How to build for reality, not hype. On 1 November 2025, Zoey turned 1 year old.
Our journey into AI has felt a lot like parenting. When you first have kids, you think a lot about societal issues. We make bold commitments about banning fast food, eliminating screentime and so on. You believe that these are the principals required to raise the best possible humans. A few years on, getting them to school, soccer, music or eating a single vegetable in a week become the real victories. You are still raising good humans, but you realise parenting is much more in the day-to-day than you ever imagined.
In much the same way, when we first began experimenting with AI agents, we expected that the prompting would be complex and the integration to be simple. After all, AI prompt engineering was new and untested while technology integration has been refined over decades. The opposite turned out to be true. Coding the bot is only a small part of the work, the real challenge lies in the backend engineering, the orchestration of data and the architecture required to make conversations possible.
The hard work sits where data and system design is and decides how well an AI agent performs. In a contact centre, for example, a human agent moves through multiple systems while talking to a customer, pulling information, verifying details and updating records. For an AI agent to do the same, it needs access to structured data through APIs and logic that can mimic human behaviour at machine speed.
The problem is that legacy architectures were built for humans using user interfaces on a screen. While a lot has been done over time to create API led decoupled architectures, we still found the bulk of our work has been focused on refactoring the backend so that the AI can operate effectively. Like the proverbial iceberg, for every one hour spent writing the prompt, another nine are spent building the infrastructure and pluming to allow the AI agent to work.
Guardrails and governance
A major lesson we learned early in our AI agent journey is that freedom is risky. As impressive as large language models (LLMs) are, they still hallucinate. Like a young child, AI wants to please you and will seldom not give an answer. When it doesn’t know, it tries to fill in the gaps with what it thinks the answer MIGHT be.
In a corporate environment where an AI agent represents your brand and makes commitments to customers, there is no room for error and creative thinking. The lesson here is that AI needs clear parameters and business rule guardrails so it only operates within the scope it has been trained for. When a customer asks a question outside that range, the right response is to hand the conversation back to a human.
We’ve found that limiting the AI’s freedom has been critical to delivering predictable and compliant outcomes. Data privacy and protection add another layer of complexity. AI systems need access to the right information at the right time but never more than that. Human agents perform better when they have all the information in front of them. With AI, the opposite is true. Each step of an AI journey must be given only what it needs to perform a specific task. Preventing unnecessary data movement reduces the risk of data leaks and ensure compliance with both regulatory and reputational standards.
The art of conversation
The second lesson is that natural conversation is, surprisingly, highly technical. Making a bot sound human depends on latency, timing, tone and the subtle details of how people really talk. Humans overlap, they interrupt and use verbal cues like ‘uh’ or ‘ah’. These details, when engineered correctly, create flow and comfort. Even simple additions like background noise or the ability to recognise emotional tone can dramatically improve engagement.
What we’ve realised is that people are willing to talk to bots when the interaction feels natural and efficient. It’s this balance between empathy and efficiency where the sweet spot lies. An AI agent doesn’t need to trick a customer into thinking it is human. It simply needs to create enough trust that the customer believes the conversation is worth having. When that happens engagement and outcomes improve. In collections, for example, customers are more honest with bots which leads to a higher follow-through rate. Without the awkwardness or taboo that comes when speaking about debt to another human being, they commit to payment plans they can actually fulfil.
Building skills not scripts
The first time you interact with AI, it can feel like you’ve truly been transported into the future. In your mind at least, science-fiction becomes science-fact as the AI effortlessly answers your simple and complex questions with amazing easy and fluency. As you start testing in process driven use cases, you tend to realise the articulate and comprehensive nature of the interaction hides a simple truth – the AI does what you tell it do and will only impose boundaries that you tell it are there.
So, like training a young but very articulate child, you start with clear boundaries and simple rules. Over time, you add layers of skill and judgment. You begin with behavioural prompting – what it can say, how it should respond, what is acceptable. Once that foundation is stable, you can build libraries of specific capabilities such as handling empathy, resolving exceptions or performing defined tasks.
This progressive learning process creates consistency, safety and scalability. From a tooling perspective, quality matters more than novelty. The open-source ecosystem is exciting and tempting, but commercially supported components promise and tend to deliver greater stability, data protection and support. When a machine is engaging directly with customers and making commitments on behalf of a business, these are not trivial issues.
Cost is also a major factor. Poorly constructed prompts can burn through tokens at an alarming rate and combined with an overly verbose LLM, the business case compared to humans can evaporate very quickly. Thus, optimisation of all components becomes critical. Each skill and its underlying architecture must be evaluated to balance cost vs complexity vs performance and requires fine tuning and careful component selection as you scale from experimental to production.
The human equation
Perhaps the most important lesson for us is that the best results come from combining AI and people, rather than choosing between them. When autonomous agents handle most of the repetitive work and humans focus on the exceptions, the overall performance improves.
It creates a better work environment, more consistent service for customers and, ultimately, stronger outcomes for clients. We’ve seen that play out in the numbers too. Around 80 to 90 percent of the repetitive work can be handled by an AI agent, the rest is looked after by humans who bring the judgment and experience that machines still lack. You don’t need AI and humans working on the same problem, that’s over-engineering human–AI collaboration.
What we’ve learned with AI agents is that when the automation takes care of the heavy lifting and people handle the edge cases, the whole system performs better — the conversations are smoother, the outcomes are stronger, and the experience improves for everyone involved. That’s why the real work with AI agents isn’t about hype or headlines, they only work when everything behind them works too.
Looking ahead
As we embark into Zoey’s second year of life, we are now truly AI parents. The world of not-an-AI-parent is long forgotten, and we now need to solve for 2-year-old problems. So, what might those problems be?
From what we can see, the core component technologies have reached a very useful level of maturity and clear winners in the enterprise space have emerged. My prediction is that the next year will focus on 3 things 1) building out of skills so the AI can increasingly do more 2) advancing the platform to balance cost vs quality 3) refactoring the current operating model as AI starts to reshape work.
And of course subjecting our friends and acquaintances with photos from our 1-year-olds birthday party…
Link: https://www.nutun.com/insights/lessons-from-ai-agents
Hans joined Nutun in March 2023 with 19 years of technology experience across multiple industries including financial services, telco, retail and insurance.
Hans previously served as the Chief Product Officer for TransUnion Africa with the responsibility of developing and maintaining innovative customer engagement and credit risk management solutions leveraging the latest digital and analytics technologies.
Hans also held the role of Managing Director of Technology Strategy at Accenture specialising in technology led business transformation, having led several major transformation initiatives across the Africa continent.