AI Agents & AutomationFree worksheetPDF, Excel, Word exportDiscrete Manufacturing

Manual Keying Correction Log Worksheet

Overall accuracy hides which fields cause the work. This log tracks corrections per field so you can see where the engine struggles and fix that field first.

Download:

Field	Documents checked	Corrections made	Correction rate	Typical error	Impact	Planned action
			9.5%
			1.5%
			7.2%
			4.3%
			14%

Total corrections

111

Worst field correction rate

14%

Corrections per 100 documents checked, all fields

6.3429

Count a correction when a person changes the value the system proposed. Do not count fields the system left blank and a person filled, unless you add a Missed field row.

Close the feedback loop

Optional: see how corrections can retrain a private model on your own servers.

Optional. The tool and every export work without this. No spam.

How to use the log

Pick a review period, such as a week, and a sample of documents that reviewers touched. For each field, enter how many documents were checked and how many times the proposed value had to be changed. The sheet computes the correction rate for each field. Add a typical error type and an impact rating so you can compare a cosmetic date error against a wrong unit price that could cause a wrong payment.

Reading the results

Sort by correction rate, then by impact. A field with a moderate rate but a blocking or payment impact deserves attention before a high-rate cosmetic field. The summary shows corrections per hundred documents across all fields, a simple trend number you can track week by week. Watch for a field that suddenly worsens, which often means a supplier changed layout or a scan source degraded.

Turning corrections into fixes

Wrong characters point to image quality or model confidence. Wrong format errors usually have a rule-based fix, such as a decimal separator or date order per country. Wrong field picked suggests a layout issue that needs a template or more training samples. Add validation rules that use ERP data, like checking a PO number against open orders, to catch errors before a person must.

Common mistakes

Do not count only the documents that failed, because that overstates error rates. Do not mix invoice fields and quality fields in one period without a separate row for each. Finally, retain old logs. When you change a configuration, the previous period is your baseline.

Frequently Asked Questions

What is a correction rate?

It is the number of times reviewers changed a field value divided by the documents checked for that field, shown as a percent. It is a practical measure of how often the engine was wrong on a field in daily use.

Is this the same as OCR accuracy?

It is closely related but based on real review behavior rather than a test set. It captures only errors that reviewers noticed, so it can understate true errors if review is shallow.

How big a sample should I use?

A few hundred documents per period gives a stable view for common fields. Rare fields, like heat numbers, need their own count of documents where the field appears, which is why checked is entered per row.

Which fields should I track first?

Start with the fields that block posting or affect payment: vendor, PO number and line, quantity, unit price, tax and total. Add descriptive or quality fields once those are under control.

Netray can capture reviewer corrections automatically and feed them back to an on-prem model, so accuracy improves without separate logging work.

Talk to an Expert Explore More Free Tools

Related Tools

AI Agents & Automation

More AI Agents & Automation tools

See all

Back to All Tools