If a proposal promises “an AI agent to manage appointments”, I would ask to see an operation from start to finish. Someone requests a slot, the system checks the correct calendar, and the booking appears where you will handle it. A convincing chat response does not prove that an appointment exists.
A rule-based workflow may be enough for that task. For example, it asks for the service type, offers permitted slots and confirms the selection. Adding a language model can help interpret a freely written request. What matters is whether the whole system does the agreed work and keeps within its limits.
The difference is who decides the steps
Anthropic distinguishes between workflows where code determines the paths and agents where the model dynamically directs the process and tool use. Its explanation acknowledges that there are different definitions of “agent”. That distinction helps you ask about architecture; it is not grounds to dismiss every system using if/else conditions.
Access to a database is not enough to call a system autonomous either. A predefined workflow can check a calendar and make a booking. And an agent does not need to store conversations permanently if the task does not require it. Memory, tools and autonomy are separate decisions.
I would compare proposals using these questions:
| Question | What I would ask to see |
|---|---|
| What information does it consult? | The specific source, permissions and date or status of the data. |
| What actions can it perform? | A list of permitted changes and those requiring approval. |
| How does it confirm the result? | The destination system’s response and the operation record. |
| What does it retain from a conversation? | Stored data, purpose, access and duration, if anything is persisted. |
| What happens when it fails? | An error case and the route to a person or a review queue. |
A test that includes an exception
In a hypothetical garage scenario, someone asks for their brakes to be checked on Monday at 18:30 because they have Spain’s vehicle roadworthiness inspection, the ITV, on Tuesday. Confirming the work requires its duration, availability and the garage’s conditions. If that information is missing, the right response is to ask for clarification or pass the enquiry on. I would not accept an answer merely because it mentions all three.
A workflow with well-prepared rules can pass the same test. An agent that checks a calendar but misinterprets the duration can fail it. I would also test an inaccessible calendar and a slot that is taken between the availability check and confirmation. You need to observe the outcome and the message the customer receives.
What I would decide before granting more autonomy
If the steps are clear and exceptions are few, I would start with a limited workflow. If each request requires choosing between different paths using changing information, I would examine a more flexible solution. First, I would put the permissions, approvals and stopping point in writing. Anthropic’s documentation includes stopping controls and human intervention; they are not a guarantee of accuracy.
A rule-based system can also have programming errors, outdated data or duplicate actions. I would not present it as infallible. I would ask the quote to include testing, failure review, usage costs and maintenance. The commercial name matters less than knowing which operation you can entrust to it and how to regain control.
Sources
- Anthropic: Building effective agents — Published on 19-12-2024; accessed on 10-10-2026. Architectural distinction and execution controls.
If you have a specific operation you want to assess, you can describe it to me through Let’s talk.