--- title: Fine-Tuning vs. Prompting for Domain Expertise description: Fine-tune for tone and format. Prompt for reasoning and policy. Use tools for facts. Here is when each approach wins for travel AI. canonical: https://nowah.xyz/blog/fine-tuning-vs-prompting-domain-expertise lastModified: "2026-08-07T08:06:13.113Z" --- # Fine-Tuning vs. Prompting for Domain Expertise Fine-tune for tone and format. Prompt for reasoning and policy. Use tools for facts. Here is when each approach wins for travel AI. When we started building our AI travel agent, the question came up in the first week: should we fine-tune a model on travel data, or just prompt a general-purpose model really well? Two years later, I can tell you the answer is neither. And both. It depends entirely on what problem you are trying to solve. There are three distinct approaches to giving an AI agent domain expertise, and each one is good at something different. Getting this taxonomy right saved us months of wasted effort. ## Three approaches, three problems ![Illustration for this section](https://pics.nowah.xyz/website-media/ai-research-037-img-1.webp) **Fine-tuning** modifies the model's weights using domain-specific training data. It changes how the model "thinks" at a fundamental level. This is excellent for tone, format, and vocabulary. It is slow to update, expensive to maintain, and unnecessary for most reasoning tasks. **Prompting** gives the model instructions and context at inference time. It shapes reasoning, enforces policies, and adapts to real-time situations. It deploys instantly, costs nothing to update, and works remarkably well for complex decision-making. But it cannot change the model's base capabilities or style at a deep level. **Tool use** connects the model to external systems that provide real-time data and execute actions. It eliminates hallucination for factual queries by grounding the model in verified data. The model does not need to "know" that a flight from JFK to NRT costs $847. It calls the search API and gets the current price. The mistake most teams make is using one approach where another would be better. I have seen teams fine-tune models on flight pricing data that goes stale in hours. I have seen teams write 10,000-word prompts to achieve a consistent tone that fine-tuning handles in a few hundred training examples. I have seen teams try to prompt their way to factual accuracy when a simple API call would give them ground truth. ## Fine-tuning wins for tone, format, and vocabulary Our AI agent has a distinct personality. It is knowledgeable but not stuffy, opinionated but not pushy, concise but not terse. Getting this right through prompting alone was a constant battle. The model would drift. Some responses would feel like a travel blog. Others would feel like a customer service script. The inconsistency was noticeable. Fine-tuning solved this. We trained on a curated set of example conversations that demonstrated the exact tone, format, and vocabulary we wanted. The result was immediate and consistent. The model stopped producing responses that felt "off brand" because the desired style was baked into the weights. Fine-tuning also works well for output format. If you want your model to consistently structure flight recommendations in a specific way, with a brief summary followed by [three options](/blog/why-three-options-not-three-hundred) with clear tradeoff explanations, fine-tuning achieves this more reliably than prompting. Prompted formatting instructions work most of the time. Fine-tuned formatting works virtually all of the time. The domain vocabulary benefit is real too. A fine-tuned model naturally uses terms like "codeshare," "fare class," "minimum connection time," and "revenue passenger kilometer" without needing them defined in every prompt. ## Prompting wins for reasoning and policy ![Supporting diagram](https://pics.nowah.xyz/website-media/ai-research-037-img-2.webp) Here is where prompting shines and fine-tuning falls short: anything that needs to change quickly or reason about novel situations. Travel policies change constantly. An airline updates its cancellation rules. A country changes its [visa requirements](/blog/ai-agents-visa-requirements-documents). A new pandemic restriction appears. If these policies are baked into model weights through fine-tuning, updating them means retraining, which takes hours to days and costs money. If they are in the prompt, updating them takes minutes and costs nothing. Reasoning patterns are the other big win for prompting. When we want the agent to think through a complex [multi-city](/blog/multi-city-flight-booking-ai-agents) itinerary, we can describe the reasoning approach in the prompt: "Consider total travel time including layovers, not just flight duration. Weight the user's stated preferences above general best practices. If two options are within 5% on price, prefer the one with the better schedule." These instructions shape how the model reasons about the specific problem at hand. They can be adjusted per query type, per user segment, or per experiment without touching the model itself. We also use prompting for safety rails. The agent should never book a flight without explicit user confirmation. It should never share one user's data with another. It should refuse to help with anything unrelated to travel. These policies are in the prompt because they are absolute rules that we need to update immediately if we discover an edge case. ## Tool use wins for facts and actions This is the one I feel most strongly about. **Never fine-tune or prompt for factual accuracy about real-time data.** Use tools. When a user asks "what is the cheapest flight to Tokyo next week," the correct answer changes every few minutes. No amount of training data or [prompt engineering](/blog/prompt-engineering-travel-agents) can give the model this information. It needs to call a flight search API that returns current inventory and pricing. Tool-augmented generation eliminates the hallucination problem for factual queries. The model does not guess at flight prices or hotel availability. It retrieves them from authoritative sources. The user sees real options with real prices that they can actually book. We have over a large set of tools in our agent's toolkit, covering everything from flight search to hotel availability to visa requirements to weather forecasts. Each tool is a structured function that the model can call with specific parameters and receive verified data in return. The model's job is to decide which tools to call, interpret the results, and present them in a useful way. The tools' job is to provide accurate, real-time data. This separation of concerns is the foundation of reliable AI travel booking. ## The cost and maintenance burden Fine-tuning has a hidden cost that is easy to underestimate: maintenance. A fine-tuned model is a snapshot. It reflects the training data at the time of training. As your product evolves, as user expectations shift, as the travel industry changes, the fine-tuned model drifts from what you need. You have to retrain periodically, which means maintaining training data pipelines, quality-checking datasets, and managing model versions. Prompts, by contrast, are version-controlled text files. You can update them in a pull request, review the changes, deploy them instantly, and roll back if something goes wrong. The iteration speed is orders of magnitude faster. For our agent, the prompt changes multiple times per week. Small adjustments to reasoning instructions, new edge case handling, refined personality guidance. If each of these required a fine-tuning run, we would move at a fraction of our current pace. ## The hybrid approach we actually use In practice, we use all three approaches in a layered architecture. The base model is lightly fine-tuned for tone, format, and domain vocabulary. This gives us a consistent personality and communication style without prompting overhead. On top of that, extensive prompts provide reasoning instructions, policy enforcement, and real-time context. The prompts adapt per query type. A simple flight search gets a different reasoning framework than a complex multi-city planning session. And surrounding everything, the tool layer provides ground truth. Real-[time pricing](/blog/real-time-pricing-data-pipeline). Live availability. Verified travel data. The model reasons about this data; it does not invent it. This hybrid approach balances capability with operational simplicity. The fine-tuned base changes infrequently (quarterly at most). The prompts change constantly. The tools are maintained as APIs with their own deployment cycles. Each layer is optimized for what it does best, and none is asked to do what it does poorly. If you are building an AI [trip planning](/blog/multi-city-trip-planning-ai-shines) agent, I would suggest starting with tools and prompts. Get the product right. Then fine-tune only when you have identified a specific, measurable gap that prompting cannot close. That gap is almost always tone or format, not reasoning or facts. Fine-tuning is a precision instrument, not a default strategy. --- Nowah is an AI travel agent that searches and books real flights and hotels through conversation — no filters, no thirty open tabs. [Plan your next trip](https://app.nowah.xyz).