The conversation that started SoloScan happened over coffee with a friend who works as an auditor in Sweden. He had been watching everyone in his life use ChatGPT to summarize, draft, analyze, search. He wanted in.
His question was simple. Could he use AI to help him work through client financial statements, payroll records, and the rest of the documents that pile up during an audit?
The honest answer was no. Not the tools he had heard of.
This article is about why that answer was no, and what changes when the AI runs on your own computer instead of someone else's.
The problem with cloud AI for confidential work
Tools like ChatGPT, Claude, Copilot, and Gemini are powerful. They are also hosted services. When you drop a document into the prompt box, that document leaves your computer. It travels over the internet, lands on a server you don't control, gets processed, and may be retained for some period of time.
For most people, this is fine. A high schooler summarizing a Wikipedia article does not care.
For an accountant or an auditor, this is the entire problem.
The documents you handle are not yours. They belong to the client. The contract you signed with that client almost certainly forbids sharing their data with third parties. The professional body you answer to has rules about confidentiality that go further than the contract. The country you work in has data protection laws (GDPR in the EU, similar regimes elsewhere) that go further still.
Uploading a client's payroll ledger to a cloud AI service is, in most jurisdictions, a clear violation of at least one of these obligations. It does not matter that the AI provider promises not to look at the file. The act of sending it is the violation.
So a competent accountant or auditor reads the news about AI, watches their friends in other professions get faster, and then sits in front of a stack of documents that they cannot legally feed into any of these tools.
What "local AI" actually means
Local AI is software that runs the model on the user's own computer. The document never leaves the machine. The internet is not involved in the analysis. The AI provider, if there even is one, has no visibility into what you are doing.
This is not new technology. Researchers have been running models locally for years. What is new is that the models small enough to fit on a laptop are now good enough to be useful for serious work.
A model like Meta's Llama 3.2 runs on a modern laptop. It reads a contract, summarizes a report, answers questions about a financial statement. The quality is not GPT-4. It is closer to ChatGPT from a year or two ago. For most document analysis work, that is more than enough.
The trade-off is honest:
- You give up some quality compared to the largest cloud models.
- You gain complete control over where the data goes.
For an accountant or auditor, that trade is not even close.
The gap SoloScan fills
Running a local model is technically possible today. Ollama, LM Studio, Jan, AnythingLLM and others let you do it. They are excellent tools.
They are also built for developers. They expect you to know what a quantization is, why a 7B model fits in RAM but a 70B does not, how to set up a vector database, how to write a system prompt.
The accountant pulling an all-nighter on Q3 reconciliations does not have time to learn any of this. They want to drop a PDF in a window and ask "what is the total VAT outstanding across these invoices."
SoloScan is the layer between those two worlds. The model still runs locally (we use Ollama under the hood). The privacy story is identical. But the experience is a desktop app that opens, accepts a document, and answers questions about it. No terminal. No configuration. No vector databases to manage.
What runs on your machine, and what doesn't
For the curious, here is exactly what happens when you use SoloScan:
- You drop a PDF, DOCX, or scanned image into the app.
- SoloScan reads the file. PDFs use a text-layer extractor first; if the PDF is a scan, the app falls back to OCR (Tesseract, also running locally).
- The extracted text goes to a local Ollama instance running on your machine, which runs Llama 3.2 (text) or Llama 3.2-Vision (scans).
- The model's answer comes back to the app and appears in the chat.
- The document and the conversation are saved to a local SQLite database in your user folder.
At no point in this flow does any client data travel over the internet.
What does travel over the internet:
- A license check (so we know your subscription is active).
- An update check (so SoloScan can install new versions).
Both are inspectable in any network monitor. Neither involves your documents.
What this does not solve
We are not selling magic. A few honest things SoloScan does not do:
- It is not as good at deep reasoning as the largest cloud models. If you need GPT-4-level analysis, no local model on a laptop will give you that today.
- Handwriting is hit or miss, and for right-to-left scripts like Arabic, scanned-image quality varies. Text-based Arabic PDFs read well; handwritten or low-quality scans may need a closer review.
- It does not replace your judgment. The AI is an assistant. The final reading of a contract or a financial statement is still your work.
We are also clear-eyed that local AI is not the right answer for every profession. If you are writing marketing copy, drafting emails, or analyzing publicly available data, the cloud tools are faster and often better. Local AI exists for the work where the privacy constraint is not optional.
Why we built this
This whole product exists because one auditor asked one question that did not have a good answer.
We assume there are many more people in that position. Accountants, auditors, lawyers, therapists, consultants, researchers, anyone whose work depends on keeping someone else's information confidential.
If that's you, SoloScan is free to download and try. We would rather you decide for yourself whether it fits than read a longer version of this article.
If it's not you but it might be someone you know, forward this along.