An enterprise AI assistant inside your closed network

WIZICO Private AI is an enterprise AI assistant that runs on the organisation’s own server without internet access. You upload your documents: policies, instructions, contracts, technical documentation, knowledge bases, regulations. An employee asks a question in plain words, by text or voice, and the assistant finds the answer in those documents and cites the source. Separate assistants with their own document bases can be set up for different teams, for example lawyers, HR or IT support. The assistant finds the facts and drafts the answer; the employee makes the decision.

A public AI chat processes requests on the provider’s servers, usually abroad, so internal data cannot be sent there. It answers from the model’s “memory”, knows nothing about your internal documents and does not show where the answer comes from. WIZICO Private AI runs inside your network without internet, answers only from the documents you upload, shows the source for every statement and says plainly when the documents contain no answer.

Any documents with text: policies, orders, instructions, contracts, technical and project documentation, knowledge bases, reply templates, regulations. PDF, Word, Excel and text files are supported, and scans are recognised on upload. Documents can be split into bases per department, with rules on who can access which base. Your administrator uploads and updates the documents: a new version replaces the old one, and once it’s uploaded the assistant answers from it.

The assistant follows the RAG (retrieval-augmented generation) approach: it first finds passages in your documents, and the language model writes the answer only from them. If there is no relevant passage, the assistant says the documents contain no answer instead of inventing one. Software checks every citation against the retrieved passages and flags numbers without a source; mismatches go to the request log for review. The final decision always rests with the employee.

Russian and Kazakh: the assistant understands questions in both languages and searches documents in each of them. For Kazakh, morphology is handled so a word is found in any form, and voice input uses a model fine-tuned on Kazakh speech. Uzbek, English and other languages are added per project, with search, word-parsing rules and speech recognition set up for them.

A workstation with an NVIDIA GPU with 8 GB of memory is enough for the pilot. For everyday use by 20–30 concurrent employees we recommend a server with one GPU of 48 GB or more, a CPU with 16+ cores, 64 GB of RAM and two NVMe drives in RAID1. If you expect more load or a larger language model, a configuration with 96 GB of video memory and 128 GB of RAM fits. We can size a server for your tasks or install the assistant on hardware you already have.

We start with a free demo: you send your employees’ typical questions and we show how the assistant answers with citations, using open documents such as the laws and standards of your industry; your documents aren’t needed at this step and are checked in the pilot. Then we run a pilot on your hardware: we install the assistant without internet, load your documents, set up roles and access and check quality on control questions. Cost depends on the number of users and departments, the volume of documents and the hardware, so we quote after a short call. Support is a separate agreement.