# Fine-Tuning > Fine-tuning further trains an existing AI model on a specific set of examples so it learns a particular style, format, or task. A construction firm might use it to teach a model its conventions for classifying subcontractor scope items. - Stages: Estimating, Preconstruction - Concepts: LLMs, Data - Published: 2026-08-28 - Canonical: https://buildr.com/library/fine-tuning ## Why it matters in construction GCs often hear "fine-tuned on construction data" in sales pitches. A precon team should ask what the model was trained to do, how many examples it used, and how that differs from giving the model the firm's documents at question time. Fine-tuning changes how a model behaves more than it changes what it can look up about a firm's projects. It can help the model map "ACT" to the right cost code or follow the firm's proposal style. It is not a reliable way to retrieve the backlog or a past bid's unit price. ## How it works 1. **Start with a base model.** A general-purpose model already understands language and some construction vocabulary. 2. **Build a training set.** Collect examples of the input and desired output. For scope classification, that might be 5,000 proposal line items labeled by estimators with the right cost code and inclusion or exclusion flag. 3. **Train.** Training adjusts the model's internal weights so it better reproduces patterns in the examples. It can take hours or days and requires computing resources. 4. **Evaluate.** Hold back part of the labeled set and measure accuracy on examples the model did not see during training. Compare it with a well-written prompt. 5. **Maintain.** If the base model changes or the firm changes its cost codes, the fine-tune may need to be updated. Prompting is flexible but may be less consistent. Retrieval supplies current facts at question time. Fine-tuning can produce more consistent behavior on a narrow task, but it is slower to update. Products may combine all three. ## Example in practice For example, a software vendor serving commercial GCs finds that its prompt-based classifier maps sub proposal line items to a standard cost code structure with 91 percent accuracy. The remaining 9 percent is concentrated in the trades estimators argue about anyway: whether "cutting and patching" belongs to the sub or general conditions, where "temporary protection" lands, how to treat combined mechanical and plumbing proposals. The vendor collects 12,000 estimator-corrected examples from across its customer base, fine-tunes a model on them, and reaches 97 percent on a held-out test set. For a GC processing 150 proposals a month with roughly 40 line items each, that reduces the number of miscategorized items an estimator needs to review. The vendor, rather than the GC, manages the training process. ## Go deeper - [GPT Models for Contractors: A Plain-English Guide](/blog/gpt-models-for-contractors.md) - [Best AI Preconstruction Software for GCs 2026](/blog/best-ai-preconstruction-software.md) ## Related terms - [Retrieval-Augmented Generation (RAG)](/library/retrieval-augmented-generation.md): Retrieval-Augmented Generation (RAG) finds relevant pages in a company's own documents and gives them to an AI language model before it answers. It grounds a construction team's responses in its actual specs, proposals, and project history instead of the model's general knowledge. - [Prompt Engineering](/library/prompt-engineering.md): Prompt engineering is the practice of writing instructions, examples, and context for an AI language model so it produces reliable output, such as telling it exactly how to classify exclusions in a construction subcontractor proposal. - [Data Readiness](/library/data-readiness.md): Data readiness is how complete, consistent, and accessible a construction company's project, pipeline, cost, and staffing records are. It determines whether AI tools can produce useful forecasts and comparisons or confident answers built on gaps. - [Large Language Model (LLM)](/library/large-language-model.md): A large language model (LLM) is an AI system trained on large amounts of text that can read, summarize, and generate language. In construction, it can help parse RFPs, sub proposals, and specifications for preconstruction teams. - [AI Evals](/library/ai-evals.md): AI evals are structured tests that measure how well an AI system performs on a defined task using examples with known answers. A construction firm can use them to test whether a tool reads its sub proposals or RFPs accurately before relying on it on bid day. ## Referenced by - [AI Data Privacy (Training on Your Data)](/library/ai-data-privacy.md): AI data privacy in construction is about whether a vendor or model provider uses your estimates, sub pricing, and project data to train models used by other companies, and which contractual and technical controls prevent that use. ## FAQ ### Should a GC fine-tune its own model? Most GCs will not fine-tune a model themselves. It requires many clean, labeled examples and ongoing maintenance. In many cases, retrieval and well-designed prompts over the firm's documents are more practical. ### When does fine-tuning make sense? When the task is narrow, high-volume, and stable, and prompting alone cannot hit the accuracy target. Classifying line items into a fixed cost code structure across tens of thousands of proposals is a reasonable case for a vendor to fine-tune. ### Does fine-tuning teach the model our project data? It teaches patterns and behavior, not facts to look up. A fine-tuned model will not reliably recall a specific unit price from a past bid. Retrieval is the right tool for facts; fine-tuning is for style and consistent judgment.