Table of Contents
A chatbot answers questions in a conversation and changes nothing outside it. A copilot works inside a person's tools, drafts and suggests, and the person decides and executes. An AI agent is given a goal and pursues it: it plans, calls tools, and acts on other systems, with humans approving only at defined checkpoints. The difference is who acts.
Key takeaways
- All three can run on the same large language model. What separates them is initiative, tool access and where a human approves.
- A chatbot's worst case is a wrong answer that a person reads. An agent's worst case is a wrong action in a live system.
- A copilot keeps a qualified person as the one who executes every action.
- OWASP lists "excessive agency", meaning too much functionality, permission or autonomy, among its top risks for LLM applications.
- For OT, CISA and partner agencies advise limiting active control by AI without a human in the loop.
What is a chatbot?
IBM defines a chatbot as "a software application that communicates with people through text or voice". Older chatbots follow scripts and decision trees. Newer ones use a language model to interpret the question and write the answer.
The defining trait is that the output stays in the conversation. The chatbot answers; a human reads the answer and decides what to do with it.
In a plant, a typical chatbot answers questions about procedures, equipment manuals or specifications, ideally grounded in the site's own approved documents and in the operators' working languages. This is one of the uses of an on-premise LLM such as DALILE, which runs on site and supports Arabic, French and English. The approach is covered in detail in our guide to RAG over maintenance manuals.
What is a copilot?
"Copilot" started as a product name and became a category. Microsoft describes GitHub Copilot as "an AI coding assistant" and Microsoft 365 Copilot as a tool that "enhances workflows across Microsoft 365 applications".
What makes something a copilot is its position. It sits inside the tool where the work happens and sees the context of that work: the open document, the record being edited, the code being written. It drafts and suggests. The person accepts, edits or rejects, and the person performs the action.
In industry, a copilot might draft a work order in the maintenance system from a technician's notes, or prepare a first version of a lab report. The planner or the engineer still checks it, signs it and submits it.
What is an AI agent?
Anthropic separates two kinds of LLM systems. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths". Agents are "systems where LLMs dynamically direct their own processes and tool usage". CISA and its partners define AI agents as software that can "process data, perform decision-making capabilities, and initiate autonomous actions".
Two things make an agent: it chooses its own next step, and it can act on systems outside the conversation. An industrial example is an agent that notices a recurring deviation in exported historian data, pulls past work orders for that equipment, opens a draft notification in the maintenance system and alerts the on-call engineer.
Autonomy has a price. Anthropic notes that agents trade latency and cost for performance, and that their autonomy brings "the potential for compounding errors". Our earlier article on agentic AI for industrial process automation covers the use cases in more depth.
How do chatbots, copilots and AI agents compare?
| Chatbot | Copilot | AI agent | |
|---|---|---|---|
| Who initiates | The user asks a question | The user, while working in a tool | A goal, a schedule or an event; the agent picks the next steps |
| Tools it can use | Search over documents, at most | Read the user's current context; write drafts in that tool | Several systems through APIs: read, create, update, notify |
| Autonomy | None outside the conversation | Low; it proposes, the user disposes | Medium to high, inside the limits it is given |
| Human approval | Not needed, nothing is executed | Every action, by design | At defined checkpoints, or only for high-impact actions |
| Typical industrial example | Answering a question about a startup procedure from the approved manual | Drafting a work order or a lab report for an engineer to check and sign | Compiling a daily deviation summary from historian, LIMS and maintenance data and opening draft notifications |
| Main risk | A wrong or invented answer that someone trusts | A person approving drafts without reading them | A wrong action in a live system, repeated before anyone notices |
How can you tell which one a vendor is really selling?
The labels are loose. Gartner has warned about "agent washing", the rebranding of AI assistants, robotic process automation and chatbots as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI are real.
Five questions cut through the label:
- Who starts the work? Microsoft's own documentation draws this line: one type of agent relies on "user-initiated interactions", while another can "trigger actions automatically, even without direct user input". If the user must always type first, you are buying a chatbot or a copilot.
- Which systems can it write to, and under which account? If the answer is none, it is not an agent.
- Does it choose its next step, or follow a fixed script? A fixed sequence is a workflow in Anthropic's terms. That is not worse; it is often what a plant should want.
- Can you see the log of a real run? Every tool call, input and result should be visible.
- How do you stop it? There should be a simple, tested way to pause or disable it without help from the vendor.
Our guide on how to choose an industrial AI vendor goes further on vendor evaluation.
When is each one the right choice in a plant?
Anthropic's advice is to find "the simplest solution possible", which "might mean not building agentic systems at all". That holds in industry.
Use a chatbot when people need answers from a large body of documents: procedures, manuals, HSE rules, technical specifications. Nothing is executed, so the risk is limited to the quality of the answer, which you control by restricting it to approved sources and showing the reference.
Use a copilot when a qualified person must remain accountable: work orders, permits, inspection reports, formulation proposals. The AI speeds up the draft; the signature stays human.
Use an agent for multi-step, repetitive work that crosses several systems and where each action is reversible or low impact. CISA's guidance notes that predictive ML models are typically used in operational layers, while LLMs are typically used in the business layers of the Purdue model, potentially on data exported from the OT network. That is a sensible starting zone for LLM agents too. For the integration side, see integrating AI with SCADA and OT systems.
What safeguards does an AI agent need before it touches operational systems?
These controls come from the OWASP entry on excessive agency, the CISA joint guidance on AI in OT and Anthropic's engineering notes:
- Least privilege and narrow tools. Give the agent only the functions it needs, with the minimum permissions, and avoid open-ended tools such as a shell.
- Authorization enforced by the target system. The maintenance system or historian decides what the agent may do, not the model.
- Human approval for high-impact actions. CISA advises limiting active control of OT infrastructure by AI without a human in the loop.
- Data pushed out, not pulled in. CISA recommends push-based or brokered architectures that move data out of OT without granting persistent inbound access, and one-way transfer where data must reach business networks.
- A distinct identity and full logs. Log inputs and outputs, and make sure the agent's identity in the logs is separate from users and other machines. Rate limits reduce the damage of a runaway loop.
- A failsafe and a bypass. Processes must be able to revert to traditional automation or manual operation, and the way to bypass or replace the AI belongs in functional safety and incident response procedures.
- Testing before production. Run the agent in a sandbox or on test infrastructure first, with stopping conditions such as a maximum number of iterations.
An agent is also a new component in the industrial control environment, so it falls under the same security program as the rest, such as the ISA/IEC 62443 series for securing industrial automation and control systems. Our agentic AI development and consulting teams can help define these limits before a pilot.
How does this map to KABAS AI projects?
Our two published projects follow the copilot pattern. At ADFC, a Sonatrach Group subsidiary, engineers define performance targets and constraints for each drilling section, Virtual Lab runs simulations and multi-criteria optimization, and the system produces a focused laboratory test playbook that the engineers validate in the lab. At ENAD Shymeca, researchers use Virtual Lab to test detergent formulations virtually before physical trials, and the project delivered a 40% reduction in physical testing cycles. In both cases the software proposes and the engineers decide. Both companies are exploring integration with DALILE, our multi-agent AI assistant, to strengthen knowledge discovery.
Sources
- Anthropic, Building effective agents (December 2024)
- CISA, ASD's ACSC and partners, Principles for the Secure Integration of Artificial Intelligence in Operational Technology (December 2025)
- OWASP GenAI Security Project, LLM06:2025 Excessive Agency
- IBM, What is a chatbot?
- Microsoft Learn, What is Microsoft Copilot?
- Microsoft Learn, Compare declarative and custom engine agents
- Staffing Industry Analysts, Gartner says agent washing is taking place (June 2025)
- ISA, ISA/IEC 62443 series of standards
Frequently asked questions
What is the main difference between a chatbot and an AI agent?
A chatbot answers questions in a conversation and does not change anything outside that conversation. An AI agent is given a goal, decides its own next steps and uses tools to act on other systems, such as creating a work order or sending a notification. The difference is about permissions and initiative, not about which language model is underneath.
Is a copilot just a chatbot inside an application?
Not quite. A copilot sits inside the tool where the work is done and uses the context of that work, such as the open document or record, to draft and suggest. The user still decides and executes every action. A chatbot is usually a separate conversation window with no access to the user's working context.
Can an AI agent control a PLC or a DCS directly?
It can be technically connected, but current guidance from CISA and its international partners advises limiting active control of OT by AI without a human in the loop. A safer pattern is read-only access to data pushed out of the OT network, with the agent proposing actions that an operator approves through the existing control and change management process.
How can I tell if a vendor's product is really an AI agent?
Ask who starts the work, which systems it can write to and under which account, whether it chooses its own next step or follows a fixed script, and how you stop it. Ask to see the log of a real run. If the user must always type first and nothing changes outside the chat, it is a chatbot or a copilot.
Which one should an industrial company start with?
Start with the simplest option that solves the problem. A chatbot over approved documents or a copilot that drafts for a qualified person covers many plant use cases with low risk. Move to an agent only for multi-step, repetitive work across several systems, and only once permissions, logging and a way to stop it are in place.



