Услуга · Data and analytics

Intelligent document processing

An option for a flow of similar documents: invoices, delivery notes, acts, applications. The software identifies what it is looking at, extracts the details and enters them into the accounting system. The person only checks and confirms instead of typing everything in.

Volume
from a hundred documents a month
A person
still checks it
Into accounting
loads by itself
Training
on your own forms

What the work includes

It only makes sense at volume. A dozen documents a month is cheaper to type than to automate.

Discuss the scope

Reading

Scans and phone photographs, including skewed ones and those of less than ideal quality.

Identifying the type

The software recognises whether it is an invoice, a delivery note, an act or a contract.

Extracting the fields

Number, date, amounts, counterparty, line items. Tuned to your own forms.

Reconciliation

Reconciliation with the accounting system: does the counterparty exist, does the amount match the order.

Confirmation

A screen showing the document itself alongside what was extracted from it.

Loading

Once confirmed, the document is created in the accounting system.

How it goes

A pilot on one document type takes a month to six weeks including tuning it to your forms.

01

We count the volume

How many documents, of what kinds, and how much time typing takes now.

02

We collect the samples

Real documents in all their variety, including frankly poor scans.

03

We configure

We train it on your forms and measure the accuracy on a held-back sample.

04

We go live

We connect it to the live flow, train the operators and keep it under watch for the first month.

Perfect accuracy does not exist and cannot be relied on. The realistic bar is eighty-five to ninety-five percent of fields extracted correctly; a person corrects the rest. The gain is still substantial, but the verification step has to stay. Removing it on the grounds that the software is clever is a sure way to end up with distorted records.

Questions and answers

Take a hundred documents of one type a month as the benchmark. Below that, setup and support cost more than typing. Count in operator hours rather than in sheets of paper.

The software usually copes with skew and low resolution. It copes worse with crumpled paper photographed on a phone in poor light. Input quality hits accuracy directly, and sometimes it is easier to agree with the supplier that they send electronic documents.

Far more sensible, and it is the first thing we suggest. Recognition is for the counterparties who have no intention of moving to electronic exchange. With the rest it is wiser to agree on direct file transfer than to invest in reading scans.

We will remove the manual typing of documents

Tell us which documents and how many a month. We will judge whether recognition pays off or it is easier to agree on electronic exchange.