The meeting is at nine. At eight, three files are still missing.
Imagine an operations team combining reports from several branches every week. One sends a table in a PDF, another a scan and a third a spreadsheet. An employee copies values, corrects units and tries to establish whether each document covers the full week or only part of it.
When a total looks unusual, the attachments have to be opened again. After the meeting, a correction arrives for one report. The new figure goes into the spreadsheet, but the summary already sent remains in circulation. Reading documents faster will not, by itself, solve the problems of versions, omissions and confidence in the result.
What holds up the work?
Starting point
The same label, a different definition
A quantity field may mean individual items, packs or kilograms. Transferring a number without its context produces a convincing-looking table of values that cannot be compared.
A gap that looks like zero
A missing file, an empty cell and a genuine zero result are treated alike. The report’s reader cannot tell whether the data is complete.
A result without a route back to the source
To answer a question about a single entry, an employee searches folders and correspondence. The report does not retain the source document or the location where the value was read.
First, a shared data vocabulary. Then, automation.
We begin by defining the report: its period, units, required sources and the calculation behind every item. Only then do we choose table extraction, OCR or an AI model. Different documents may require different methods.
Every extraction passes consistency checks. An uncertain document enters a review queue with a specific reason for being held. The final summary retains links to its sources, a version identifier and information about whether all expected data is included.
What the new workflow could look like
Collect and identify the documents
Files arrive through an agreed channel: a folder, form or integration. The system identifies the sender, reporting period and document type, and detects repeat delivery of the same file.
Read values into a common format
Fields are mapped to a shared structure. We keep both the original entry and the converted value, so a change of unit or date format can be explained.
Hold data that needs attention
A missing page, unknown unit, inconsistent total or questionable reading sends the document to a person. The interface shows the value alongside its source and allows the correction to be explained.
Approve and publish a version
An authorised person checks the completeness summary. A corrected source creates a new report version showing the changes, rather than silently modifying a previously approved result.
AI helps read. Rules protect the meaning.
A model can locate a field in a document with a changing layout, but its answer does not prove the value is correct. We combine extraction with checks on data types, reporting periods and relationships between fields. A model’s stated confidence is no substitute for verification. Sensitive or business-critical entries may require approval regardless of how they were read.
- A missing value remains missing unless the process explicitly defines otherwise.
- A rejected document has an owner and a visible reason.
- An employee’s correction retains the previous value and the identity of the person who changed it.
Who can see the source document?
A report for a manager does not necessarily need to expose every attachment to the whole team. We separate access to files, correction of extracted values and approval of results. We establish source retention periods, the handling of password-protected documents and permitted processing locations. When using external services, we need to define which data may be sent to them before building the integration.
Let us start with one part that works.
We start with one recurring report and a limited set of formats. We need samples of ordinary documents, corrections and difficult cases, plus someone who can approve a correct result. The first stage ends with a comparison of reports based on the same data, not simply with PDF extraction being switched on.
How will we know it is better?
This is a measurement plan, not a promised result. We agree the baseline before implementation.
Time spent preparing the report
We measure active time spent collecting, copying and checking data. We separate it from waiting for missing documents and from the system’s processing time.
Effectiveness of the checks
Using an agreed sample, we check extraction accuracy and the number of errors that passed validation. The proportion of automatically processed files is not a sufficient measure on its own.
Ability to explain the result
We check whether a user can trace a report entry back to its source and explain why two versions of the report differ.
Questions worth asking
Can scanned documents be read too?
Yes, OCR can be used, but scan quality affects the result. We examine samples from the real workflow: skewed pages, stamps, poor contrast and handwritten additions. Unreadable documents need a route for improvement or manual review.
What happens if a supplier changes its report layout?
An unknown layout should go for review rather than produce apparently valid data. We define how changes are detected and field mappings updated. Before a document returns to automatic processing, we compare its extraction with an approved sample.
