatic.
The solution
Caelum AI is built as five document-processing modules sitting on a shared AI document-reading layer, with multi-tenant client, credit and audit infrastructure underneath. Each module handles a document type end to end and emits the same workflow-ready output.
New capability is added as a module on the existing stack. Every engine shares one reading layer — AI and OCR with a human-review step where confidence matters — and one tenant-isolated infrastructure with clean billing separation.
AI document reading
Each document is read, classified and normalised by an AI and OCR layer, with a human-review step — reliable extraction that keeps the accountability bookkeeping needs.
Finance-aware extraction
Supplier, dates, gross/net/VAT and line items are pulled out, with VAT-band grouping, DRC and CIS handling — accounting-ready detail generic OCR could not produce.
Per-page bank classification
Every page is matched to a specific bank-and-layout variant, then read with its best-fit approach across UK, US and India.
Multi-receipt splitting
A single PDF of many receipts is split and grouped per receipt before extraction, with a tenant-isolated cache — for the bundled scans month-end produces.
Structured output and exports
Results are emitted as structured data plus Excel, CSV or Tally, with a separate error file for rows that need review — data enters the accounting platform without reshaping.
Usage & cost visibility
Administrators see per-client volume, page counts, processing time and cost, with Excel export and drill-down — a billable, high-volume operation, measured.
Results
Manual statement-entry time fell by roughly 95%, extraction accuracy reached 99%, and tens of thousands of documents are processed every month. Staff now review and confirm statements that Caelum AI has already read.
Statement entry
Before: each statement read and typed by hand, line by line.
After: read automatically; staff confirm each statement — roughly 95% less time.
Accuracy
Before: manual keying and generic OCR introduced errors and rework.
After: 99% extraction accuracy, with a human-review step for control.
Bank coverage
Before: bank-specific layouts dropped or mis-read transactions.
After: per-page classification reads each layout across UK, US and India.
Output & visibility
Before: data reshaped by hand; volume and cost per client hard to see.
After: Excel, CSV or Tally emitted directly, with per-client usage and cost in one view.
Headline figures reflect targets and outcomes for the deployment described — ~95% time reduction and tens of thousands of documents per month as supplied for this engagement, and 99% extraction accuracy as stated for the platform. They are not presented as independently audited.
Technologies used
A shared AI reading layer, a per-page bank engine, region-specific pipelines and finance rules applied during reading — on multi-tenant infrastructure with per-page credit billing.
Reading layer & bank engine
AI/ML extraction and OCR with human-in-the-loop review; per-page classification with a best-fit reader per layout and an exception-bank path.
Regions & finance rules
Dedicated UK, US and India pipelines for layout, date and currency; VAT bands, DRC and CIS applied automatically during reading.
Tenancy, access & billing
Per-tenant credentials, credit pool, storage and audit; dashboard users and a server-to-server API; per-page credits with junk and failed pages not charged.
Caelum AI is a Mindrops product — see the Caelum AI product page for how the same capability is delivered via dashboard, API and partner embed.
Frequently asked questions
What is Caelum AI?
Which documents and bank formats can it read?
How accurate is it, and how is quality controlled?
What results did the accountancy practice see?
Could Mindrops automate our document processing?
Contact us
Tell us about your high-volume document process. Mindrops builds AI document-processing systems that turn receipts, invoices and statements into clean, accounting-ready data, with a human-review step for control.