Why ChatGPT is not an employee
An agent built with Nova does not wait for someone to open a chat window.
“We already have AI.”
We hear this in almost every other customer conversation, usually shortly after we have put the word “agent” on the table. What people mean, as a rule, is that a few licenses for ChatGPT or Microsoft Copilot have been bought, that the marketing department now writes its copy faster, and that someone in IT recently built a script with Claude that actually works.
All of that is fine, and we do not want to spoil anyone's licenses. But it has about as much to do with what we mean by an AI agent in a company as a very well-read intern has with a working accounts department. Both are useful; they simply solve entirely different problems.
Because this misunderstanding is now so widespread that it stops projects before they have even started, we want to sort out properly who plays which role in the current AI circus, and where the real line runs between all of that and what we do with Nova.
Who is who?
First, a distinction that almost always gets lost in public discussion: there are language models, and there are products built on top of them. GPT is a model, ChatGPT is a product; Claude is both the name of the model family and of Anthropic's app. The model is the engine, the product is the car around it, including the steering wheel, the heated seats and the question of where the trip data goes.
The well-known names are all products for people who want to ask, write, research or program something. They differ in temperament and strengths, but they belong to the same species:
| Product | Provider | What it is at its core |
|---|---|---|
| ChatGPT | OpenAI | The generalist that took the subject mainstream: chat, text, images, research, all in one window |
| Claude | Anthropic | Strong on longer texts, analysis and above all code; with Claude Code also a developer tool |
| Gemini | Deeply integrated into Google's world of search, mail and Docs | |
| Copilot | Microsoft | The assistant in Office and Windows, where the work happens anyway |
| Grok | xAI | Closely tied to X, with a humor of its own and access to whatever is being discussed there right now |
| Perplexity | Perplexity AI | An answer engine that searches the web and delivers its sources along with the answer |
What they all share is the basic principle: a person asks a question or gives an instruction, the system answers, and then it waits patiently for the next input. Without a prompt, nothing happens. These tools are brilliant conversation partners, but they have no job, no rota and no responsibility for what happens in the company after the conversation.
And then came the agents
For a good year now the industry has had a new favorite word, and with good reason. With Claude Code, the agent features in Claude or agent mode in ChatGPT, the systems no longer just answer, they act: they read files, write and test code, operate the browser, create documents and work through a multi-step task while you go and get a coffee.
We use these tools ourselves every day, and anyone who has watched Claude Code finish in an hour a refactoring that used to take a day does not want to go back. Even so, it is worth looking closely at who these agents actually work for.
The answer is fairly clear: for exactly one person. The agent sits on an employee's laptop or in their account, it is started by that employee, it works with their permissions and their files, and it stops when the task is done or the quota is used up. It is an extremely capable personal assistant, and one that needs its boss before it can start at all.
For individual productivity that is enormous leverage. For running a company, for the processes that have to run every day regardless of who is on holiday, it is only half the way.
What a business agent really is
An agent as we build it with Nova does not wait for someone to open a chat window. It has a fixed role in an operational process, is triggered by events (a new goods receipt, a change request in the ITSM, an alert from the SOC, a month-end close) and then does its work, at three in the morning too, and also when the colleague in charge is lying by the pool in Mallorca.
That sounds like a small difference, but it changes everything that follows. An agent in a process needs its own identity with precisely defined permissions. It works directly in the systems where the work takes place, that is the ERP, the ticket system, the firewall configuration or the IAM, and it has to document every decision so that an auditor can still follow it two years later. Where it reaches the limit of its mandate, it escalates to a person, with everything that person needs to decide quickly.
In practice it looks like this, for example:
- In logistics, an agent matches order, delivery note and invoice, clears up discrepancies with the supplier and puts only the real problem cases on the table.
- In IT operations, an agent checks every change request against maintenance windows, dependencies and approval rules before it even reaches the CAB.
- In security, an agent recertifies firewall rules and permissions, finds the rule from 2019 that nobody can explain any more, and asks its owner.
- For NIS2, DORA or TISAX, an agent continuously collects the evidence that would otherwise be scraped together in a panic three weeks before the audit.
The difference from personal agents can be summed up at a glance:
| Personal agent (Claude Code, ChatGPT & co.) | Business agent (Nova) | |
|---|---|---|
| Trigger | A person gives an instruction | Events, schedules, thresholds in the process |
| Works for | One person | A process or an organizational unit |
| Permissions | The user's | Its own role, least privilege, reviewed separately |
| Result | A text, a piece of code, a file | A completed work step in the target system |
| Traceability | Chat history | Tamper-proof audit log of every decision |
| When nobody is there | Nothing happens | Keeps running, escalates by rule |
In short, a business agent does not automate the typing, it automates the work itself: workflows, checks, controls and the many small decisions that today sit in Excel lists, email loops and the heads of individual employees.
And if I still want to talk to ChatGPT?
You can. Through open interfaces such as the Model Context Protocol, Nova agents can technically also be addressed from ChatGPT, Claude or Copilot, and later so can the agents customers build with Nova themselves. So anyone who insists on asking in the familiar chat window how many changes were rejected this week, and why, can do so.
We still advise against it, at least for anything beyond a harmless status query. The reason is simply that every question and every answer then passes through a third party's infrastructure, and you have no influence over its model changes, its logging or its data retention. For a company under KRITIS, NIS2 or DORA that is not a matter of taste, and a chat window that talks to the internet and to your own core processes at the same time is, from a security point of view, an opening for prompt injection that nobody should invite in voluntarily.
So the chat window is at best the remote control. The actual work, the rules, the permissions and the data belong where the company is in control, and that is exactly why there is Nova Gate to secure every inference and Nova OneSite to run the entire AI in-house.
Where Nova comes in
Nova is therefore not another chatbot but a family of agent products, each of which docks onto a different part of operations:
- Nova Change brings digitalization, automation and AI transformation into the business processes.
- Nova Run takes over routines, checks and workflows in IT operations.
- Nova Defense defends actively at the network, system and application layer, automated through to the response, and provides the CISO.
- Nova Gate routes requests to the right model, detects confidential data locally and lets it out only masked, or not at all.
- Nova OneSite is your own AI on site, a single place for your own data instead of spread across the cloud.
So the next time you hear “We already have AI”, feel free to ask how many processes that AI completed last night without anyone asking it to. If the answer is “none”, we should talk.