What does it take to build reliable, intelligent virtual agents that delivers real value? That’s the question I posed to Anne Jenkins, VP of Solution Architecture at Teneo, in a wide-ranging conversationwide-ranging conversation covering conversational AI, agentic systems, implementation challenges and why hybrid approaches may hold the key to success.
The challenge: understanding what agentic AI really is
The interest in agentic AI is still high, but confusion remains as to what it actually is. That’s why I put together this deep dive that explains in detail what AI agents actually are. Anne explains that many vendors are stretching the term to suit their marketing, often applying it to simple LLM chatbots that do little more than classify a query and hand-off to a human agent.
In fact, even if you look at OpenAI’s recent Agent Builder platform: in the demonstration agent, provided to show you a preview of what’s possible, the first example given is an ‘agent’ that simply classifies a user utterance to an intent, and saves that intent as a variable. Call me crazy, but this is not an agent.
In fact, I entirely concur with Anne. This is why I put this piece together a while ago explaining what’s actually going on under the hood of many of these supposed enterprise agentic AI platforms.
True agentic AI, Anne argues, requires memory, autonomy and the ability to execute a series of actions based on a goal. It’s not just about responding fluently, it’s about doing something useful, many times independent of specific instruction.
Yet many implementations still fall into the trap of treating LLMs like magic wands. This often leads to brittle systems that sound convincing, but break under pressure, particularly in complex business use cases where consistency and accuracy are non-negotiable.
The solution: hybrid AI that combines rules and language
Anne shared some perspectives on Teneo’s approach, which is tracking similarly to other market leaders in the enterprise conversational AI space. There’s maturity building in that market where providers aren’t viewing LLMs as the whole solution, but rather a tool in a broader system.
The Teneo platform blends rule-based logic, natural language understandingnatural language understanding (NLU) and LLMs to build what Anne calls intelligent, dynamic and robust agent experiences.
For example, in a demo shared during the conversation, Anne showed how users can teach Teneo new things by speaking naturally. Turning a conversational AI platform into a conversational UI system itself. The system captures examples and generates multiple dynamic system responses, each one consistent in content, but varied in language.
Importantly, it caches those responses for efficiency, rather than calling an LLM each time, reducing cost and latency, and increasing reliability. This is a technique I think you’ll see become commonplace over time.
This ability to intelligently control when and how to use LLMs, combined with a structured memory and workflow, is crucial for creating scalable, sophisticated use cases that maintain accuracy, context and compliance.
Implementation: how to build AI that works in the real world
Anne is rightly points out that the key to successful implementation is putting the customer at the centre:
“The need for good conversation design hasn’t gone away.”
Instead of jumping straight into prompts and model selection, we’ve always recommended that companies should start by mapping out what users are trying to achieve, what the ideal experience could be and where automation can support that journey.
Anne recommends taking small, focused steps:
- select a couple of priority use cases
- prototype with real data
- pilot with real customers
This helps avoid common pitfalls, like building overly broad systems or assuming that generative AI will ‘just work’. It also shows that the tried and tested design principles and approaches haven’t changed, just because we have some new tools in our arsenal.
Another important point Anne raised, which I do honestly think is fairly unique, is the ability to hold conversation memory across sessions. Having a structured memory gives virtual agents awareness of things like prior calls, user preferences and authentication status. This enables continuity and personalisation, something LLMs on their own (and many other conversational AI platforms, in fact) struggle with.
Whether you’re starting or scaling, the requirements don’t change
For companies already implementing AI, you’ll appreciate that having the ability to connect to multiple LLMs is crucial. This means you can select the best model for each task. Even the best model for a particular sub-task or capability. You’ll also realise that hybrid workflows that combine human-like interaction with reliable business logic are going to become table stakes. It’s the only way large enterprises can reliably operate.
For those not yet in production, Anne’s advice is clear: invest in education, pilot before scaling, and choose tools that allow flexibility.
“The worst thing is buying something when you don’t know what you’re buying,” Anne says.
Ask for live demos, test thoroughly and make sure your AI platform can support change, including plugging-in new models and technologies in the future.
Generative AI isn’t a shortcut
Building agentic AI isn’t about handing over your business to ChatGPT. It’s about combining natural language, structured workflows and memory in a way that solves customer problems from end-to-end with minimal effort.
As Anne puts it:
“Generative AI can do incredible things, but it’s not a magic shortcut. You still need to know what experience you want to create and design for it.”
With the right tools, mindset and support, AI can move from flashy demo to everyday value one conversation at a time.
To learn more about Teneo and its hybrid AI approach, visit teneo.aiteneo.ai and to catch up on the full conversation with Anne, check out the VUX World PodcastVUX World Podcast.
