Standalone prompt
A single defined request and output for human review.
Example: Extract requested website pages and missing details from a short inquiry.
AI agents and choosing the right workflow
BotCentralHub explains how to choose an AI approach for a specific task. Compare four approaches, go through the decision framework, and determine where a human should verify the result.
What we mean by an AI agent here
In this guide, an AI agent refers to a system where the model selects the next steps and uses available tools based on the goal and intermediate results. In a predefined workflow, the sequence is determined by prebuilt logic. A standard chat interface does not distinguish between these approaches.
Reference material: Anthropic: Building effective agents ↗
Four approaches to working with AI
These are building blocks that can be combined depending on the task.
Predefined workflow
Predetermined steps connecting processing and review stages.
Example: Create a summary from an inquiry, check for completeness, and draft a response.
Routing
The input is assigned to a category and passed to the corresponding procedure.
Example: Distinguish between a new inquiry, a technical issue, and a site update. Escalate ambiguous messages to a human.
Tool-using agent
The model selects the next action based on prior results and granted permissions.
Example: Find the root cause of an error in a test project, propose a fix, and verify the outcome.
How to choose an initial approach
-
Can the task be precisely described by rules?
Consider a standard script or non-AI automation. -
Is processing one input into one output sufficient?
Start with a standalone prompt. -
Are the subsequent steps known in advance?
Design a predefined workflow. -
Do different types of inputs require distinct procedures?
Add routing and an escalation path for ambiguous cases. -
Must the next step change based on new findings?
Consider an agent with restricted tools and stopping conditions.
Note: These questions serve as a starting point. Approaches can be combined. Validate suitability using representative examples from your own task.
Illustrative example: website inquiry processing
Model proposal — not a presentation of a deployed BCH system.
Situation
A small web agency receives messages with new inquiries, feature requests, and issue reports. The goal is to prepare background materials for a response.
Proposed workflow
- Extract verified details from the message and flag what is missing.
- Categorize the message as a new inquiry, a website modification, or a technical issue. Pass mixed or ambiguous messages to a human.
- Prepare the relevant summary or follow-up questions based on the category.
- Have a human review the content, add commercial terms, and decide on sending.
Why start with a routed workflow here
Both categories and subsequent steps are known in advance. We therefore choose a predefined workflow with routing for this proposal. The model prepares draft materials; in this example, it has no authority to set prices, promise deadlines, or send messages.
What to verify during testing
- Did the system preserve all data provided in the message?
- Did it recognize missing information?
- Did it escalate mixed or ambiguous messages to a human?
- Did it avoid generating unbacked prices or deadlines?
- Did sending remain under human control?
What to define before granting tool access
- Which data the system may access and modify.
- Which actions require human approval.
- When to stop execution or escalate a problem to a human.
- How to cap time, step counts, and costs.
- How to log executed actions and review results.
Permissions must be enforced by application logic and tool settings. Instructions in a prompt alone are not a replacement.
Reference material: OpenAI: A practical guide to building agents — PDF, English ↗
Continue to specific implementation
In Automation, you will find principles for connecting individual steps to applications. The Lab presents the Neutral Personal Agent concept and questions of decision-making control.