After two years of building enterprise-grade conversational AI agents and powering tens of millions of real customer interactions, we’ve learned some hard truths—what held up in production, and what didn’t. These lessons were not formed in theory, but validated (and challenged) in real enterprise environments. Before diving into them, I want to thank our early
adopters and believers. Without their trust, openness, and willingness to build alongside us, we would never have had the opportunity to pressure-test our assumptions in the real world.

What we were right about

• A deterministic backbone is essential to balance LLM intelligence and hallucination. This was true two years ago—and it remains true today, even with the latest LLMs.

  
  

  LLMs are powerful, but they are fundamentally unreliable when asked to autonomouslyreason across multiple issues, long conversation histories, and policy-heavy enterprise workflows. Left on their own, they optimize for plausibility, not correctness.

  
  From day one, we designed a deterministic backbone to own flow control, state, and decision boundaries. LLMs are invoked deliberately and narrowly—primarily as classifiers or extractors, each with a clear purpose. Only when the system reaches high confidence do we allow the conversation to progress or trigger the next LLM call. Building this way requires discipline. We constantly have to resist the temptation to “just let the LLM decide” to reduce engineering effort. At the same time, we are equally careful not to over-engineering rigid rules that strip away conversational flexibility and turn AI agents into robotic chatbots. Enterprise-grade performance lives in the middle. The deterministic backbone provides correctness, governance, and debuggability. LLMs deliver language, flexibility, and human-like interaction

  
  

• Human-level quality in customer experience is both achievable and sustainable. From the beginning, we made a bold commitment: our AI agents would be benchmarked against real call-center performance. If we failed to meet human-level quality, our partners would get a full refund. That promise shaped everything—from architecture to evaluation to how fast we iterate post production.

  
  So when we learned that one partner’s in-house call center was running at a 95% CSAT, I was genuinely concerned. Unfiltered CSAT at that level is rare. Matching and sustaining it felt like a stretch.

  
  What happened surprised us. Not only did we reach that bar, we have maintained it for over 18 months in production. Last year, that same partner went on to win Singapore’s Best Customer Service award, with our AI agent operating as a core part of their customer support stack.

  
  

  This doesn’t mean AI is always better than humans. It isn’t. There are cases where experienced agents outperform AI, and cases where AI is more consistent, faster, and less error-prone. The real insight is reliability at scale. By grounding our system in deterministic control, rigorous evaluation, and continuous benchmarking against human outcomes, we’ve been able to keep a promise many teams quietly walk back: delivering human-level quality not just in demos, but day after day in real operations. That consistency—not novelty—is what makes AI transformation stick for our partners.

What we were wrong about

• Language localization is not automatically solved by LLMs. At the beginning, we assumed modern LLMs could handle language localization well enough. In real production, this turned out to be wrong—especially in multilingual markets.

  
  In places like Hong Kong, there is no single dominant language. A single customer message may mix multiple languages in one input. Aligning the customer’s language, the knowledge-base language, and the AI’s response language is not trivial. Generic language detection often fails, which leads to incorrect answers or responses that do not follow brand policies. We eventually had to build custom language-switching logic instead of relying on the model alone.

  
  Voice made this even clearer. Across all voice models we tested, Malay speech was more sensitive to noise and interruptions than English. We initially thought we could reuse the same playbook instructions for both languages. That assumption was wrong. To reach acceptable quality, Malay voice agents needed tighter scripts and stronger guidance of the conversation flow than English agents.

  
  The lesson is simple: language localization is not just a model problem. It is a system design problem.

  
  

• One playbook does not fit all channels.We initially believed that with sufficient control at the interface layer, a single playbook
  could resolve the same customer issue across all channels. The underlying logic, after all, should remain the same.

  
  What we underestimated were the fundamental differences between channels. Email, for example, does not support the same conversational dynamics as chat. Asking one question at a time—perfectly acceptable in chat—can feel like deliberate deflection in email. Rich UI elements such as carousels or quick-reply buttons work well in chat but are unavailable outside of it. Even authentication flows must vary: the level of strictness depends on the security characteristics and accuracy of each channel.

  
  

  We still believe chat is the channel where AI agents can deliver the best experience. But enterprise customer experience is inherently multi-channel. Our responsibility is not to force customers into the channel that suits AI best, but to design agents that work well wherever customers choose to engage.

These lessons are not conclusions—they are waypoints. Enterprise conversational AI is still early, and the problems worth solving are getting harder, not easier. The journey has only just begun. We will continue to share what works, what breaks, and what we learn along the way—openly and honestly—as we keep building alongside our partners.