15,000+
source conversations
Real conversations informed the structure of the knowledge base.
Anonymised case / customer support
A first-line AI agent for multi-channel support.
The team turned manual handling of repeated enquiries into a controlled process with AI responses, operator escalation and a separate control panel.
Discuss your processRecorded outcome
Figures from the internal completion material.
90%
automated handling
The share attributed to the agent without operator involvement in the internal completion report.
under 1 min
first response
Before implementation, a response could take tens of minutes or wait until the next working day.
3x lower
operator workload
Operators focused on complex questions and parallel conversations.
Starting situation
What required a change.
Enquiries from multiple channels were handled manually during working hours.
Repeated questions consumed operator time on every occurrence.
Evening and night enquiries waited for a response and created funnel leakage.
Operators lacked one control surface for active conversations.
Solution architecture
A controlled process, not a single tool.
Orchestration
n8n manages messages, conversation states, timeouts and handover between the agent and operator.
AI agent
A specialised prompt and vector knowledge base keep answers within the agreed product scope.
Channels
Meta integrations receive enquiries and return answers to the original conversation.
Human in the loop
A trigger or complex request hands the conversation to an operator with its context.
Operator panel
A Telegram panel shows active conversations, queues and claimed requests, and allows the bot to be disabled.
Return to automation
After inactivity, the workflow can return the conversation to the agent under a defined rule.
Control and safety
A person remains at the critical decision point.
Escalation
A user can request a person, while the system also identifies complex requests.
Context
Conversation history is passed to the operator without asking the customer to repeat everything.
Control
An operator sees every status and can manually change the agent mode.
Evidence status
What the supplied package supports.
Evidence source: an internal project completion deck reviewed on 29 July 2026. It records 90% automated handling, 10% escalation, response time under one minute and a threefold workload reduction. Raw conversation logs and the calculation method were not included, so these values are presented as internal project results, not an independently verified benchmark.
Claim boundary
What is intentionally excluded.
The client name, product category, domains, access details, prompts and conversation content are not published. Claims of unlimited scale or increased conversion are excluded because the supplied material contains neither a load test nor a numeric conversion measure.
Your process
First capture the baseline. Then choose the tool.
The first conversation defines the task, data, cost of error and human control boundary.
Describe the task