Deployment Reports
Operational transformations. Documented.
Three production deployments. Each structured as an operational transformation report — problem definition, architecture, orchestration design, and outcome metrics.
Inbound lead processing and pipeline qualification migrated from agent-initiated activity to a structured orchestration system.
A residential property agency processing 340+ inbound enquiries per month was operating without a defined intake structure. Enquiries arrived across multiple channels and were processed by individual agents on an ad-hoc basis. No qualification framework existed. Follow-up sequences were not executed systematically. Pipeline state was not measurable at any point in time. A workflow orchestration system was deployed across a 14-week period. The system now handles signal capture, classification, qualification sequencing, lead scoring, follow-up execution, and pipeline reporting as background processes. Agent involvement is triggered only at defined escalation conditions.
Operational Problem
The agency had no structured intake layer. Enquiries from property portals, web forms, and direct WhatsApp contact entered the same shared inbox and were processed in whatever order an agent happened to open them. No classification logic existed to separate high-intent signals from general enquiries. Qualification was conducted through unstructured WhatsApp conversations initiated at agent discretion — buyer budget, timeline, and location data were captured inconsistently or not at all. The absence of structured qualification data meant lead prioritisation was not possible. Pipeline state could not be reported because the underlying data did not exist in any structured form.
Existing Workflow — Before State
Inbound signal arrives on portal or website — system sends email notification to shared team inbox
Agent opens inbox at an undetermined time — no defined response window, no assignment logic
Agent initiates WhatsApp contact from personal device — no template, no CRM log, no tracking record
Qualification conducted through natural conversation — data not extracted to any structured record
If site visit agreed, agent books manually via phone call — no calendar integration, no confirmation sent automatically
Lead record updated in CRM at agent's discretion — entries incomplete, timestamps frequently absent
No automated follow-up for leads that do not respond — silence treated as disqualification by default
No mechanism for the principal to view pipeline state at any point in time
Operational Bottlenecks
A 6.2-hour mean response time meant most first contacts occurred after the initial enquiry window had closed. Leads in this context typically contact multiple agencies simultaneously; a delayed response reduces the probability of securing the conversation before a competitor does.
77% of inbound leads received no structured follow-up. Leads that did not respond to initial contact were typically abandoned. No re-engagement mechanism existed.
Without structured qualification data, all leads occupied equal pipeline priority regardless of buyer readiness. Agent time was allocated without objective basis.
The principal had no mechanism to view lead volume, pipeline stage distribution, conversion rates, or team activity. Operational decisions were made without supporting data.
AI Infrastructure Architecture
Deployed Pipeline
Signal Ingestion Layer
Intent Classification Engine
Qualification Flow Orchestrator
Lead Scoring Engine
Sequence Execution Layer
Escalation Router
Observability & Reporting Layer
Seven components handle distinct stages of the lead lifecycle. No component requires direct agent initiation. The agent team interacts with the system only when the Escalation Router delivers a briefing requiring human judgement.
Signal Ingestion Layer
Captures and normalises inbound signals from all source channels into a unified event stream
Intent Classification Engine
Assigns each signal a routing decision: respond immediately, enter qualification, or archive as non-convertible
Qualification Flow Orchestrator
Conducts multi-turn conversations to extract budget range, purchase timeline, and location criteria as structured data
Lead Scoring Engine
Calculates a score from extracted qualification data against a configurable rubric. Triggers escalation at threshold breach.
Sequence Execution Layer
Executes follow-up sequences for non-responding leads across a configured 14-day window. Timing and content adjustable without code changes.
Escalation Router
Assembles and delivers a structured briefing to the assigned agent when escalation conditions are met
Observability & Reporting Layer
Logs all events to a structured store and delivers scheduled pipeline briefs to the principal
Workflow Orchestration Design
Each inbound signal enters the Signal Ingestion Layer within seconds of receipt and is passed to the Intent Classification Engine. The engine assigns one of three routing outcomes: immediate response, qualification sequence entry, or non-convertible archive. Leads entering the qualification sequence are handled by the Qualification Flow Orchestrator, which conducts a multi-turn conversation to extract budget range, purchase timeline, and location criteria. Extracted data is written to the lead record in structured form and passed to the Lead Scoring Engine. A score is calculated against a configurable rubric. Leads scoring above the configured threshold (currently 82) are passed to the Escalation Router. Leads below threshold enter the Sequence Execution Layer, which executes follow-up steps across a 14-day window at configured intervals. If no response is received within the full window, the lead is archived with a complete activity log. Sequence timing and message content are adjustable in configuration without code changes.
Conversational Intelligence Layer
The conversational layer uses GPT-4o with a system prompt that encodes qualification objectives, response constraints, and agency-specific context. Prior message history is retrieved for each conversation turn — the model does not operate on single-message input. The system presents under a named agent identity representing the agency; it does not identify itself as automated. Response generation takes 30–60 seconds from signal receipt under normal operating conditions. Concurrent conversations do not compete for processing capacity — throughput scales with inbound volume. Every message sent and received is logged against the lead record with timestamp and message content.
Human Escalation System
Escalation conditions are defined in the deployment configuration and can be modified without code changes. Default conditions are: lead score above 82; stated budget at or above the configured threshold; explicit request for human contact detected in conversation; a compliance-relevant phrase detected. When any condition is met, automated messaging pauses immediately. The Escalation Router assembles a briefing containing the full conversation transcript, extracted qualification data, current lead score, and a plain-language description of the escalation reason. The briefing is delivered to the assigned agent via WhatsApp within 30 seconds of trigger. If no agent activity is detected within 4 hours, the system logs the inactivity and, where configured, re-notifies. Automated sequencing does not resume during an open escalation.
Monitoring & Analytics Layer
$ workflow-monitor --env=prod --study=VC-CS-2401
[SIG] 2024-09-12 09:41:03 inbound_signal received source=whatsapp id=MSG-8821
[CLS] 2024-09-12 09:41:03 classification=residential_enquiry confidence=0.94
[QFL] 2024-09-12 09:41:04 budget_range=detected timeline=short location=confirmed
[SCR] 2024-09-12 09:41:04 score=78 threshold=65 → qualifying
[SEQ] 2024-09-12 09:41:05 acknowledgment_sent follow_up_scheduled=T+4h
[OBS] 2024-09-12 09:41:05 log_written escalation=false
[SIG] 2024-09-12 09:43:17 inbound_signal received source=whatsapp id=MSG-8822
[CLS] 2024-09-12 09:43:17 classification=general_query confidence=0.89
[QFL] 2024-09-12 09:43:18 insufficient_signals → response_only_mode
[SEQ] 2024-09-12 09:43:18 template_response_sent re-engagement=T+24h
── status ─────────────────────────────────────────────
signals_today : 34 mean_response : 44s
escalations : 1 queue_depth : 0
Every system event is written to a structured log entry in MongoDB Atlas: signal receipt, classification decision, each conversation message sent and received, sequence step execution, escalation trigger, agent response, and lead outcome. The log schema is consistent across all lead records and is queryable without custom tooling. A state manifest is maintained in real time: count of leads by status (active, in-sequence, escalated, archived), mean time in each state, escalation queue depth, and sequence completion rate by step. Alerts are generated when response latency exceeds 120 seconds, when a sequence delivery fails, or when the classification engine returns a confidence score below the configured minimum.
Dashboard Visibility
Two scheduled briefings are delivered automatically: a daily brief at 09:00 containing pipeline state, overnight signal volume, and open escalations; a weekly summary on Friday at 18:00 containing conversion rates by lead source, sequence completion rates, and lead scoring distribution. An operational view is available in real time, showing current queue depth by status, escalation queue, and scoring distribution. All underlying data is exportable in structured format. Data ownership remains with the agency; the system operator does not retain data beyond the deployment term.
Operational Outcome Metrics
Mean signal response time
Structured follow-up rate
Qualified leads per month
Site visits booked per month
Agent time on manual follow-up
Pipeline reporting
Infrastructure Evolution Opportunities
Property matching layer — serving listings to qualified leads based on extracted criteria, without agent curation
Multi-branch configuration — deploying the same workflow system across branch locations with a consolidated reporting view
Scoring weight adjustment — updating scoring rubric weights using historical conversion data as it accumulates
Channel expansion — extending signal capture to email and SMS enquiry sources beyond WhatsApp
Long-Term Optimisation Roadmap
Scoring weight adjustment using 90-day conversion outcome data; initial calibration of rubric against observed lead quality
Multi-branch deployment with per-branch queue management and a consolidated principal dashboard
Property matching layer: serve extracted buyer criteria against listing inventory to generate personalised follow-up content
Start a deployment
Begin with an operational audit. No commitment required.
We map your current workflow structure, identify where the operational layer is absent, and calculate the measurable impact of closing that gap before proposing a build.
Request an operational audit