Concept

AI agent, assistant or chatbot: what is the difference?

A chatbot enables a conversation. An assistant supports a task. An AI agent can choose next steps and carry them out using permitted tools. The labels overlap: a chatbot may front an assistant that coordinates several agents. To choose a system, inspect the work it actually takes on, what it may change and when a person intervenes. [S1][S2][S4]

Last updated: 6 November 2026 | Next review: November 2026 Sources and illustrative MING method Machine-readable record

Which task should the system take on?

Start with the result your team needs. Ask to see how the system gets there and when a person takes over.

Three tasks for a supplier demo
Your taskWhat the demo must showClarify beforehand
Report the status of a delivery.The current status and the source used.Is looking up the status enough, or does someone need to investigate a problem?
Investigate a missing delivery.Which information it checked and how that led to the next step.What may the system trigger on its own? Who handles conflicting information?
Prepare a proposal from an order email.An order proposal matched to the customer and products, with references to the relevant details in the email.Who checks the proposal and decides how to process it further?

The first two tasks use our illustrative parcel example. The third draws on the STREIT customer story: Tabea, an AI agent at office supplier STREIT, prepares order proposals. Its exact permissions are not fully described in public. The table is our suggestion for your demo, not a complete description of this client system.[S4][S5]

Prepare a task for my supplier demo ↓

An illustrative example

The same chat can show a status or prepare an investigation.

“My parcel hasn’t arrived. Can you check?”

Show the tracking status

  1. Read the tracking record
  2. Show the latest status in chat
  3. A person takes over any further steps.

Investigate the missing delivery

  1. Check the order and choose what to inspect next
  2. Prepare an enquiry to the carrier
  3. Service team approves
  4. Open the case and show its reference.

Missing or conflicting information goes back to the service team.

Two chosen configurations, not a ranking of chatbot, assistant and agent. Opening an investigation needs human approval in this example. Proposed by MING Labs [S4].

A chat can show the parcel status or investigate a missing delivery

“My parcel hasn’t arrived. Can you check?” Imagine a customer typing this into a retailer’s chat. In one setup, the system retrieves the tracking status. In another, it investigates the missing delivery and prepares a carrier enquiry. Both have a chat window. This is our illustrative example, not a client case.[S4]

A chatbot can be the interface to an assistant or an agent

The chat window tells us how the customer talks to the system, but not who does the work behind it. An assistant might help a service employee answer the parcel question. An agent could choose which delivery record to inspect next, use a permitted tool and check what it found. Those steps can depend on the results. A fixed workflow can also call tools; tool access alone does not make it an agent.[S2][S4]

These descriptions overlap. SAP shows an assistant coordinating specialised agents. Microsoft describes Copilot Studio agents as AI-powered assistants and offers different ways of building them. The product name does not settle what a particular configuration can do.[S1][S3]

Our comparison of agentic and generative AI uses a delivery query to show which work each setup can take on.

Ask the supplier to handle the same missing parcel

In our example, the investigation system checks the order, then chooses which permitted delivery information to inspect. It might find a tracking gap or conflicting records. It prepares an enquiry only when the evidence supports one. A service employee must approve it before the system opens a carrier case. Missing or contradictory information goes back to that employee.[S4]

For a useful demo, ask the supplier to show which records the system checked, which finding led to the next action, what it changed and who approved it. Repeat the task with a wrong order number and missing tracking data. A fluent reply alone does not show that the underlying task was completed.[S4]

Bring the same parcel question to your next supplier demo

Before the demo, write down what your team needs from this parcel question: a current status, or an investigation ready for approval. Give suppliers the same task and compare what their systems actually do. The narrower setup may be enough if customers only need the status. If exceptions need investigation, our guide to choosing between an agent and a fixed workflow helps you examine those decisions. When an agent joins the team, the role-introduction guide helps you name its owner and limits. These are proposed checks, not measured savings or proof of superior performance.[S4]

For concrete tasks, our business examples examine three published company and vendor accounts, separating reported work from the conditions and permissions you would still need to check. They are not an independent product comparison.

Bring a task you can check to the demo

Adapt the template to your process and give every supplier the same conditions. Check what happens when information is missing or contradictory.[S4]

View and copy the template
Save as a text file

Task for an AI supplier demo

Our task:
[A recurring process in our team.]

The result we need to work with:
[What must be ready at the end? Which source or supporting record should accompany it?]

The documents and systems the application may use:
[Use only information approved for the demo or anonymised information.]

The actions the application may take on its own:
[For example, read information and save a draft. Specify permissions for messages, changes and approvals separately.]

When a person must take over or approve an action:
[Name or role, the reason for the handover and the information that person needs.]

Please demonstrate the same task three times:
1. With complete, consistent information.
2. With one piece of information missing.
3. With two conflicting pieces of information.

Please show:
- The sources used and the passages supporting each detail.
- Whether the next steps are predefined or chosen based on what the system finds.
- What the system actually changed or triggered.
- When it stopped and who received the unresolved case.

Our assessment:
Compare the information shown with the approved original documents yourself. A source reference alone does not prove that the information was used correctly.
[What was usable? What did we need to check or correct? Which question remains unanswered?]

MING Labs: a proposed evaluation aid, not a product certification or a promise of results.

Compare AI offers using a task from your own team

Bring one task and the offers you are considering. We can discuss the result the system should deliver, what it may change and how your colleagues will check the work.

Discuss my AI task

Tell us what you have in mind. We will respond personally.

Sources

[S1]
KI-Agenten vs. KI-Assistenten SAP · 2026-07-24 Supports: An assistant can coordinate specialised agents
[S2]
Building effective agents Anthropic · 2024-12-19 Supports: Predefined workflows, dynamically chosen steps and human checkpoints
[S3]
Agents overview Microsoft Learn · 2026-10-01 Supports: Different operating approaches under overlapping product labels
[S4]
Illustrative parcel example and proposed demo checks MING Labs · 2026-11-06 Supports: Example, permissions and handovers; no measured client outcomes, Task selection and copyable demo assignment as a proposed evaluation method
[S5]
STREIT: an AI agent in customer service MING Labs · 2026-11-06 Supports: Tabea reads order emails, matches customers and products and prepares order proposals; exact permissions not fully published Public customer story checked on 6 November 2026

Frequently asked questions

Is a Copilot automatically an AI agent?
The name is not enough. Inspect the particular product and configuration: does it choose next steps, follow fixed paths or only answer questions? Microsoft documents different approaches within Copilot Studio. [S3]
Can people use an AI agent through a chatbot?
Yes. The chat window is the interface; the system behind it can use tools and agents. The conversational format alone reveals neither its permissions nor its freedom to choose actions. [S1][S2]
Can an agent still require human approval?
Yes. Human checkpoints are compatible with agentic systems. In our example, the system chooses its investigation steps but needs approval before opening a carrier case. [S2][S4]
All Insights