Artificial Intelligence is everywhere.
Every week there are new language models, Copilot announcements, AI agents, and demonstrations showing how quickly the technology is evolving. Within the Business Central community, we’re seeing AI used for documentation, AL code generation, support assistants, UAT creation, knowledge search, and workflow automation. Looking at social media, it almost feels as though every organization already has an AI strategy.
Over the past several months, I’ve been on a similar journey, building AI agents specifically for Microsoft Dynamics 365 Business Central. Like many others, I started by experimenting with different models, prompts, and tools, believing that selecting the right AI model would be the biggest challenge.
I couldn’t have been more wrong.
The AI model was rarely the problem.
The real challenges appeared the moment I tried to make these agents useful in real Business Central projects.
One of the first issues I encountered was context. An AI can explain how posting works in Business Central, but if a customer reports a posting error, the AI has no understanding of that customer’s Posting Setup, General Ledger configuration, custom extensions, dimensions, or business rules. Without access to that information, it can only provide generic suggestions. The responses sound convincing, but they’re often based on assumptions rather than facts. I quickly realized that AI without business context behaves more like an intelligent search engine than an experienced consultant.
Another lesson came from trying to build one AI assistant that could handle everything. Initially, I wanted a single agent capable of reviewing AL code, generating documentation, creating UAT scenarios, troubleshooting issues, answering functional questions, and supporting implementations. While it could perform all of those tasks to some degree, it wasn’t consistently reliable in any of them. That experience completely changed my approach. Today, I build specialized AI agents with clearly defined responsibilities, dedicated instructions, focused knowledge, and the right tools for a specific job. The improvement in consistency and quality was immediate.
I also discovered that prompt engineering wasn’t the breakthrough many people believe it to be. Early on, I spent considerable time refining prompts, hoping better wording would produce dramatically better results. While prompts certainly matter, they weren’t what transformed the quality of my AI agents. The biggest improvements came from designing structured instructions, defining investigation workflows, limiting assumptions, and teaching the AI how an experienced Business Central consultant approaches a problem instead of simply asking it to generate an answer.
Perhaps the biggest technical challenge wasn’t intelligence at all it was organizational knowledge. Modern AI models already understand ERP concepts, software development, and Business Central fundamentals. What they don’t understand is how your organization implements Business Central. They don’t know your development standards, reusable AL components, implementation methodology, documentation templates, customer-specific processes, or lessons learned from previous projects. I realized that without this knowledge, AI can only provide generic guidance. With it, AI begins to behave much more like a consultant who understands the organization’s way of working. That realization eventually led me to the idea of an AI Knowledge Hub.
Trust became another unexpected challenge. During demonstrations, AI often looks incredibly impressive, but live customer projects are very different. Consultants don’t need an AI that occasionally provides brilliant answers. They need an AI that reasons consistently, follows structured investigation steps, knows when it lacks sufficient information, and asks questions before making assumptions. I found that consultants quickly lose confidence in an AI that sounds certain but produces unreliable recommendations. Building trust turned out to be significantly harder than building intelligence.
As my AI agents became more capable, another realization emerged. AI isn’t something you deploy once and leave unchanged. Business Central evolves, business processes change, coding standards improve, and organizations continuously generate new knowledge. AI agents require ongoing refinement, better instructions, updated knowledge, feedback from consultants, and governance to remain useful. In many ways, they need to be managed just like any other member of the consulting team.
Looking back, I no longer believe that the success of a Business Central AI initiative depends primarily on selecting the best Large Language Model. Models will continue to improve, and new ones will appear every year. The real competitive advantage will come from designing specialized AI agents, building structured organizational knowledge, establishing clear governance, and creating AI that consultants genuinely trust during real customer engagements.
That’s why I don’t think most Business Central AI projects will fail because of AI itself.
They’ll struggle because organizations underestimate everything that surrounds AI and, in my experience, that’s exactly where the greatest opportunity for innovation lies.
Leave a Reply