AI agents for companies are useful when they have access to the right data and clear limits. An AI agent is a language model with tools: read a table in your ERP, search a knowledge base, draft a document, book an appointment. Faced with a request, it decides which tools to use and in which order, then answers or proposes an action. A useful agent has good access to your data; a safe agent can only do what you allowed, and logs every step. Our article AI agents: what they are and what they really do has examples from SMEs.
Artificial intelligence
AI agents for companies: custom development inside your software
We build AI agents for companies that work inside your software: they answer questions on management software data, prepare reports and exports, route requests and fill in documents for someone to confirm. Each agent only uses the tools we give it, with the user's own permissions, an action log and a controlled cost per request. You get a quote after a free analysis.
Book a 30-minute call
Updated on
What an AI agent is, in practice
A real example: the Aurelia Hub assistant
In Aurelia Hub, software for jewellers in Germany, the assistant answers questions and exports data. It queries 26 data sets through declared tools, always read-only; it inherits the user's permissions; sensitive columns are never described to the model; exports are private, signed and expiring; and a set of 40 test questions measures answer quality whenever the prompt or model changes.
How we build an AI agent
-
Task and boundaries 1 week
What it does, on which data, what it must never do, who confirms.
-
Tools 2-week sprints
Tools with permissions and limits, like any other software function.
-
Test set In parallel
Real questions with expected answers to measure errors before production.
-
Production and review Ongoing
Conversation log, cost per request, regular review of wrong answers.
Costs, models and data
Every request has a model usage cost; we estimate it during the test phase and cap it with time and spending limits. We use Anthropic, OpenAI or Google models through business APIs that do not train on customer data, with EU processing where needed. See also AI for companies and the AI phone system.
Got a task for an agent?
Message us on WhatsApp with the question you hear ten times a day or the document you retype by hand.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers with text it knows. An agent uses tools: it reads ERP data, drafts, books, routes. That is why it needs permissions and limits.
Can an agent change data in my software?
Only if we decide so together, usually with human confirmation. Many useful agents are read-only or prepare drafts.
How do you check the agent is not wrong?
With a test set of questions and expected answers, a conversation log and human confirmation for actions that matter.