Custom Agent
A custom agent is a self-configured AI program that performs a fixed task for a specific person or company. For this, it is given its own instructions, its own data, and access to selected tools such as calendars, databases, or email.
Many AI programs can answer questions, but they do so for all users in the same way. A custom agent is such a program that someone has set up for a very specific purpose. You give it fixed instructions, for example: answer customer inquiries about our bicycles. You also give it its own documents, such as price lists or manuals. And you allow it to use certain programs, such as the appointment calendar or the ordering system. This turns a general tool into a specialized helper with a clear mandate.
Why companies don’t stick with standard AI
A general AI knows the internet, but not your company. It doesn’t know how much a spare part costs at your business, which colleague is responsible for complaints, or what your warranty terms are. Yet exactly this knowledge determines whether an answer is useful. A custom agent closes this gap because it is given the internal information.
The second reason is reliability. A general model answers differently depending on how the question is phrased. For customer service, this is a problem. Fixed instructions narrow the scope: the agent should never promise discounts, always ask for the order number, and hand off to a human for legal questions. This makes the behavior more predictable.
Economically interesting is the time saved. A task that costs an employee ten minutes takes the agent seconds. With a thousand transactions per week, this adds up. That’s why custom agents now appear as their own business segment in software companies' quarterly reports.
Instructions, knowledge, and tools
At the core of a custom agent is a perfectly ordinary large language model, i.e., an AI that understands and generates text. This model itself is usually not changed at all. What is changed is what is presented to it before the user even asks anything. This fixed preamble is called a system prompt and describes role, tone, and prohibitions.
Added to this is its own knowledge. Company documents are stored in a database. When someone asks a question, the system first searches for the matching text passages and presents them to the model along with the question. This process is called retrieval, meaning targeted lookup. The advantage: if a price list changes, you swap out the document instead of retraining the AI.
The third building block is tools. These are interfaces to other programs that the agent is allowed to call itself. It can thus schedule an appointment, look up an order, or send an email. Only this ability to act distinguishes an agent from a pure chatbot that merely talks. A common misconception is that a custom agent is specially trained. In the vast majority of cases, it is merely configured, and that takes hours instead of months.
From support chat to your own GPT
Custom agents are most commonly encountered in customer service. If a chat window on a shop page knows your order number and states the delivery status, exactly such a system is often behind it. Internal helpers are also widespread: an agent for the HR department answers questions about vacation requests based on the actual company agreements.
For private individuals, the same idea exists in a simpler form. At ChatGPT, such self-built variants are called GPTs; with other providers, Assistants, Gems, or Projects. You describe in plain English what the helper should do, upload a few files, and you’re done. A student can build an agent that only quizzes from their own biology notes this way.
In business news, you usually encounter the term in connection with enterprise software. Providers like Microsoft, Salesforce, or ServiceNow sell toolkits for such agents. One point remains important: a custom agent is only as good as the documents it receives. If the internal documents are outdated, it gives outdated information with full confidence.