Read the purchase order once

One failed field restarted the document

An integration consultancy maintained purchase-order extraction for a distributor. The workflow read incoming documents, extracted supplier and line-item details, and prepared an order draft for an employee to approve. The distributor paid the provider and hosting charges.

Each comparison period covered 600 documents and 1,800 original pages, with the same supplier layouts and document mix. Handwritten orders and unfamiliar layouts went to manual processing outside the comparison.

When a field failed validation, the original recovery path restarted OCR and extraction. Some documents restarted several times. An incorrect item code could therefore trigger another charge for pages that had already been read successfully.

Keep the usable work and repair the field

The revised route stored OCR output against the document version and reused it during field repair. It retained valid fields and reconsidered only the failed field and its dependencies. A full validation then checked quantities, line totals, and required fields.

A tenant-scoped order-intent identifier prevented a retried job from creating duplicate work. A new order received a new identifier, even when its content resembled an earlier order. OCR ran again only if the source changed or the original output was unusable.

Every original page still incurred its initial OCR charge. The saving came from avoiding repeated reads during recovery.

An order still had to pass every check

A successful order draft required correct supplier details, item codes, quantities, and totals, ready for approval within 15 minutes. A material field correction counted as a failure even if an employee later repaired it.

The release criteria required at least 93% task success, no incorrect quantity or total bypassing validation, and a 95th-percentile processing time below eight minutes. The approver retained access to the original document and extraction history.

The operating results

Two 30-day periods, with the same eligible volume, task mix, provider rates, and acceptance criteria. All amounts are in Indian rupees, excluding tax.

95th-percentile time is the time within which 95% of tasks completed. Technology cost per successful task includes the cost of unsuccessful work and retries.

Technology cost fell by ₹3,420 over the 30-day period. Provider charges fell 30.0%, while total technology cost fell 23.8% because hosting stayed fixed. Cost per successful order draft fell 25.4%.

Initial OCR covered 1,800 pages in both periods. Repeated page reads fell from 900 to 90. At ₹2 per billable page, the 810 avoided reads saved ₹1,620. A document can be retried more than once, so retrying-document counts and repeated-page counts are different measures.

Model extraction and repair charges fell by a further ₹1,800 as targeted repair replaced full-document extraction during recovery. That difference comes from model usage, separately from the OCR calculation. Together, the two reductions account for the ₹3,420 lower bill.

The distributor received the saving. The consultancy's maintenance fee and the employee approval requirement stayed the same. Stored OCR output fit within the existing hosting allowance; a workflow exceeding that allowance would need to include the additional storage cost.

What this would mean over a year

At 600 documents per month, the ₹3,420 period difference projects to ₹41,040 in gross annual technology savings, split between ₹19,440 in OCR and ₹21,600 in model charges. Annual technology spend would move from ₹1,72,800 to ₹1,31,760.