Small businesses run on documents: supplier invoices, price lists, packing lists, contracts, certificates, email threads. Large language models (LLMs) — ChatGPT, Claude, Gemini and their peers — have learned to read all of it and answer questions about the content. The temptation is obvious: hand the paperwork over to a machine and get back to selling. But the technology has a clear boundary: in some tasks an LLM saves you hours every single day, while in others one unnoticed error costs more than all the savings combined. Let's map out where that boundary runs.
Where LLMs genuinely read documents well
The strength of language models is working with the meaning of text, not with exact numbers. In practice, that means:
- Sorting and routing. Identifying what just arrived — an invoice, a proforma, a price list, a complaint, or a newsletter — and filing it in the right folder. The model rarely gets this wrong, and a mistake costs minutes.
- Summaries. Condensing a three-screen supplier email into five lines: what they're asking for, by when, and what has changed.
- Draft data extraction. Pulling line items, quantities, and prices out of a PDF invoice into a table — as a draft that a human will verify.
- Version comparison. Spotting how the new price list differs from the previous one, or which contract clauses have changed.
- Draft replies. Turning your short note into a polished email to a supplier in English.
What these tasks have in common: the result is easy to check at a glance, and any error is reversible.
Where you can't trust it: money and customs
Payment details and amounts
An LLM generates text — it doesn't read it byte by byte. The model can mix up similar-looking digits, "complete" a plausible account number, or drop a row in a long table — and it will do all of this in a confident tone. A separate risk is payment-detail fraud: a system that simply "reads email" won't tell a genuine supplier's message from one sent off a look-alike domain. That's why banking details, final amounts, and the currency must be verified by a human before any payment goes out — against the original document, not the AI's retelling of it.
HS codes and customs documents
For a business running import operations, product classification determines the duty rate, certification requirements, and the risk of a customs reassessment. An LLM will happily suggest an HS code, and it will look convincing — but the model carries no liability and may not know the current section notes, classification rulings, or the practice of a specific customs office. The same goes for certificates of origin, delivery terms, and duty reliefs: AI is good for preparing questions for your customs broker, not for replacing the broker's signature.
The working rule: AI proposes, a human approves
It helps to split your document workflow into two zones:
- The green zone — sorting, summaries, drafts, reminders. Errors here are reversible: let it run automatically.
- The red zone — anything touching money, customs, or legal obligations. AI prepares the groundwork; a human makes the decision.
The second rule is source verification. If the system extracted an amount from an invoice, a link to the actual PDF and the exact spot the number came from should sit right next to it. In the red zone, a retelling without the primary source doesn't count.
Where a small business should start
- Start with email: classifying supplier messages and generating summaries is the fastest win at minimal risk.
- Turn on invoice data extraction in "draft plus review" mode, not as an autopilot.
- Define your red zone up front: a written list of operations the AI never performs without human confirmation.
- Measure the payoff in hours: our calculator is a quick way to estimate how much time processing incoming mail takes today.
We break down typical automation scenarios in our blog, and answers to common questions are collected in the FAQ.
This is exactly the logic behind HORUVIA: the AI operator sorts supplier email and prepares data from invoices on its own, but always leaves money and customs decisions to the owner.