Agentic-CS

Inspiration

Customer service is often the first point of contact between a business and its customers, but many small and medium-sized businesses still depend on fragmented tools, repetitive manual responses, and conventional chatbots that cannot safely complete real tasks.

We were inspired to build Agentic-CS after noticing that most AI customer-service solutions focus only on answering questions. They often lack reliable business context, operational integrations, human approval controls, and a clear audit trail. For businesses handling orders, payments, refunds, support tickets, and sensitive customer information, generating a convincing answer is not enough—the system must also be accurate, controllable, and accountable.

Agentic-CS was created to explore a different model: an AI customer-service operating system that can understand a customer’s request, retrieve approved evidence, interact with business systems, and involve human staff whenever an action carries risk.

What it does

Agentic-CS is a multilingual, guardrail-first AI customer-service platform designed for Malaysian SMEs.

It allows businesses to connect their approved knowledge, customer records, orders, payments, and support workflows to an AI agent. The agent can:

• Answer customer questions using verified business knowledge
• Provide citations and supporting evidence with its responses
• Retrieve relevant customer, order, payment, and ticket information
• Identify the customer’s intent and determine the appropriate workflow
• Escalate uncertain or sensitive cases to human support
• Request human approval before performing high-risk actions
• Maintain a replayable audit trail of AI decisions, evidence, tool calls, and approvals
• Support multilingual customer conversations
• Help businesses prototype digital customer experiences through an AI-assisted Design Studio

Instead of operating as an isolated chatbot, Agentic-CS connects customer conversations with actual service operations while allowing businesses to retain control over important decisions.

How we built it

We designed Agentic-CS around a modular agent architecture consisting of several major layers.

The conversational layer handles customer messages, multilingual interaction, session context, and intent recognition. The knowledge layer uses retrieval-augmented generation to search approved business documents and provide evidence-based responses with citations.

The agent runtime coordinates reasoning, knowledge retrieval, business-system tools, approval requests, and support handoffs. Integrations expose structured information such as customer profiles, orders, payments, and tickets without allowing the language model unrestricted access to backend systems.

A policy and guardrail layer determines which actions can be completed automatically, which require human approval, and which must be rejected or escalated. Every important step is recorded by the audit layer, allowing administrators to replay an interaction and inspect the evidence, decisions, tool calls, and human interventions involved.

We also built an administration interface for managing knowledge sources, monitoring conversations, reviewing approval requests, configuring policies, and auditing AI runs.

Challenges we ran into

One of the largest challenges was balancing agent autonomy with operational safety. A useful customer-service agent needs access to business data and tools, but unrestricted access can create serious risks. We therefore had to design clear boundaries between read-only operations, low-risk automated actions, approval-required actions, and prohibited actions.

Another challenge was preventing hallucinated or unsupported answers. Retrieval alone does not guarantee reliability, so we needed to connect each response to approved evidence, define minimum evidence requirements, and escalate cases when the available information was insufficient.

Maintaining context across longer customer conversations was also difficult. The agent needed to remember relevant details without allowing outdated or unrelated information to influence later decisions.

Integrating multiple business domains—customers, orders, payments, tickets, knowledge, approvals, and audit records—required a consistent tool interface and shared data model. We also had to make the internal agent process understandable to non-technical business operators through a clear administration and replay interface.

Finally, multilingual customer service introduced challenges involving translation quality, mixed-language messages, business terminology, and maintaining the same safety policies across different languages.

Accomplishments that we're proud of

We are proud that Agentic-CS goes beyond a demonstration chatbot and functions as an early customer-service operating system.

The platform can generate evidence-based responses, connect conversations to structured business information, perform controlled tool calls, request approval for sensitive operations, and hand cases over to human support.

We are especially proud of the replayable audit system. Administrators can inspect how an answer was produced, which sources were retrieved, which tools were called, what policies were triggered, and where human approval was involved. This makes AI behavior more transparent and gives businesses a practical way to investigate errors.

We also developed a reusable architecture that can support different companies without rebuilding the entire agent for every business. Knowledge, workflows, policies, integrations, and interface configurations can be adapted independently.

Most importantly, we demonstrated that an AI customer-service agent can be both capable and controlled. Autonomy and human oversight do not need to be opposing design choices.

What we learned

We learned that building an agentic product is fundamentally different from adding a language model to a chat interface.

The quality of the model matters, but the surrounding system is equally important. Reliable agent behavior depends on structured tools, permission boundaries, evidence management, approval workflows, observability, failure recovery, and clear escalation paths.

We also learned that an agent should not always attempt to answer or act. Knowing when to request clarification, defer to a human, or refuse an unsupported action is an essential part of intelligence.

Another important lesson was that auditability must be designed from the beginning. Logs created only for technical debugging are not enough. Business users need a human-readable explanation of what the agent knew, what it decided, and why.

Finally, we learned that SMEs need systems that are easy to configure and operate. Advanced AI capabilities provide limited value when businesses cannot understand, govern, or adapt them to their actual workflows.

What's next for Agentic-CS

Our next step is to move Agentic-CS from private beta toward real-world pilot deployments with Malaysian SMEs.

We plan to expand the platform with:

• Additional communication channels, including WhatsApp, web chat, email, and social messaging
• More integrations with CRM, e-commerce, payment, logistics, and ticketing platforms
• A visual workflow and policy builder for non-technical administrators
• Improved multilingual and Malaysian-language support
• Automated quality evaluation and regression testing for agent behavior
• Stronger identity verification and role-based access controls
• More detailed analytics for resolution rate, escalation rate, response quality, and customer satisfaction
• Configurable agent personalities and service policies for different businesses
• Industry-specific templates for retail, education, hospitality, professional services, and e-commerce
• Safer transaction workflows for refunds, order changes, bookings, and account-related requests

Our long-term goal is to make Agentic-CS the trusted AI customer-service operating layer for SMEs: a system that can understand customers, coordinate service workflows, and take useful action without removing human control.

Built With

• codex
• openai