Specialist know-how can beat raw model size — but where should that expertise actually live: training, tools, or the people using it?

This looks at how adding specialist knowledge (medicine, engineering, a company's internal know-how) changes what an AI system can do, and at where that knowledge works best: in the model's training, in the tools around it, or in the people using it. The corpus keeps coming back to one finding. How expertise is structured often matters more than how big the model is. One example fine-tuned a mid-sized model on 24,000 reasoning problems built by tracing paths through a medical knowledge graph. That model reached state-of-the-art results across 15 medical specialties. It worked because it learned to combine small pieces of medical knowledge, not because it had more parameters (Can knowledge graphs teach models deep domain expertise?). A broader thread argues that AI built to learn only from raw data, without explicit rules, pays a price. It ends up harder to interpret, picks up biases that nothing corrects, and breaks more easily on unfamiliar cases. Adding a small amount of structured knowledge closes much of that gap (Does refusing explicit knowledge harm AI system performance?).

Expertise also doesn't have to live inside the model. In an industrial case study, a team wrote their specialists' rules and design principles into the scaffolding around an LLM agent, meaning the prompts, checks and workflow steps. Output quality rose 206%, and non-experts produced work that experts rated at expert level (Can codified expertise let non-experts match specialist output?). The model didn't improve. What changed is that tacit know-how became explicit and reusable. That makes the question less about whether the AI knows the domain and more about whether anyone has written the domain down. Read this way, the four main ways to inject knowledge are choices about where expertise should sit: retrieval at query time (RAG), baking it into the weights, swappable adapters, or prompt design. Each trades flexibility against cost, and combining them beats any single one (How do knowledge injection methods trade off flexibility and cost?).

Here's the counterintuitive part. Sometimes expertise works better when you don't spell it out. In medical AI, sophisticated diagnostic reasoning emerged from reinforcement learning on hard problems with only a right-or-wrong reward. Nobody had to show the model an expert's step-by-step reasoning (Can simple rewards alone teach complex domain reasoning?). A related approach rewards models for both correct answers and sound explanations. It embeds domain knowledge better than standard fine-tuning, because it rewards coherent understanding rather than copying expert wording token by token (Can reinforcement learning embed domain knowledge more effectively than supervised fine-tuning?). The flip side: agents trained only on expert demonstrations are capped by what those experts thought to demonstrate. They never get to fail, explore, and learn beyond the examples (Can agents learn beyond what their training data shows?). Expert knowledge can raise the floor and also set the ceiling.

Specialization has costs too. The corpus describes a capability cliff. Models trained too narrowly fail badly outside their domain, while models trained too broadly give confident but wrong answers in high-stakes settings (How do you build domain expertise into general AI models?). Every training technique has a sweet spot tied to its domain, and the visible gains often hide losses: reasoning that is less faithful, skills that don't carry over, and less flexibility about output format (How do domain training techniques actually reshape model behavior?). One proposed way around the depth-versus-breadth trade-off is to let a model mix in different expert skills on the fly at inference time, without retraining the whole thing (Can models dynamically activate expert skills at inference time?).

Finally, expertise in the model doesn't guarantee results in the world. Agents that look capable still complete only about 30% of real workplace tasks. Smart routing between models and well-designed interfaces often matter as much as the knowledge itself (What breaks when specialized AI models reach real users?). So the full answer is that domain expertise changes what AI can do mostly through where it's placed: in a training curriculum, in the scaffolding, in reward signals, or in how the system meets its users. The largest gains in this corpus came from turning experts' unwritten judgment into structure a system can use, more than from adding knowledge.