Local AI for Chilean receipts and tax documents

IT Lab has its own experience and specific criteria for evaluating and deploying local AI on Chilean tax and administrative documents. In IT Lab's judgment, legible documents are handled well by local models; deteriorated material and judgment calls require controls and human review.

Legible documents are processed well by models running inside the company, on accessible hardware.
The central risk is a model filling in a field with something plausible instead of declaring what it cannot read.
IT Lab's evaluation criteria are designed to detect that risk before any work is delegated.

The business question

A Chilean small or mid-sized company receives receipts, invoices, vouchers and supporting documents on paper, as photographs or as files, and someone keys them in for accounting. The business question is which part of that work a model running inside the company can take on, without sending the documents to a third party, and which part must stay under a person's control.

What IT Lab observes in this type of document

The IT Lab bench includes tests with documents of this kind, and the lab also supports projects with real documents from Chilean organizations. In general terms, IT Lab's experience is as follows:

Documents in good condition are processed well by local models, in times ranging from seconds to a little over a minute per case.
Deteriorated, ambiguous or uneven material is not resolved in a stable way, and no model evaluated clears it without additional controls.
Decisions that depend on judgment, and not only on reading, should remain with a person.

These observations are IT Lab's judgment based on its own measurements. The detail of what is measured and of the cases used is not published.

The central risk and how an own criterion addresses it

A vision model can fill in a field with a plausible figure instead of admitting it cannot see it. Unlike badly recognized text, that figure looks correct and can reach the accounting records without anyone noticing. In IT Lab's judgment, this is the main risk of the task, more than an omitted reading.

IT Lab answered with its own evaluation criterion, which rewards a model that declares it cannot read and treats an invented value as a serious fault. That criterion, and the controls that accompany it, are part of the experience IT Lab brings to each deployment. Its detail is not published.

What this means for a Chilean company

In IT Lab's judgment, the approach supported today has three parts:

What to automate. Reading and keying in legible documents, which is where the administrative staff's time savings concentrate.
What to review. Whatever the model declares illegible, deteriorated material and decisions that require judgment go through a person. The process's control rules catch a good share of the values that do not add up.
What traceability to keep. Each value stays linked to its source document and to the person who reviewed it, so any figure can be explained.

Confidentiality is the second argument. Tax documents contain personal and company data. With local AI that data does not leave the organization, which makes it easier to address Chilean data-protection and cybersecurity regulation (its scope is for the organization's legal advisor to determine).

On hardware, measurements were taken on a machine that IT Lab specified and built specifically for testing: a Ryzen 7 5700X processor, 31 GiB of RAM and an AMD Radeon RX 9070 XT graphics card with 16 GiB, running Linux. The card was chosen as the most affordable option today with a good cost-benefit ratio that a small company can buy. The models evaluated have versions optimized for NVIDIA cards, so the general pattern is expected to be comparable on that hardware; the published figures correspond to this machine.

Frequently asked questions

Can local AI read receipts and invoices? Legible documents are processed well by models that run inside the company, on accessible hardware.

What is the risk when reading tax documents? That a model fills in a data point with something plausible instead of stating what it cannot read.

Do the documents leave the company? With local AI, the data in tax documents does not leave the organization.

Is human review needed? Yes for deteriorated documents and anything requiring judgment, which need controls and human review.

Limitations and scope

Conclusions are IT Lab's judgment based on its measurements; they are not a certification.
This piece describes the experience in general terms; it does not detail fields, cases or controls.
Each organization has different documents, and results on its own documentation should be assessed before anything is delegated.
The bench does not measure costs. Running local AI carries real costs (energy, upfront hardware investment, maintenance and time) and, in IT Lab's judgment, current prices of paid services may not reflect their real cost.
Measured on a single machine; not a purchase recommendation.

Measurement sheet

Equipment: the machine described in the previous section, with 16 GiB of graphics memory.
Quantization: 3 or 4 bits, depending on the model.
IT Lab Bench, edition 1.