Article by Nadine Soyez

Currently, I see a lot of hype and buzzwords about AI automation. Everything seems to be labelled as an AI workflow or, even more fancy, an AI agent. In my network and at my clients’ workshops, I experience that people are confused and overwhelmed from that noise. That’s why I’d like to provide some clarity and walk you through the different approaches so that you can understand what they mean for your business and where they truly add value.

 

AI chats: Reactive, dialogue-based, human-led

 

Systems such as ChatGPT, Claude or Gemini are frequently used as AI chats. These are reactive applications that respond to human input. The interaction follows a simple pattern: the user poses a question or enters a prompt, the AI responds, and the conversation ends. No further action is taken by the system.

The interaction remains entirely under human control. There is no process automation or integration with external tools. For this reason, AI chats are best suited to simple queries that require a high degree of human involvement and offer only a low level of automation.

 

Examples:

• You generate a text or image.
• You ask for an explanation of a concept.
• You let your meeting minutes be summarised.

 

 

Manually AI-assisted workflows: Support, but not autopilot

 

Manually supported AI workflows go a step further. Here, systems like ChatGPT, Claude or Gemini are embedded into working processes, offering assistance throughout tasks. You remain in control but receive targeted support from the AI. Although the level of autonomy remains low, the productivity gain is significant. This type of support is ideal for creative, text-based, or analytical tasks where human oversight is desired, but intelligent input is welcome.

 

Examples:

• You provide a structure → the AI helps you write the report.
• You upload meeting notes → the AI extracts action points.
• You begin writing an email → the AI suggests more polite wording.
• You write a paragraph and ask for alternatives → the AI gives you several improved versions.
• You provide a project structure → the AI helps draft the project plan.
• You describe a project risk → the AI suggests mitigation measures.
• You prepare a presentation → the AI generates an agenda or key messages.

 

 

Fully automated AI workflows: Standardised autopilot, reliable but inflexible

 

Fully automated AI workflows operate differently. They follow a predefined, rule-based process with minimal human intervention. A trigger initiates the workflow, which proceeds through a fixed sequence of steps with largely predictable outcomes. You set up each step explicitly defined in advance, including what actions the AI should perform. These workflows do not adapt to new circumstances or make independent decisions, making them ideal for structured, repeatable, and low-risk tasks.

It can include different large language models like (Chat)GPT, Claude, Gemini. There are no dynamic decisions: the workflow won’t decide at runtime which model to use. It follows your predefined logic and uses the models in specific steps with no autonomy. It is up to you to give the instruction of which model should be used when. There is no adaptation to output quality or errors.

 

Examples:

• IT support: Incoming tickets are automatically categorised and routed to the relevant team
• E-commerce: A customer places an order via chatbot; the AI guides them through the process and automatically sends a confirmation email
• Project management: Status updates are collected from tools and automatically turned into weekly reports
• HR: An employee submits a leave request, which is automatically approved and forwarded
• Marketing: Campaign data is compiled and sent as a PDF report to a predefined group
• Project management:  A project status report is generated every Friday and automatically sent to the project manager.
  If milestones are delayed, the system sends an automatic reminder to all responsible team members.

 

 

AI Agents: Autonomous problem-solvers with a goal

 

AI agents offer a high level of autonomy. Unlike fully automated AI workflows, they do not follow rigid sequences. Instead, they independently pursue defined goals. You specify the goal and the agent’s role; the AI then plans, executes, and makes decisions on its own.

AI agents can access external systems, search for information online, or update data files. Technically, this is supported by protocols such as the Model Context Protocol (MCP), which standardises tool and data access. These agents can plan and execute multistep processes, make their own decisions, and adapt to new situations.

Like the fully automated AI workflow, it can include different large language models in the task-solving process. But here lay the differences: the agent can choose the model based on context or performance. Different large language models may be fine-tuned for different tasks, e.g., one optimised for reasoning, another for summarisation, and another for coding. If one model fails or gives inconsistent results, the agent can fall back to another. Agents can also compare outputs from multiple large language models and choose the best one or synthesise a combined result. They treat the models as intelligent collaborators in the planning and execution loop.

 

Examples:

• IT Self-healing agent: Detects server issues, analyses logs, chooses an action (e.g. restart or escalation), and documents the incident
• Onboarding agent: Identifies a new employee, schedules onboarding sessions, sends checklists, notifies departments, and tracks responses.
• Sales autopilot agent: Scans the CRM for inactive leads, sends personalised reactivation emails, and automatically schedules follow-up meetings
• Marketing agent: Not only compiles campaign data but analyses performance and optimises the campaign autonomously
• Project Coordinator Agent detects delays, analyses dependencies, and proposes a revised timeline, including communication with stakeholders.
• Meeting Action Agent automatically scans all weekly meetings, extracts incomplete to-dos, and generates an updated task list for the team.
• Prioritisation Agent reviews ongoing tasks based on effort, deadlines, and resource availability, and dynamically reorders them.
• Resource Planning Agent identifies project bottlenecks, analyses workload distribution, and recommends capacity reallocation complete with justification.

One hint when you use Microsoft Copilot Studio to build your AI workflows and automation: For marketing reasons, Microsoft names every automation and workflow „AI agent“ even though it is not always an AI agent as described. This can be confusing.

 

 

Overview of Different AI Approaches Using a Project Status Report Example

 

AI Chatbot: You can ask an AI chatbot how to set up a project status report. The AI will generate an answer based on the knowledge in its trained model. However, it cannot access external or project-specific data. Its response is general and based only on its training.

Manually Assisted AI Workflow: With this approach, you can ask the AI not just how to set up the report, but also to generate it for your specific projecd. For this, you manually have to provide the necessary data and prompts to the AI to start the process. For example, you give the project data, and the AI writes the report. If you want to send it by email, you have to do that step yourself. The AI can help write the email, but not send it.

Fully Automated AI Workflow
This workflow can automatically access the necessary project data, generate the report, and send it via email without human interaction. You can set a trigger like “every Friday afternoon,” and the system will execute the full process.
This setup is structured and predictable: you define the steps, the data used is known, and the outcome (e.g., a standardised report) is consistent. Next actions, like adjusting the timeline or project plan if needed, are up to you.

AI Agent
An AI agent goes beyond automation. It not only generates and sends the report every Friday but can also analyse project data, identify issues, and proactively make decisions like adjusting timelines or modifying the project plan. The outcome is not strictly predictable: the agent makes autonomous decisions based on goals and context.

 

 

When should I use which approach?

 

Choosing the right AI solution depends heavily on the task, level of risk, and complexity involved. The more autonomy you give the AI, the less control you retain, which can increase risk but also boost scalability.

To decide which approach fits best, ask yourself these questions:

1. How complex is the task?

Is it simple, structured, and straightforward? Or does it involve multiple steps and complex decisions?
→ For straightforward tasks, an AI agent is unnecessary.
→ As the task becomes more complex, automation begins to make sense.

2. What is your risk tolerance?

How much control are you willing to hand over to AI?
→ For low-risk but complex tasks, AI agents can operate independently.
→ But for high-stakes scenarios (such as hiring, firing, or loan approvals), AI should assist, not decide. These tasks require a human-in-the-loop approach.

When you use AI agents, you must act in an ethical way.

 

Decision Guide

• For simple questions or conversation → use AI Chat
• For specific tasks requiring human guidance → use a Manual AI-assisted Workflow
• For repeatable, structured, predictable tasks that the AI can run itself → use a Fully Automated AI Workflow.
• For complex multi-step tasks with a clear goal, low risk, and ethical correctness, that the AI can run itself, use → an AI Agent.

 

 

Summary

 

Ultimately, it all depends on the nature of the task and the problem you are trying to solve. Not every use case needs an AI agent. Before building or deploying one, analyse the problem and evaluate the potential return on investment. In many cases, a simpler AI approach such as a chatbot or a workflow is not only sufficient but also faster, cheaper, and safer.

Start with the challenge, assess the complexity and risk, and then choose the most appropriate AI strategy for the job.