Automating Customer Support with AI Agents and Data Streaming

Support teams live in a world of repetitive questions, fragmented tools, and growing customer expectations. Customer service agents bounce between customer relationship management (CRM) systems, ticketing, email, and chat while customers wait, often repeating the same information across channels. Batch-based systems are unscalable for AI: Context is always a step behind, escalations pile up, and it’s difficult to intervene in time.

At Thunai, we’ve built an agentic AI platform that deeply understands every interaction in real time—and can act on them safely across the tools support teams already use. From capturing live customer conversations to updating CRM records and automating workflows, our mission is to be the leading agentic AI platform for automating and enhancing professional productivity and customer support, with a core goal of maximizing Level 1 (L1) task deflection while improving customer experiences.

To deliver on that mission, we needed more than just powerful large language models (LLMs). We rely on a real-time data streaming backbone to keep our AI agents continuously up to date with fresh, reliable context. That’s why we’d joined the Confluent for Startups program, which helps early-stage companies like Thunai adopt data streaming from day one.

Transforming Live Customer Conversations into Agent Workflows

At Thunai, data quality and fresh context are not nice-to-haves—they’re existential requirements for agentic AI. Our platform’s core promise is that it can not only summarize and understand interactions but also take actions on behalf of users and support teams.

AI is the core base of our operations. We depend on deep contextual understanding of meetings, user requests, and multichannel interactions to drive agentic automation that can safely act across the tools people already use.

Our agentic AI platform is designed to:

• Optimize customer support by solving prevalent issues and providing maximum deflection for L1 tasks (e.g., basic troubleshooting and helpdesk tasks).

• Automate routine interactions and accurately interpret user intent, significantly reducing the load on human agents and improving response times.

• Increase user productivity by seamlessly automating routine tasks, integrating across tools like CRM and to‑do lists, and providing deep, real‑time contextual understanding for every interaction.

Under the hood, this relies on context engineering. Poor data quality would lead to misinterpretation, incorrect actions, failed automations, and a low L1 deflection rate, which would undermine our entire business goal.

That’s why we treat real‑time meeting transcripts, CRM updates, and user interactions as streaming context that must be captured, enriched, and made available to Thunai’s AI brain as events, not nightly batches. Thunai continuously ingests and analyzes:

• Live customer conversations (voice, chat, email, WhatsApp, and more)

• Real-time transcripts from calls

• Events from CRM and ticketing systems

• User actions and workflow events across tools

All of this flows into Thunai Brain, our platform’s continuously updated knowledge base that powers all generative AI (GenAI) and agentic decision-making. Data streaming is what transforms a static knowledge base into a living system of record for every interaction.

Confluent provides high-quality fresh context that’s paramount for Thunai's AI products because our core functionality—agentic automation and L1 task deflection—depends entirely on an accurate, real-time understanding of the user's situation and intent.

How Data Challenges Hindered Agentic Customer Support

Before we adopted Confluent as our data streaming backbone, a combination of business and technical challenges made it difficult to realize our vision for agentic AI in customer support.

Business Challenges

• Reliance on batch processing led to slow, reactive decision‑making. By the time data arrived, the moment to act had often passed.

• Data was spread across CRM, enterprise resource planning (ERP), logistics, and support tools, creating siloed workflows and preventing a unified, real‑time view of the customer journey.

• Agents spent time switching between systems, which led to high operational inefficiency and context switching fatigue.

• Lack of real‑time insights prevented proactive detection of issues like delays, escalations, and growing customer frustration.

• Customer experience suffered because information across channels was inconsistent and out of sync.

• Increasing customer interactions exposed scalability constraints in our old architecture, especially during spikes in activity.

• Escalations were poorly prioritized due to limited real‑time context, making it hard to distinguish urgent issues from routine noise.

Technical Challenges

On the technical side, batch‑oriented pipelines and legacy integration approaches were a poor fit for real‑time AI:

• Existing systems struggled to ingest and process high‑throughput workloads coming from calls, logs, and user actions.

• Batch pipelines introduced high latency, making real‑time AI analysis effectively impossible.

• Real‑time transcription, sentiment analysis, and keyword detection all required low‑latency streaming that our legacy infrastructure couldn’t support.

• Thunai’s knowledge store couldn’t stay continuously updated with batch workflows, which meant the AI was reasoning over stale context.

• Traditional ETL tools lacked the fault tolerance, ordering guarantees, and durability we needed for voice and log streams.

• Our infrastructure could not reliably handle traffic spikes from calls, events, and agent workflows without degradation.

• We needed event replay and reprocessing for model retraining, debugging, and error recovery, but batch systems weren’t built for this pattern.

• We lacked a unified way to connect multiple systems (ERP, CRM, inventory, etc.); existing integration methods were brittle and hard to evolve.

AI‑Specific Concerns: Context and Safe Actions

Because our primary goal is maximum L1 task deflection and seamless agentic automation, we also had to confront AI‑specific challenges.

Contextual Misinterpretation (Hallucination Risk): Our agents can perform actions such as updating a CRM or creating a ToDoist task. Their ability to do so correctly depends on accurately understanding deep, real‑time context from meetings and user queries. Misinterpreting subtle nuances can lead to incorrect—and potentially damaging—automated actions.

Action Failure and Rollback: Automating actions across third‑party tools requires robust integrations. Failures in execution, API downtime, or misformatted data can break workflows and create a negative user experience. Ensuring transactional integrity and the ability to safely roll back actions is a complex but critical requirement.

These combined pressures made it clear that we needed a streaming‑first architecture that could keep Thunai Brain continuously updated, minimize data loss, and support safe agentic actions in real time.

Why We Use Confluent for Real-Time Context

To power the real‑time, context‑aware nature of our agentic AI platform, we turned to Confluent as the backbone for data streaming. We use it to ingest the streams that feed Thunai Brain and support systems.

Several aspects of Confluent’s data streaming platform were essential for us:

Real-Time Data Streaming Capability

We needed to ingest live data from voice calls, logs, and user actions across multiple tools and channels. Confluent provides real‑time ingestion, which is critical for live transcription, real‑time sentiment, and keyword analysis that we apply on top of these streams.

Low Latency and Real-Time AI Feedback

We use Confluent to stream live transcripts into Thunai Brain during a call, enabling our agents to see and act on context as it’s generated. This allows us to support real‑time agent assistance, suggestions and recommended actions, and sentiment‑aware escalation cues for human agents when needed.

Scalability and Reliability

Confluent’s high throughput and fault tolerance help us handle spikes in voice data, bursts of user events and interaction volume, and high availability requirements for production support environments.

Continuously Updating the Knowledge Base (Thunai Brain)

We treat Thunai Brain as a continuously updated knowledge base where processed data is streamed into it in real time, not batched in periodic loads. This keeps our AI continuously learning and up to date so it can drive accurate L1 deflection and trustworthy automation.

Governance, Data Management, and Security

Confluent also provides schema management and governance features that help us ensure data quality and schema evolution without breaking consumers, compliance, or safe data handling.

Mature Streaming and AI Integration

Confluent’s fully managed Apache Kafka® and Apache Flink® services—and Confluent Intelligence capabilities—give us a path to:

• Real‑time, context‑aware AI

• Event-driven Streaming Agents and real‑time model inference

• A simplified architecture that avoids stitching together many separate systems

Today, we’re using Confluent as our primary streaming backbone, and our agents act on real‑time data from our knowledge base, Thunai Brain.

Data source systems act as producers, continuously publishing interaction events to Kafka topics, while Thunai’s agents consume these events in real time. The consumed data is then normalized, enriched, and stored as vector embeddings, keeping the knowledge base continuously updated without batch processing.

Reference Architecture: Automated, Agentic Customer Support Built on Confluent Cloud

Here’s a high-level look at the reference architecture:

We transcribe audio, vectorize the resulting text, and store these embeddings using MongoDB Vector DB as part of our knowledge infrastructure. We also rely on multiple LLMs—including Gemini 2.5 Flash, Claude Sonnet, and DeepSeek (for research)—on top of this streaming context and vectorized knowledge.

GenAI and Agentic AI Use Cases Unlocked at Thunai

With data streaming, Thunai has built multiple key GenAI and agentic AI use cases:

Automating L1 Customer Support

Most Level 1 support queries are automated end to end by Thunai instead of a human agent, enabling approximately 70–80% deflection rates for customers. Thunai’s AI taps into Thunai Brain to generate context-aware responses grounded in up-to-date knowledge from prior interactions, documentation, and real-time events. The same logic works across channels such as WhatsApp, email, voice, and chat, keeping context consistent.

Real-Time Agent Assistance

During live calls, streaming data is used to generate transcripts, detect sentiment, and identify escalation risks in real time. Agents then provide next-best-action guidance teams, with suggestions, cues, and knowledge snippets powered by the streaming knowledge base rather than static FAQs.

Operational Monitoring and Analytics

Real-time event streaming powers dashboards for ticket backlogs, service-level agreement (SLA) breaches, and issue spikes. Thunai tracks the full interaction timeline across channels, enabling better prioritization and root-cause analysis. Events are tied to consumption for billing and usage metering, all powered by the streaming layer.

Productivity and Meeting Automation

Beyond customer support, Thunai uses the same architecture to automate professional productivity workflows, such as joining meetings and summarizing action items. 

Because these experiences are grounded in real-time streaming data, Thunai can act with high confidence and minimal hallucination risk.

Business Impact: Better Experiences While Safely Automating 70-80% of L1 Support

By adopting Confluent and building Thunai Brain and AI agents on top of real-time, contextualized, and trustworthy data streams, Thunai has realized meaningful gains for both its customers and engineering teams.

Business Impact

With a streaming‑first architecture underpinning Thunai:

• We are able to automate most L1 queries instead of routing them to human agents, achieving around 70–80% deflection for our customers.

• Resolution time has decreased drastically from a few hours to a couple of minutes in many cases, because the AI operates on live context rather than stale data.

• Overall, this has led to a better customer experience in which users receive timely, consistent responses across channels.

Technical Impact

On the technical side, we’ve seen:

• Minimizing data loss, thanks to Kafka’s queueing behavior; if there’s an error, processing can stop safely without losing messages, and we can resume or replay as needed.

• A more robust error‑handling model, where the queue system ensures that problematic events don’t propagate downstream unchecked.

Together, these improvements give us the confidence to let Thunai perform agentic actions—such as CRM updates and automated task creation—on behalf of users with appropriate safeguards in place.

A Foundation for Future AI Expansion

With Confluent and agents in place, we have a strong foundation to expand into more advanced streaming and AI capabilities over time—without rearchitecting from scratch.

The result is a platform where agentic AI is not just an overlay but is deeply integrated into the fabric of real-time customer interactions and operational workflows.

Get Started: Build Agentic AI on Streaming Data

Thunai’s journey underscores a key principle for AI developers: If your business depends on real-time understanding and action, your AI needs a streaming foundation, not just batch pipelines. 

Start by ingesting events into Kafka, making streams your system of record for interactions across channels. Then continuously enrich and feed context into your AI knowledge base, which unlocks the ability to use system events to trigger automation, alerts, and workflows in real time. With Confluent’s fully managed platform, you can stand up this backbone quickly, and then use contextualized data to build reliable agents—without taking on infrastructure heavy lifting.

If you’re seeking AI for customer support or professional productivity, learn more about Thunai’s approach and products here. 

Sign up for Confluent Cloud to explore how Confluent’s data streaming platform can help you power real-time, agentic AI at scale.

Apache®, Apache Kafka®, Apache Flink®, Flink®, and the Flink logo are trademarks of the Apache Software Foundation in the United States and/or other countries. No endorsement by the Apache Software Foundation is implied by using these marks. All other trademarks are the property of their respective owners.