AI assistants & agents

General tools like ChatGPT are trained on broad world knowledge but do not know your company. When an employee asks about an internal procedure, they either guess or say they do not know. Staff pasting confidential documents into personal accounts is also a serious data risk. An enterprise AI assistant connects the same language model technology to your own data, under your own rules.

The differences are concrete: the assistant shows which document an answer came from, never sees documents the user is not cleared for, says it does not know rather than inventing an answer that is not in the sources, and every conversation is stored in a log your company controls. Model provider terms are chosen so that your data is not used to train their models.

In short, the difference lies less in the model itself than in the structure built around it: data connections, permissions, logging and evaluation. Building that structure is exactly what we do.

RAG (retrieval-augmented generation) means the assistant searches your document store before it writes an answer. Documents are split into small passages and indexed by meaning; when a question comes in, the most relevant passages are retrieved and handed to the model with an instruction to answer only from them. The model uses your current knowledge and cites its source under the answer.

RAG quality depends more on data preparation than on model choice. If three versions of the same policy sit in the archive, the assistant will hesitate too. That is why we make document ownership, update flows and permission tags part of the project. Difficult sources such as tables, scanned PDFs and e-mail archives get their own processing steps. When a document changes, the index refreshes automatically, so the assistant does not keep describing an outdated procedure.

An assistant answers questions; an agent gets work done. An AI agent plans several steps on its own using defined tools: it reads a customer request, finds the record in the CRM, checks the order status, drafts a reply and opens a task if needed. A document-reading agent takes an incoming invoice or application form, extracts the fields, checks them against your rules and prepares them for entry into your system.

Agents are powerful but should never be left unbounded. For each agent we define which tools it can reach, which actions it may only suggest and which it may carry out, and at what threshold it hands over to a person. For hard-to-reverse steps such as moving money, sending formal correspondence to customers or deleting records, approval always stays with a human. Every action is logged with its reasoning.

The most common mistake is starting with ‘let’s add AI’ without a clear business goal. The second is opening up an assistant that looked impressive in a demo but was never measured against real questions. The third is ignoring cost and latency; calling the largest model for every question can be both expensive and slow. In a well-designed setup, simple questions go to lighter models and complex ones to stronger models.

The marks of a good project are these: accuracy was measured against a test set before launch, answers cite their sources, the assistant says so when it does not know, users can flag a bad answer in one click and those flags are reviewed regularly. Who asked what, and at what cost, is also visible. Without this discipline an AI project stays an experiment; with it, it becomes a lasting part of daily work.