Built an inquiry management system that screens website submissions, identifies legitimate client enquiries, classifies them with AI, and routes them to the right person while automatically handling unwanted submissions.
The challenge
A consulting business receives a mix of genuine client inquiries, vendor pitches, promotional messages, spam and other irrelevant submissions through its website. Manually reviewing every submission creates repetitive work for the team and increases the chance of a legitimate inquiry being overlooked. The goal was to automate the first level of inquiry screening while making sure genuine potential clients still reached the right person.
The solution
Built an automated inquiry screening and routing system that evaluates each website submission, determines whether it is a legitimate client inquiry and routes it accordingly. The workflow connects the website inquiry form to an AI screening process then uses the result to determine what happens next.
Workflow diagram
System architecture
What I built
Built in Lovable with a Tally inquiry form integrated directly into the site.
Creates a professional client-facing intake point for all inquiries.
Submissions are analysed by Make AI Toolkit and GPT against defined criteria to determine whether each submission represents a legitimate client inquiry or an unwanted submission.
Removes manual screening entirely.
The AI classification result determines which path the submission takes. Legitimate inquiries go to the team. Unwanted submissions are handled automatically.
No manual sorting required.
Legitimate inquiries trigger a professionally formatted email containing the full inquiry details and the AI assessment.
The right person receives the right information immediately.
Rejected submissions are removed from Tally automatically through an API request.
Unwanted submissions never become a manual task for the team.
The workflow includes error handling to manage failures gracefully and retry logic to ensure reliable execution.
Built to handle real-world conditions not just ideal scenarios.
Automation engine
Operational impact
Measurable results
Manual inquiry screening eliminated through AI classification
Speed at which legitimate inquiries reach the right person
Risk of genuine inquiries being missed during busy periods
Screenshots
The Northstar Operations website and client inquiry experience that feeds the automated screening workflow.
Purpose · Shows the professional front-end entry point for new operations inquiries.
The full scenario from the Tally trigger through AI screening, the router and both the Gmail and API deletion branches.
Purpose · Shows the complete screening and routing engine in one view.
A completed scenario run where the AI classification routed an unwanted submission to the Tally deletion request.
Purpose · Shows the rejected-submission branch executing successfully.
A completed scenario run where a legitimate inquiry passed the filter and triggered the department notification email.
Purpose · Shows the legitimate-inquiry branch executing successfully.
The formatted notification email delivered to the Head of Department with full inquiry details and the AI assessment.
Purpose · Shows the output a legitimate inquiry produces.
Key outcomes
A manual first-level screening process where every website submission — genuine inquiry, vendor pitch, promotion or spam — had to be read by a person before anything could happen.
A Make.com scenario triggered by Tally submissions that classifies each inquiry with Make AI Toolkit and GPT, then routes legitimate inquiries to Gmail and deletes unwanted submissions through the Tally API, with error handling and retry logic.
During testing a routing issue was discovered where legitimate inquiries were being sent down the wrong path. By inspecting the execution history and tracing the data structure the filter problem was identified, corrected and retested successfully. This reinforced the importance of validating data structure between workflow steps not just the logic on paper. A workflow can look correct on paper and still fail because the data moving between steps is not structured the way you expect.
All data shown in screenshots is sample data used for demonstration purposes.