# An internal AI assistant for businesses that reads your company > An internal AI assistant for businesses, connected to contracts, accounting and internal docs, that answers with a verifiable source, not guesswork. URL: https://thenichesociety.ro/en/ai-engineering/document-assistant An internal AI assistant for businesses doesn't guess like a generic chatbot from the internet — it answers with a verifiable source, drawn from the company's real contracts, documents and databases. **We built one such assistant** for a real accounting use case, with 20 tools for reading and verification. ## The difference between searching and knowing An assistant on your company's documents isn't a general chatbot you paste files into — it's a system that indexes the company's real knowledge (contracts, internal manuals, ledgers, policies) and answers by citing the source, not with a plausible guess. The difference matters most where a wrong assumption costs money or reputation: a misread tax rule, a misread contract clause, an internal rule applied incorrectly. A business chatbot built without internal sources stays generic — the real difference comes from connecting it to the company's own documents and databases. ## The reader, the brain, the workflows A solid internal assistant has three distinct layers: the data reader (the technical part, invisible to the user — reads native formats, even without an API), the domain brain (the industry-specific rules and knowledge, the real differentiator), and the workflows (reconciliation, checking, diagnostics — the part that translates directly into hours saved). A language model without domain knowledge gives answers that sound plausible and are wrong — we saw this directly in a pilot, where a wrong assumption about how often a tax return had to be filed was only caught at human review. Technically, the mechanism behind it is called RAG for businesses (retrieval-augmented generation): the model searches your data first, then formulates the answer based on what it found. You find the same isolation rule on the Consumer Voice platform, where [isolating data per client instance](https://thenichesociety.ro/en/case-studies/consumer-voice) was a deliberate decision, not a technical accident. ## What stays a human decision The right positioning for an internal assistant is "copilot," not "autopilot": it reads, calculates, reconciles, checks filings before submission, prepares import files — but the treatment decision, the signature, and the filing stay with the human. The argument is not just caution, it is commercial: „the AI does your job on its own” is a hard sell to someone whose professional liability is on the line. What the client is actually buying is trust and speed, not an abdication of responsibility. An internal AI assistant for businesses stays a copilot, not an autopilot: it proposes an answer with a verifiable source, but the human decides whether to send it on or sign it. For companies where isolation by dedicated database is not enough either, we explain separately what [self-hosted AI, with your data on your own infrastructure](https://thenichesociety.ro/en/blog-ai-suveran-self-hosted-ue) means, not in an external provider’s cloud. ## Many important systems have no API A good portion of the systems used daily in Romania (desktop accounting, older ERPs) don't have and won't get an API anytime soon. That doesn't block an internal assistant — it just means reading happens directly from the native formats (local databases, bank statement PDFs), taking care to never write directly to the data. The golden rule: direct reads yes, direct writes never — any change goes through the source system's official import flow. A well-built AI document assistant becomes, over time, a kind of second brain for the company — the collective memory of its processes and decisions, instantly searchable. For data that can't leave the company, the assistant runs on a [private AI infrastructure](https://thenichesociety.ro/en/ai-engineering/ai-infrastructure), on-premise or in the EU. ## Professional secrecy isn't optional For any assistant that works on client data (accounting, legal, contractual), per-client data isolation is a design requirement, not something bolted on later. Every deployment has its own space, its own keys, with no mixing between portfolios. For an accountant with 30–80 companies in their portfolio, that means designing the configuration around the whole portfolio from the start, not around a single company with ad hoc expansion later. Each client’s data stays isolated in its own separate database — the same rule we apply on multi-client platforms like Consumer Voice, not only on internal assistants. ## From a pilot case to an internal product A good internal assistant starts from the company's real knowledge sources, not from a generic list of "what AI can do." - 01We map the company's real knowledge sources (documents, databases, internal policies) - 02We build the reader + the domain rules specific to your industry - 03We test on real cases, with explicit human validation at every critical step - 04We hand over the system + documentation, with installation/setup included ## A fixed price, after a short discovery call. - second brain architecture - connection to the company's data - written acceptance criteria - access to the company's documents - answers with a source - monthly maintenance - everything N1 includes - cost/quality routing - role-based permissions + audit Every project has a different context, workflows and infrastructure, so the price is set after a short, paid discovery and does not change along the way. ## Frequently asked questions ### What is an internal AI assistant (second brain)? It's a system connected to the company's real documents and databases, which answers questions by citing the exact source, instead of generating a plausible answer from the model's general memory. Technically, it's based on RAG (retrieval-augmented generation) and connectors that read the company's native data. ### What is RAG, and why does it matter for a company? RAG means the model first searches the company's real documents, and only then formulates the answer, citing the source. It matters because it eliminates generic AI's biggest risk — plausible hallucination — on specific information that doesn't exist in the public data the model was trained on. ### Can an internal chatbot work on systems with no API? Yes — when the source system has no API (common with desktop accounting or older ERPs), the assistant reads the native data formats directly, under the strict rule of never writing into them directly. Every change goes through the system's official import flow. ### How much does an internal AI assistant cost for a company? A simpler assistant, built on a single set of documents, can cost far less. ### Does company data get used to train the AI model? No, if you use a serious enterprise/API provider — Anthropic, for example, doesn't train on commercial data sent through Team/Enterprise/API. That said, where the data is processed (data residency) and whether a self-hosted solution is needed for full sovereignty remains a separate discussion. ### Can the internal AI assistant make decisions on its own? The recommended position is "copilot, not autopilot": the assistant reads, calculates, checks and flags, but the treatment decision and the final sign-off stay with the human. It's as much a commercial choice as a prudent one — professional liability stays human. ### What does data isolation between clients mean for an AI assistant? It means every client or company has its own data space, its own keys, and zero mixing with other portfolios — a requirement especially for accountants or consultants managing dozens of companies, where professional secrecy and GDPR aren't optional. ### What kinds of documents can an internal AI assistant actually read? Contracts, invoices, internal correspondence, procedures, records from SAGA or another accounting system — any structured or semi-structured source we connect explicitly, with clear role-based permissions, not everything that happens to exist on a file server. ### Does the assistant answer in Romanian? Yes. Current models work well in Romanian, and the assistant answers in the language it's asked in, with a reference to the document the information came from. ### Let's see what can be automated in your business. A free 30-minute session: we'll tell you what can be automated, how long it takes and what it costs, with a fixed price after discovery. We reply the same business day.