Advice on leveraging AI in business is abundant, but much of it is self-serving, ours included, so judge us on our track record: 13 years of delivering successful projects to repeat clients who value our integrity and expert guidance. This article outlines our strategic recommendations for AI integration.
The “Language” in Large Language Models (LLMs) is critical. Your industry and business use specialized terminology that differs from everyday usage. Since LLMs are trained on general data, they can easily misinterpret your specific jargon. Your first task is to provide the context for the AI, much like onboarding a new employee.
Knowledge engineering begins with elicitation to produce a domain model. This model is a strategic asset and gives humans and AI the common ground they need to work together well. Beyond simple chat interfaces, a domain model allows AI to assist with employee onboarding, help experts analyze potential business changes, underpin automated testing, and constrain AI-generated reports.
A domain model gives the AI your vocabulary. The next question is what kind of capability you are building on top of it.
Teaching the AI means giving it knowledge - “know how” , and sometimes wisdom. Information can be thought of as data with context. Knowledge, on the other hand, is active "know-how": the ability to apply information. Historically, this expertise resided in human minds or was rigidly encoded in software. LLMs represent a genuine shift, enabling the creation of AI agents with specific skills. Music theory, for instance, is information, but knowing theory alone is insufficient; improvising a melody is a skill grounded in knowledge. Such skills can now be taught to AI agents.
To see how this plays out in practice, consider an airport hotel. It uses standard hospitality terms (check-in, occupancy) alongside airline jargon (layover, red-eye) and internal status levels (Gold, Silver, Bronze). When managing a high-status customer delayed by weather, the "correct" response depends on many factors. This is where wisdom comes in. The situation requires a meta-level of knowledge to guide judgment and find optimal outcomes in unanticipated situations. There's no pre-canned answer here, which is exactly what makes it worth teaching: the AI has to work through different strategies and pick the best one. This is where AI can excel, but also where it can hallucinate; grounding decisions in a domain model converts hallucinations into innovation and insight.
Individual skills and judgment calls only become valuable when they are repeatable. The goal is to capture your operating procedures so humans and AI can collaborate on both execution and improvement. Unlike the pre-LLM era, which required expensive software engineering to maintain systems, you can now use agents that access your domain model. These agents must, however, be constrained and supervised to ensure they learn and improve within safe parameters.
In practice, this might mean an agent proposes a new escalation path after handling a tricky case, a human reviews it, and the system folds it into how it operates going forward. The procedure improves because someone paid attention to what worked.
Supervision raises an obvious question: how reliable does the AI need to be? AI doesn't need to be perfect; it just needs to outperform what you’re doing now enough to be worth it. While you can constrain AI to be entirely predictable, doing so sacrifices the reasoning capabilities that make it special. Instead, humans and AI need a shared set of rules:
A chronological decision log lets both sides review this transparently.
When correctly implemented, AI can:
While AI can automate many tasks, it is not a wholesale replacement for humans. Real wisdom, knowing when to deviate from standard procedure, remains a human trait that is difficult to encode into agents.
If you’re weighing where to start with AI integration, we're happy to talk it through. Get in touch at [contact link].
