Dr Nadine Kroher presents at AI Public Sector Week

Beyond Chatbots: Where AI Goes When Language Isn’t Enough

‍

At DigiWeek, our Chief Scientific Officer, Nadine, took a step back from the current noise around AI and asked a simple question: what happens when language isn’t the problem you’re trying to solve?

‍

Because while large language models dominate the conversation, they only represent a small slice of what AI actually is.

‍

‍

‍

The AI everyone sees vs the AI that runs everything

‍

Most people experience AI through a chat interface. You type something, it responds. It writes emails, summarises documents, helps with presentations. Useful, accessible and very visible.

‍

That visibility is exactly why it’s become the default mental model for but it’s also misleading.

‍

A huge proportion of AI systems don’t generate text at all. They sit quietly behind the scenes, optimising, predicting, allocating and controlling. You don’t prompt them but you interact with them constantly.

‍

Every time Amazon delivers a package faster than expected, AI is at work.

‍

Every time Spotify gets your playlist right or Netflix suggests something worth watching, that’s machine learning.

‍

Every time traffic flows just a bit more smoothly or a factory line runs efficiently, there’s AI in the loop.

‍

‍

Where LLMs actually fit

‍

Nadine broke this down clearly. AI is a broad, often loosely used term. Machine learning is more specific, systems that learn patterns from data rather than following hard-coded rules. Deep learning is a subset of that, using neural networks with many layers.

‍

And LLMs sit inside that world as one particular type of system: generative models designed to produce language (and now images, video and more). They are powerful but they are not universal.

‍

If your problem is about generating content, they’re often a great fit. Hoowever, If your problem is about making decisions over time under uncertainty, they usually aren’t. That’s where reinforcement learning comes in.

‍

‍

Reinforcement learning, intuitively

‍

Rather than learning from static data, reinforcement learning is about learning through interaction.

‍

An agent observes a situation, takes an action, and receives feedback in the form of a reward or penalty. Over time, it learns a strategy that maximises that reward.

‍

Nadine illustrated this with a simple example: teaching a system to play a racing game.

‍

The system sees the world as a set of states (the screen, speed, position), takes actions (steer, brake, accelerate), and learns from outcomes (win, crash, go faster). Through repeated simulation, it develops a policy, a way of acting in any given situation.

‍

This is fundamentally different from predicting the next word in a sentence.

‍

‍

From games to the real world

‍

The obvious question is whether this translates beyond games.

‍

It does, but not in a trivial way.

‍

The hardest part isn’t training the model. It’s framing the problem correctly: defining the states, the actions, the environment and most importantly, the reward.

‍

When that’s done well, reinforcement learning becomes incredibly powerful.

‍

Nadine shared three examples that bring this to life.

‍

‍

Smarter battery charging

‍

Fast charging is convenient, but it degrades batteries. Charging slowly preserves battery health, but frustrates users. This is a classic trade-off.

‍

Using reinforcement learning, an agent can learn how to dynamically adjust charging behaviour based on real-time signals like temperature, voltage and charge level, as well as past charging history.

‍

The result is a strategy that balances speed and longevity in a way fixed rules simply can’t.

‍

‍

Optimising aircraft maintenance

‍

Airlines operate under tight constraints. Aircraft must undergo maintenance based on flight hours, cycles and calendar time. Schedules change constantly due to delays, weather and operational issues.

‍

Planning this manually is complex and often suboptimal.

‍

Here, reinforcement learning can learn when each aircraft should be sent for maintenance, balancing cost, availability and safety constraints over long time horizons.

‍

In practice, these systems have shown they can outperform human planners on both efficiency and reliability.

‍

‍

Responding to wildfires in real time

‍

Wildfire response is one of the most dynamic and high-stakes environments imaginable.

‍

Fires spread unpredictably, resources are limited, and decisions made now can have consequences hours later.

‍

Using simulation models of fire spread, reinforcement learning agents can be trained to allocate resources such as helicopters, ground crews and firebreaks in real time.

‍

Importantly, these systems don’t replace human expertise. They augment it, offering recommendations and reasoning that help teams make better decisions under pressure.

‍

‍

The real takeaway

‍

The point of Nadine’s talk wasn’t that LLMs are overhyped or unimportant, It was that they’re only one tool.

‍

AI is not a single capability. It’s a toolbox of different approaches, each suited to different types of problems. Language generation is just the most visible one right now.

‍

The more interesting work often happens elsewhere. In systems that optimise, adapt and decide, often invisibly.

‍

If we only focus on chat interfaces, we miss where AI is already creating the most value. And more importantly, we risk applying the wrong tools to the wrong problems.

‍

At Passion Labs, this is the space we spend most of our time in. Not just asking what AI can say, but what it can do.

‍