Enterprise AI Assistant (AICortex)
A conversational assistant working on the company's real documents, with references to sources. The processing part splits documents into chunks so the assistant's knowledge stays current.
- Domain
- Conversational AI and information retrieval
- Target customers
- Senior managers, admin staff, teams that need fast access to organizational information
- Main capabilities
- Live conversation with document references, tool calls into other services with human approval, per-company data isolation
AICortex is the organization's AI assistant platform: a conversational assistant equipped with the company's real data and documents, which shows its answers with references to sources and can connect to other platform services through defined tools to return live information or, in the future, carry out an action. The product has two parts: the main service for conversation and assistant execution (AICortex), and a behind-the-scenes processing service (AICortexWorker) that reads the organization's documents, splits them into chunks and prepares them for intelligent search, so that the assistant's knowledge base is always up to date.
An important and candid note on the current state: the infrastructure of this product has been built end to end and validated with real tests (a real user sign-in, a real browser, a real call to another platform service), but it has not yet been made fully available to all users. The assistant's service-to-service connection to the platform's identity and authorization system was only recently completed, and the assistant page is not yet linked in the dashboard's main menu. The knowledge base today is built from a small number of sample documents, to prove the correctness of the whole chain. Connecting to the organization's real and complete documents is the next stage. Conversation replies currently come from a trial engine until the connection to the final AI engine is established. This is a controlled decision for gradual rollout and is not a technical defect.
Key capabilities
Live streaming conversation
The assistant's reply is shown as a live stream (word by word) on the dashboard's assistant page, rather than all at once after a long delay.
Answers retrieved from real documents with source references
Each answer is built from a hybrid search (keyword search and semantic search together) over the company's documents and is displayed with a reference marker to the source document, so the user knows where the answer came from.
Data isolation per company
Public content is visible to all companies, but each company's private content is visible only to that company and never leaks to another company. This isolation has also been proven in a real test.
Tool calls into other platform services
To answer, the assistant can read live information from other platform services (such as looking up the type of an accounting voucher), using exactly the identity and access level of the user who asked the question.
Action control and human approval
Any action beyond simply reading information (that is, any action that would change something) is automatically paused and waits for explicit approval from a person. The assistant never performs, on its own, an action that changes the state of data.
Precise access control by user and company
The assistant knows and does only what the asking user is permitted to. Even the ownership of each conversation is checked against its current company, so a user who belongs to several companies cannot leak a conversation's messages into another of their companies.
Verifiable event history
Every message and every action of the assistant is recorded in a tamper-proof history chain, suited to the audit and compliance needs of a financial platform.
Knowledge base update pipeline
The behind-the-scenes processing service continuously reads new documents, splits them into meaningful chunks (taking the structure of Persian text into account), builds an intelligent index and adds it to the assistant's search engine. No document is processed twice, and an error in one document does not stop the rest.
Replaceable AI engine
The architecture was designed from the start to connect to different AI engines. Today it works with a local trial engine, and it can be connected to a real commercial or cloud engine without redesign.
Business value
To find a simple answer (a rule, a procedure, a number from a system), staff usually have to search through several pages and several documents. AICortex turns this search into a simple question in natural language and returns the answer with a clear source, without stepping outside the platform's security and authorization framework. Because every action beyond reading information needs human approval, this assistant is a safe support tool, and it is not an autonomous system that makes decisions without oversight. The entire path (identity, authorization, company data isolation, history recording) uses the platform's own security infrastructure, so the product does not operate as a separate, unsupervised AI tool next to the core systems. It is part of the existing security and compliance framework.
What sets it apart
- Security and authorization are implemented in the main platform layer, and not in a separate layer. In a real test, even deliberate incorrect attempts (wrong password, unauthorized company) were correctly rejected.
- Answers always come with a reference to the source document, instead of generic answers with no root in the organization's real data.
- Complete isolation of each company's data in search and in conversation ownership, even for a user who belongs to several companies.
- No state-changing action runs without explicit human approval.
- Verifiable history for every message and every action, matching the compliance needs of a financial platform.
- A design independent of any specific AI engine, ready for easy replacement of the engine in the future.
See this module on demo data
In a demo session we walk through your organization's scenarios on Dara's demo environment and answer your technical and finance teams' questions.