It won't surprise you to know that at Unit4 we believe we are ahead of the curve when it comes to AI. I'm sure you've heard that from many technology vendors. As proof of our leadership in this space, we often point to the fact that back in 2016 we launched Wanda, one of the first chatbots for mid-market organisations.
What you may not know is that about two years ago, as a leadership team, we had a frank discussion about whether to shelve the concept along with our investment in the idea of ambient ERP. This vision of the technology would see it operating autonomously in the background, only interacting with users when it required human intervention, and the user could have a simple conversation in the application they used every day, such as Microsoft Teams.
The problem was take up. Our customers weren't keen on the approach. Maybe it was the technology, maybe it was how we were selling it.
Then along came ChatGPT and everything changed.
Looking back now, hindsight makes it obvious what was happening. Users and technology vendors were only at the start of a process of experimentation.
If we are all honest about AI, we are only now beginning the serious work of understanding how it affects the way we work. It is reshaping workflows which will affect organisational structures. Some roles may go away, but other roles will be required. And AI requires different skills which will make humanities degrees just as relevant as engineering ones.
Throughout this period of change, we are working out the business cases for AI on the job. Anyone who tells you they have fully worked them out is lying. This will evolve rapidly, but we should be honest, vendors and users alike, that the next year will be one of experimentation to find out where AI improves productivity, gives back time and enables users to focus on more rewarding work.
It is why vendors like Unit4 are offering customers commitment-free access to AI. Our initiative, AI for Your World, enables mid-market organisations to experiment without requiring a subscription. Developing Wanda taught us that it is only through collaborative trial and error that we can create a feedback loop which informs our product development cycle.
Let me illustrate it through an example based on our engineering teams' research. Since the start of the year, we have been modelling what effect using AI coding tools would have on the way we developed products. What we have concluded, in part, validates the marketing hype. The modelling indicates we could accelerate product development with products being developed 5–15 times faster.
That, though, is only the headlines. Integrating AI into our development cycles is causing us to reassess the essential approach to engineering new products. If AI is taking on more of the code production work that is an obvious benefit in terms of resource savings, but that efficiency would be squandered if we didn't expend more effort on planning.
It requires a much deeper level of interrogation of the problem the engineering teams are looking to solve. It is why I believe individuals with humanities degrees, especially philosophy, will be well placed to assess the challenges an organisation wants AI to address.
This is one of the key learnings we are feeding back to our customers as they look to find the right business case for AI.
However, the feedback loop is two-way. With Wanda we learnt early on from our customers that business context is critical. A project means one thing in professional services, another in the public sector and something else in non-profit settings. Unless the AI understands the context for a user's question, it will not know what information to extract from the ERP, and it will very likely hallucinate. Knowing this led us to build very rich descriptions of all the data operating in our next generation platform, ERPx. All the objects have been described through semantics and an ontology. On top of this we have layered all of the business processes in a format that allows the LLM to reason. In total we have described 1,200+ data contracts to reduce the risk of the LLM hallucinating.
Governance and compliance are also important to our customers leading us to build a Central Business Logic (CBL). This layer contains all of the transactional business logic. Rules are defined in the CBL as deterministic rules so that if the LLM or the ERP tries to update an invoice this mechanism will make sure that it is safe to do so. If it is the wrong response, the CBL will feed back to the system with LLM interpretable error messages it can act on within the same loop. This way we guard the transactional integrity.
Another area where the feedback loop is critical is understanding how customers think about their businesses. For example, a CFO in a professional services company may look at the P&L differently to someone in a non-profit organisation. This information must be captured and even though we provide functionality out of the box a lot of the work must be done in the on-boarding process to distil how the business sees its operations. This is a completely different way of thinking about software as previously it was just a configuration exercise. With AI, you can still have custom fields, but you must tell the AI what they are, otherwise the LLM will be very confident, but totally incorrect.
Sovereignty is a more recent, and important requirement for a number of our public sector and non-profit customers. They need to know that data is not leaving the EU to train AI models in other jurisdictions. That's why we host the models we use within the EU, in collaboration with European partners.
What 30 odd years working in the technology industry has taught me is that we need to overcome our aversion to experimentation. CIOs have been cautious about experimentation as it suggests something is incomplete and cannot be relied upon in production environments. Yet, as we look to get to grips with AI, we must overcome our natural European caution and explore different ways to use the technology. Yes, it involves risk taking but in the safe environments we are setting up we want to encourage such trial and error as it is critical to building all of our understanding of the best use cases for AI.
This article was originally published on computerweekly.com.
