AI in construction: what works on site today, and what is still smoke
Artificial intelligence applied to construction spans everything from voice transcription to delay prediction. Some of those applications already work well and can be used on a real project today. Others work under controlled conditions and fail in the field. And several are sold with confidence and still do not exist reliably. This article separates the three categories: the value lies both in what works and in what is better not to buy yet.
What already works
Voice transcription
The most mature application and the one with the most immediate impact on site, because it attacks the exact point where the chain breaks: capture.
A site engineer dictates what he sees and the system turns it into text. It works well even with moderate background noise.
Its real limit is not general technology but vocabulary. A generic model trained on neutral Spanish butchers construction jargon: cimbra, encofrado, formaleta, castillo, dala, cadena, colado, vaciado, fundida, varilla, fierro. And those are precisely the words that carry the information. It requires domain vocabulary adapted per country — specific work that does not come free with the model.
Data extraction from documents
Reading a budget, an item catalog or an invoice and turning them into structured data. It works well with reasonably tidy documents and saves days of manual entry.
It fails with poor-quality scans, irregular table structures and very unconventional formats. It always requires review.
Visual detection of elements
Identifying in a photo whether workers wear helmets, whether the front is tidy, whether the scaffold has its protections. It works with acceptable precision for narrow, well-defined tasks.
It serves safety monitoring and photo pre-filtering. It does not serve to verify construction quality, which is a far more complex judgment.
Search and summarization over your own information
Asking in natural language about what happened on the project and getting an answer with sources. This one works well and is underrated: a good share of the time lost on site goes to hunting for information that exists but nobody can find.
What works halfway
Quantity measurement from photos
Estimating how many square meters of wall went up from a photo. It works under controlled conditions: good lighting, proper angle, visible scale reference.
On a real site, with sun in the lens, obstructions and improvised angles, the error margin is too high to feed a progress billing. It serves as an approximate check, not as measurement.
Delay prediction
The models exist and the signals they use are reasonable: staffing drops, productivity below budget, accumulating constraints.
The problem is that they demand quality history, and most builders do not have it. A predictive model fed with month-end memory-reconstructed data produces predictions with the quality of those data. The right sequence is to measure well first and predict later, not the other way around.
Report generation
Turning structured data into written reports. It works well for the prose, and the numbers need watching: a language model can impeccably write a wrong figure. The numbers must come from a calculation, not from the generation.
What is still smoke
Automatic construction scheduling. The idea of a system that builds the optimal schedule sounds great and collides with reality: a project’s sequence depends on local constraints, subcontractor availability, supervision criteria and decisions that are in no dataset. Current attempts produce schedules no experienced planner would approve.
Automatic detection of construction defects. Distinguishing a structural crack from a shrinkage crack, or judging whether a pour came out right, requires technical judgment and information a photo does not contain. What exists today detects the presence of a crack, not its meaning.
Automatic price estimation. Suggesting unit prices from historical data sounds reasonable until you consider how much a price depends on specific conditions: access, height, local material availability, crew capability. It serves as an initial reference; presenting it as a budget is dangerous.
Anything promising to replace technical judgment. A site engineer decides with incomplete information under time pressure, integrating factors nobody wrote down. That is exactly the kind of work AI does worst.
The evaluation criteria
Three questions that work against any proposal.
What happens when it is wrong? If the error is visible and cheap to fix — a transcription the user reviews — the risk is low. If the error is invisible and expensive — a mismeasured quantity entering a billing — the risk is high and mandatory human verification is required.
What data was it trained on? A model trained on US construction will fail with confined masonry, Latin American vocabulary and the region’s unit-price mechanics.
What does a human do in the loop? Every serious application of AI on site has a human confirmation point. If it does not, either the risk is very low or somebody did not think it through.
Where to start
If you are bringing AI into your operation, the sequence that makes sense is this:
- Capture, with voice transcription. That is where the real pain is and where the return is immediate.
- Structuring, turning that capture into data by concept and location.
- Analysis, comparing executed against budgeted.
- Prediction, only once you have six months of good data.
Most companies want to start at the fourth step because it is the one that sounds impressive. Without the first three, it is a model predicting over invented data.
Frequently asked questions
- Will AI replace the site engineer?
- Not in any foreseeable horizon. What it can replace is the administrative share of the job, which today eats about a third of the workday. The expected outcome is not fewer engineers: it is engineers doing more technical supervision and less data entry.
- How reliable is transcription with construction noise?
- With moderate noise and adapted vocabulary, quite reliable. With intense noise — a concrete pour, a jackhammer nearby — it drops noticeably. That is why the confirmation step is not optional.
- Do I need connectivity to use AI on site?
- For processing, yes, because the models run on servers. But capture can happen without signal if the channel queues and uploads when there is network. The difference matters: the engineer should not have to wait for signal to record.
- How much does implementing AI on a project cost?
- It depends on scope, and the license cost is usually the smallest component. What really costs is preparing the base data — item catalog, budget — and sustaining the capture habit. Any proposal that omits that is underestimating the project.
- How do I know if a vendor is exaggerating?
- Ask them to show you an error. An honest vendor knows where their system fails and explains it without hesitation. Whoever claims it always works has either never deployed on a real site or is not being frank.
Chilean, designing for Latin America. Field research surfaces what actually matters to clients, and that becomes products non-technical people adopt on their own — legal, education, accounting — and that show up in productivity from week one.
See all their articles →
