AI proposes the intent, business context, missing information and a possible automation opportunity. It can propose a follow-up question. It does not assign the Hot/Warm/Cold band.
← Back to the homepageWORKFLOW CASE STUDY 01 / SOYLU SISTEM
From a business enquiry to an inspectable decision.
A lead qualification workflow built and tested by Soylu Sistem using a real API and controlled synthetic data. The recorded results show system behaviour in defined test scenarios; they are not client performance results.
Before: the same manual work for every enquiry.
The test scenario assumes a sales coordinator reads about 40 enquiries each weekday, copies details into a spreadsheet, asks follow-up questions and decides who should receive a discovery call.
RECORDED TEST · SYNTHETIC DATAFrom enquiry to review recommendation
“We receive about 40 enquiries every weekday. A coordinator reads each one, copies details into a spreadsheet and decides who gets a discovery call. We want consistent qualification and faster routing. We are ready to act within 30 days.”
Condensed from the synthetic input used in an actual API test. Step through the model proposal and the rule result.
- Intent
- AI workflow automation
- Current tool
- Spreadsheet
- Volume
- About 40 enquiries / weekday
- Timeline / readiness
- 30 days / ready
Proposed opportunity: structure enquiry details, standardize follow-up questions and prepare strong opportunities for human review.
A model proposal, not a score or a business commitment.
Priority under the demo rules97 / 100
High priorityFit40/40
Clarity20/20
Urgency12/15
Potential value15/15
Readiness10/10
Next step: human review for discovery-call readiness. Not a purchase probability or a client outcome.
One recorded run, not a live monitor. No lead was stored or handed off.
What we built.
A visitor describes a need in natural language. The server checks the input, requests a structured proposal from the model and validates it. When required information is missing, the interface asks a question. Once the fields are complete, backend rules calculate a priority and a review recommendation.
- Natural-language business enquiry
- Input checks and an obvious email/phone guard
- Server-side AI extraction
- Output schema and field validation
- Intent guard and missing-field recomputation
- Rule-based score for automation enquiries
- An on-screen review recommendation, not an automatic handoff
An AI proposal is not a system decision.
The backend validates the output fields, recomputes what is missing and blocks automation scoring for non-automation intents. Eligible enquiries then pass through the demo scoring rules.
What the test actually showed.
In the recorded run, the required fields were complete. The backend produced 97/100 and a Hot priority, then recommended human review for discovery-call readiness. The result was not persisted or automatically sent to anyone.
The same validated facts produce the same rule score. Different model wording can affect field lengths and the resulting score. This is an initial example rubric, not a calibrated sales or purchase probability.
Inspect the recorded test response
Original API response with synthetic test data. Model and token usage describe this recorded run only.
{
"mode": "synthetic-demo",
"persisted": false,
"aiProposal": {
"intent": "ai-automation",
"extractedFields": {
"business": null,
"workflowArea": "lead-sales",
"problem": "Manual enquiry qualification and routing are slow, and the goal is more consistent qualification and faster routing of strong opportunities.",
"currentProcess": "A sales coordinator reads each enquiry, copies details into a spreadsheet, asks follow-up questions, and decides who should receive a discovery call.",
"existingTools": "Spreadsheet",
"volume": "About 40 inbound enquiries every weekday",
"desiredOutcome": "Consistent qualification and faster routing of strong opportunities",
"frequency": "daily",
"timeline": "30-days",
"readiness": "ready",
"constraints": null
},
"modelMissingFields": [],
"automationOpportunity": "Automate enquiry detail capture and qualification, standardize follow-up questions, and route strong opportunities for discovery calls.",
"confidence": 98,
"proposedNextQuestion": ""
},
"backendDecision": {
"status": "qualified",
"knownFacts": {
"problem": "Manual enquiry qualification and routing are slow, and the goal is more consistent qualification and faster routing of strong opportunities.",
"currentProcess": "A sales coordinator reads each enquiry, copies details into a spreadsheet, asks follow-up questions, and decides who should receive a discovery call.",
"existingTools": "Spreadsheet",
"volume": "About 40 inbound enquiries every weekday",
"desiredOutcome": "Consistent qualification and faster routing of strong opportunities",
"workflowArea": "lead-sales",
"frequency": "daily",
"timeline": "30-days",
"readiness": "ready"
},
"missingFields": [],
"qualification": {
"dimensions": {
"fit": 40,
"clarity": 20,
"urgency": 12,
"potentialValue": 15,
"readiness": 10
},
"total": 97,
"band": "Hot",
"nextAction": "Human review for discovery-call readiness."
},
"nextQuestion": {
"field": null,
"question": null,
"source": "complete"
}
},
"telemetry": {
"model": "gpt-6-luna",
"usage": {
"inputTokens": 580,
"outputTokens": 269,
"totalTokens": 849
}
}
}What happens when the normal path fails?
Ask a targeted next question instead of producing a premature score.
Do not generate an automation score; recommend human review.
Reject the input before the model call.
Show the failure without pretending that a new qualification succeeded.
Scope and safeguards.
- Use synthetic/test data only. The guard does not detect names or every possible type of personal information.
- The email/phone guard is a limited check, not a replacement for an approved real-personal-data workflow.
- Conversation state lives in page memory and is cleared on refresh. There is no durable lead record, CRM or automatic follow-up.
- Human review is a recommendation. No meeting is booked and no email or WhatsApp message is sent.
- The model’s confidence value is a self-estimate, not measured extraction accuracy.
Apply the pattern to a different workflow.
Start by describing one recurring task and the tools involved. We can then separate what needs rules, what benefits from AI and what should remain a human decision.