Automation
Receives events, passes data, and checks predefined rules. For example, it verifies mandatory fields or hands off an approved draft for delivery.
AI workflow automation with human review
BotCentralHub helps design workflows that connect business tools, AI processing, and human review. Using an illustrative inquiry model, we show what triggers the automation, what data it requires, and what happens when an error occurs.
What is passed between individual steps
For each step, specify the input, expected output, and continuation condition. Missing documentation, rejected proposals, and service failures each require their own procedure.
AI
Processes unstructured text and prepares drafts. In our example, it extracts inquiry details and formulates follow-up questions.
Human
Reviews the meaning, recipient, and commercial terms. Decides whether a specific response may be sent.
Model workflow: from inquiry to approved response
Model proposal for a small web agency. Not a presentation of a deployed BCH integration.
A new inquiry arrives via a web form. The goal is to prepare a response and send it only after review by a responsible team member. The design assumes that both inquiry receipt and message delivery have accessible technical interfaces.
-
Inquiry receipt
A new record triggers the workflow. The system logs the inquiry identifier and its status so that repeated delivery of the same event is not treated as a new order.
-
Input verification and preparation
The system verifies the presence of the message and usable contact information. It prepares only the information required to process the request for the AI. Contact details for delivery are kept separate from the materials provided to the model.
-
AI processing
The AI drafts a summary of requirements and flags missing information. It does not fabricate prices or deadlines. It treats customer text strictly as source material, not as authorization to alter workflow rules.
-
Structural check and draft preparation
The application validates whether the output contains the required fields and correct data types. It prepares a draft response from the usable materials. Format validation alone does not confirm factual correctness.
-
Human approval
A team member reviews the original inquiry, draft, and recipient. They may edit, approve, or reject the response. Approval applies to a specific text version and recipient; any subsequent modification requires fresh approval.
-
Delivery and outcome logging
The delivery step may only use an approved version. The system logs the outcome of handing off to the dispatch service. Service acceptance does not guarantee delivery to the recipient's inbox.
What is passed forward
Inquiry identifier, confirmed requirements, missing details, response draft, and processing status. Delivery additionally requires the recipient, approved version, and approval record. Only relevant system components and authorized personnel have access to this data.
What happens when the process cannot continue
Incomplete or ambiguous input
Without contact details or usable requirements, the case is passed to a team member. A missing budget or deadline can be included among follow-up questions. The system must not infer them.
Invalid output or service unavailability
Invalid output does not proceed to delivery. In this model proposal, we allow at most one retry for AI processing. Upon further failure, a human takes over the case.
Rejection or missing approval
A rejected response is not sent. If approval is not received within the set timeframe, the case remains unsent and requires manual resolution. Timeout does not constitute consent.
Repeated event or unclear delivery status
The system checks existing records and dispatch status. When the delivery service does not return a definitive response, the message cannot be automatically resent without review. In this example, an unclear status is handled by a team member.
How to translate the design into tools
In n8n, a Wait node can pause execution until a subsequent event is received. An Error Trigger enables running dedicated error handling. These mechanisms are building blocks; approver permissions, binding to a specific response version, and duplicate delivery protection must be designed and verified across the application.
Technical reference & documentation:
What to verify before launch
- Each step has a defined input, output, and responsible person.
- Only data required for the task is sent to the AI.
- Approval is enforced by the application and tied to the correct version and recipient.
- Repeated inputs do not trigger unintended duplicate actions.
- Errors, rejections, and timeouts each have a defined procedure.
- Testing covers standard inquiries, incomplete inputs, rejections, and ambiguous delivery results.
Start in a test environment with mock data. Document what testing demonstrated and which edge cases remain to be addressed.
Start with a single repeating process
Prepare a brief description of your current procedure, an anonymized sample input, desired outputs, and current tools. Include points where errors currently occur or where human decisions are required.