Does your AI need construction terminology training?

How an open question becomes completed work

A site office employee writes: „TBC: scope of work at the entrance. Question sent to the subcontractor, waiting for a reply”. „TBC” means „to be clarified”. AI files the case under „completed” and summarises it as „The scope of work at the entrance has been clarified”. Someone who knows the notes sees the error at once, because sending a question does not close the case. They have to go back to the note and correct the report before the project manager relies on it.

One such correction takes a few minutes. At several hundred notes a month, checking a finished summary starts to take longer than preparing it by hand. So the first question is: where does the mistake come from.

What to fix before fine-tuning

A glossary is enough when staff use their markings consistently. With very regular notes a plain rule is enough: a note starting with „TBC” never goes into the completed category. Not every note needs free interpretation by a language model.

When two teams use the same letters differently, the model needs to know which project the text concerns. Training one shared meaning would then lock in the error for the second team. Check the reading of photos and documents separately: if the sheet says „not closed” and AI receives only „closed”, the problem arose before the sentence was interpreted, and no training will fix it.

What fine-tuning for construction terminology can teach

Fine-tuning teaches the model a recurring distinction, for example a case waiting for a reply against a closed one, also when note authors use different words. Google describes tuning models on labelled examples and recommends refining the prompt first. Training does not replace access to current records: a model trained to read „waiting for a reply” still needs the next message to know that the reply has arrived.

In this scenario, the glossary gives most of the effect for a fraction of the cost. Fine-tuning adds 1.8 hours a month, which justifies training only at a larger scale or when errors are expensive for the project, for example when a missed question to a subcontractor delays acceptance.

Will the improvement matter to the office?

Give the assessment to someone who knows the language of the notes and uses the reports. A nicer-written report with the same category error saves them no time. Also decide who updates the glossary after rollout when a new abbreviation appears, because that affects the cost of upkeep. A similar problem with fixed categories is covered in the post about fine-tuning for service ticket classification.

We build an app on site notes as part of our AI apps, from €3,000 net, and model adaptation as part of our private LLMs. For the free analysis, a description of the report built from the notes and of the mistake that comes back most often is enough.