Conclusion First: “The Job of Reading Paper Invoices by Humans Can Now Come to an End”

Sorting invoices, comparing estimates, aggregating inventory data—there are still tasks in small and medium-sized enterprises where humans are manually entering data into Excel. Outsourcing these tasks can cost hundreds of thousands of yen per month, and sending them to a BPO can exceed 3 million yen annually. Many business owners think, “AI is not feasible for our scale.”

However, as of June 2025, the situation has changed. OCR technology for reading documents, automated processing agents, and models for analyzing tabular data—these three components can now be obtained for “almost free to under 50,000 yen per month.”

Specifically, these components are Jina AI’s document parser “jina-ocr-v1,” an open-source local LLM agent harness, and TabPFN-3.5 (a foundational model for tabular data). This combination is what we are discussing.

Let’s take a closer look at what each component does, how much it costs, and what changes it brings to the field.

—

Part 1: Jina-OCR-v1—Throw in an Invoice and Get Structured Data Back

Jina AI’s “jina-ocr-v1” is a document parser with 3.4 billion parameters.

Its function is simple. When you feed it PDFs, scanned images, invoices, tables, or graphs, it returns the data in clean Markdown format, structured and ready to use. It operates end-to-end with no preprocessing required.

There are three key points:

1. The GPU required for operation is lightweight.
The decoder part has 570 million parameters and can run on an NVIDIA L4 (which costs about 10,000 to 20,000 yen per month in the cloud). Even a used PC with an RTX 4060 is sufficient. There’s no need to rent an A100 for 200,000 yen a month.

2. Open weights are available.
The model weights are publicly available, meaning it can run on your own server or local PC. There’s no need to send data to an external API, so there’s no worry about leaking invoice data from business partners. This is significantly important for small and medium-sized enterprises.

3. The output is in Markdown, making it easy to connect with subsequent processes.
The structured Markdown can be directly passed to an LLM. Instructions like “extract the total amount and payment due date from this invoice” can be executed seamlessly.

However, there are some caveats. Currently, there are licensing restrictions for commercial use. While research and non-commercial use are free, you need to confirm with Jina AI if you plan to incorporate it into production operations. This is something to watch for in the future.

That said, the costs of using it via API are drastically lower compared to traditional OCR services (which typically cost between 100,000 to 500,000 yen per month). Even with API usage, processing thousands of invoices monthly can be kept to around 1,000 to 10,000 yen.

—

Part 2: Local LLM Agent Harness—Automating “Read, Decide, and Process Next”

Simply reading documents with OCR is only half of the automation process.

You need to create a chain of decisions and actions like “input the amount from the read invoice into the accounting software” or “compare estimates from three companies and notify the cheapest vendor on Slack.”

This is where the open-source agent harness comes into play.

What is an agent harness? In simple terms, it’s a framework that equips an LLM (language model) with “tools” to autonomously execute tasks. While the model alone can only “generate text,” passing it through the harness allows it to have execution capabilities such as “reading files,” “calling APIs,” and “writing to databases.”

As of 2025, there are plenty of options available:

• Ollama + Open WebUI: A standard for running local LLMs, with easy context window expansion.
• LangGraph / CrewAI: Excels in coordinating multiple agents.
• Smolagents (Hugging Face): Lightweight and suitable for small tasks.

The important thing is that all of these are free. There are no software licensing fees; you only need to cover the hardware costs to run the models.

If running locally, a setup with Ollama using Qwen3 or Llama 4 Scout is practical. A PC equipped with an RTX 4060 or 4070 (which can be found used for 150,000 to 200,000 yen) can comfortably run 8B class models. In the cloud, you can rent GPU instances from Lambda Labs or Vast.ai for 10,000 to 30,000 yen per month.

“OCR reads the invoice Markdown → LLM extracts amount, date, and vendor → automatically fills in the spreadsheet → notifies on Slack if there are anomalies.”

This entire flow can operate without human intervention. The previously subjective “eye checks” by the accounting staff become a reproducible system.

—

Part 3: TabPFN-3.5—Finishing Table Data Analysis in “Seconds”

The third component is somewhat different in nature.

TabPFN-3.5 is a foundational model for predictions and classifications on tabular data (like CSV or Excel tables). It has been significantly improved from the previous TabPFN-3 and achieves state-of-the-art accuracy on standard table prediction benchmarks.

What’s beneficial for small and medium-sized enterprises?

1. Minimal learning required.
Traditional machine learning involves gathering data, preprocessing it, selecting models, and tuning hyperparameters, which can take weeks. TabPFN-3.5 can return predictions simply by passing in data. It’s a pre-trained foundational model, so it can be used directly without additional training.

2. High accuracy with small data.
Data in small and medium-sized enterprises is significantly less than that of large corporations. Traditional machine learning often struggles with “insufficient data for accuracy.” TabPFN-3.5 has high generalization performance with small datasets, allowing for practical predictions with just a few hundred rows of data.

3. Strong performance with non-i.i.d. data.
Real-world business data, such as time series data (monthly sales, inventory trends) or relational data (combinations of customers and products), does not follow a “clean independent and identically distributed” pattern. TabPFN-3.5 has a clear advantage here.

Here are some specific use cases:

• Inventory forecasting: Predicting next month’s required inventory based on past shipping data, preventing cash flow issues due to excess inventory.
• Sales forecasting: Monthly sales predictions that account for seasonal fluctuations, improving procurement decision accuracy.
• Customer churn prediction: Identifying customers at high risk of leaving based on changes in transaction frequency and amounts.

TabPFN-3.5 is open-source and free to use. It can even run on a CPU, eliminating the need for additional GPU costs.

—

So, How Much Will It Cost?—An Example Configuration Under 50,000 Yen per Month

Let’s break down the estimated costs:

With a local PC setup, the cost can be under 10,000 yen per month. Even using cloud GPUs, it can be kept to 45,000 yen.

Just a year ago, attempting to do the same would have required a SaaS contract costing 300,000 to 500,000 yen per month, and if you engaged an SIer, you would have to prepare for an initial setup cost of 5 million yen. Costs have dropped to less than one-tenth. This is not just a slight reduction; it represents a structural change.

—

What Happens Beyond “Lower Costs”?

When costs decrease, what changes?

The cost of trial and error decreases.
At 50,000 yen per month, you can try it for three months and stop if it doesn’t work out. That’s just 150,000 yen in tuition. It’s not the same as paying 5 million yen to an SIer only to find out it wasn’t what you expected.

Dependency on individuals is eliminated.
Invoice formats that “only veteran Tanaka in accounting can read” can now be processed systematically. Whether Tanaka is on leave or leaves the company, operations will not stop.

Small and medium-sized enterprises can wield the same tools as large corporations.
Until now, utilizing AI has been the privilege of large companies with “data, money, and human resources.” However, the combination of open-source and local execution is breaking down that wall. In fact, nimble decision-making small and medium-sized enterprises can implement solutions faster. While large corporations take three months for internal approvals, smaller companies can act as soon as next week.

—

Caveats: It’s Not a Silver Bullet

To avoid being irresponsible by just hyping things up, here are some realistic caveats:

1. Confirm the commercial license for jina-ocr-v1.
While the weights are open, the conditions for commercial use should be clearly verified. API usage is fine, but running it commercially on your own server is currently in a gray area.

2. Building the harness requires a minimum level of technical skill.
Setting up Ollama and configuring agents is not as easy as using no-code tools. You need at least one person in-house with basic Python knowledge. If not, it’s practical to hire an external technician for the initial setup (costing around 100,000 to 300,000 yen).

3. Don’t expect 100% accuracy.
Both OCR and LLMs can make mistakes. Initially, it’s wise to maintain a human check in the flow and gradually expand the automation scope.

—

So, What Should You Do?

First, select one type of “paper document that humans are currently reading by hand” in your organization. It can be an invoice, a purchase order, or an inspection report.

Next, try reading 10 documents using the jina-ocr-v1 API. Look at the outputted Markdown and see if you think, “This is usable.” The cost up to this point is almost zero.

If it seems usable, set up a local model with Ollama and create a flow for extraction → formatting → entry. You can introduce TabPFN-3.5 once you have accumulated data.

Don’t try to do everything at once. Start with one document and 10 samples.

The components are in place. Now, all that’s left is to test them.