← All insights

Insight / AI and infrastructure

Published

Local passport processing: an anonymised case study

How a local passport-processing workflow can support human review, with an explicitly illustrative model of its potential labour value.

Rows of server racks in a data centre
Photo: Kevin Ache / Unsplash

A client asked us to help with passport information processing while keeping the client’s identity and document details private. The work centred on a local processing model that could support the reading of required fields before a person checked them. That distinction matters: a quicker extraction step has little business value if staff spend the saved time finding and fixing mistakes downstream.

This is a real project, published with the client’s identity withheld. The figures below are assumed inputs and illustrative projections, not measured results from this client. They show the kind of business case we would test against the client’s records before reporting an achieved outcome.

The work around the model

The useful comparison covers the whole hands-on task: receive a passport image, read the fields needed for the business process, enter or transfer them, check the record and resolve exceptions. With local assistance, the model prepares candidate information for a person to review. The person must still be able to correct a field or stop a record before it reaches the next system.

At the level we can describe publicly, the local processing step sits between document intake and human field review. The source document and candidate fields need a defined path through review and export. This account does not specify the model, OCR software, hardware, storage layout or connected services. A complete data-flow review should confirm whether telemetry, support services, backups or later exports received the image or extracted fields. Running a model locally alone does not establish the full privacy or security outcome.

An illustrative business projection

Suppose a team handles 600 passports a month. Assume the existing process takes 8 minutes of hands-on work per document and a locally assisted process takes 3 minutes, including human review and the average time spent on corrections and exceptions. Value staff time at an assumed AUD $45 per hour loaded labour cost. These are modelling inputs, not observations or a quote for this client.

Assumed inputValue
Documents processed600 per month; 7,200 per year
Hands-on time before8 minutes per document
Hands-on time with local assistance, including review3 minutes per document
Loaded labour cost used for the illustrationAUD $45 per hour
Illustrative measureAssumed manual scenarioAssumed locally assisted scenarioDifference
Hands-on time per document8 minutes3 minutes5 minutes released
Hands-on work per month80 hours30 hours50 hours released
Hands-on work per year960 hours360 hours600 hours released
Labour value per documentAUD $6.00AUD $2.25AUD $3.75
Labour value per monthAUD $3,600AUD $1,350AUD $2,250
Labour value per yearAUD $43,200AUD $16,200AUD $27,000

The calculation is 600 documents × 5 minutes ÷ 60 = 50 hours of potential capacity a month. At the assumed loaded rate, that is 50 × AUD $45 = AUD $2,250 a month, or AUD $27,000 a year if volume and time per document remain constant. This is a valuation of staff capacity, not a realised cash saving or profit. Cash savings would require a documented reduction in overtime, contractor expense or another actual cost. If the team instead uses the time for other work, the benefit depends on the value of that work. Neither result can be inferred from the projection alone.

How to replace the projection with evidence

We would time comparable samples before and after the change, recording the number and types of documents, the start and end of each hands-on task, the human review required, and the time spent correcting or escalating records. We would also measure end-to-end turnaround separately; reducing hands-on minutes does not necessarily shorten a queue.

Quality belongs in the same comparison. Track field-level errors, correction and escalation rates, missed records and the review standard applied to each sample. A faster average is not an improvement if more incorrect information reaches the next system. Record the sample size and the mix of easy and difficult documents so a change in workload is not mistaken for a model effect.

Finally, compare any verified benefit with implementation, review, hosting, maintenance and operating costs. Those costs and the client’s actual loaded labour rate are needed for a net return. The same review should confirm the real data path, access and retention arrangements before making a stronger claim about privacy. Our guide to local AI and data control covers those broader infrastructure decisions.

The lesson from this project is to evaluate the reviewed workflow, not just the model’s extraction speed. The table gives a transparent starting point for the conversation; measured client results will replace it when the underlying timings, quality and costs have been checked.

External project website

Open this project site?

is hosted at . It will open in a new tab.