The construction industry is not short of data. Decades of project files, defect logs, cost records, survey reports, maintenance histories. The problem is that almost none of it was collected with AI training in mind. And feeding decades of inconsistent, incomplete, and biased records into an AI model does not produce smart outputs. It produces confident wrong ones.
This is the garbage-in-governance-out problem. And it is the hidden risk sitting underneath almost every AI deployment in construction today.
Why construction data is structurally compromised
Think about how construction data gets created. A project manager fills in a delay log. They record the delay. They rarely record all the contributing factors, because that takes time and may create uncomfortable conversations. A surveyor produces a condition report. They flag what they can see. They do not flag what they did not look at, or what was in a locked room, or what a previous surveyor told them informally was not worth investigating.
Those gaps compound over time. And they create a specific problem called survivorship bias.
Survivorship bias means your dataset contains only the buildings and projects that made it into the record-keeping system. Buildings that were demolished, projects that were abandoned, companies that went bust, all take their data with them. The classic example is Second World War bomber damage analysis. Engineers analysed planes that returned and armoured the parts that had been hit. The planes that did not return had been hit elsewhere. The data only showed survivors.
In construction, this means an AI trained on historical project records will learn from completed, documented projects. It will have little to no data on the failures, the cover-ups, the informal decisions that were never written down. When it predicts risk on a new project, it is drawing on an incomplete picture of reality.
The Data Quality Problem
60% : of AI projects abandoned due to poor data quality, per Gartner
69% : of companies say poor data blocks reliable AI decisions (Huble survey)
49% : of executives cite data inaccuracies as a barrier to AI adoption (IBM IBV)
Biased datasets, biased outputs
The survivorship problem is compounded by demographic and geographic bias in construction data. Most well-documented construction records come from larger firms, public sector projects, and urban developments. Rural construction, small contractors, heritage buildings, and certain building typologies are systematically underrepresented.
If an AI tool for defect prediction was trained primarily on London commercial developments, it will perform poorly on a Victorian terrace conversion in the East Midlands. The model does not know what it does not know. It will still produce an answer. That answer will look authoritative. But its confidence has not been earned.
A 2024 analysis of AI bias in construction found that datasets are routinely biased by historical construction practices, incomplete demographic representation, and inconsistent measurement methods across sites and eras. The result is systematic errors in risk assessment, cost estimation, and defect prediction.
"An AI does not know what it does not know. It will still produce an answer. That answer will look authoritative. But its confidence has not been earned."
The governance framework you need
The UK government's DSIT and the Open Data Institute have both published guidance on AI-ready data in 2025 and 2026. The DSIT four-pillar framework covers technical optimisation, data and metadata quality, organisational governance, and legal and ethical compliance. It was written for public sector datasets but the questions it asks are just as relevant for a construction firm trying to evaluate whether its project history is suitable for AI training.
For built environment firms, auditing AI training data means asking vendors specific questions. What projects were included in the training set? What geographies, building types, and project sizes? What years? What was the data collection methodology? Are the defect logs based on intrusive inspection or visual only? Were failed projects included?
The construction industry is not short of data. Decades of project files, defect logs, cost records, survey reports, maintenance histories. The problem is that almost none of it was collected with AI training in mind. And feeding decades of inconsistent, incomplete, and biased records into an AI model does not produce smart outputs. It produces confident wrong ones.
This is the garbage-in-governance-out problem. And it is the hidden risk sitting underneath almost every AI deployment in construction today.
Why construction data is structurally compromised
Think about how construction data gets created. A project manager fills in a delay log. They record the delay. They rarely record all the contributing factors, because that takes time and may create uncomfortable conversations. A surveyor produces a condition report. They flag what they can see. They do not flag what they did not look at, or what was in a locked room, or what a previous surveyor told them informally was not worth investigating.
Those gaps compound over time. And they create a specific problem called survivorship bias.
Survivorship bias means your dataset contains only the buildings and projects that made it into the record-keeping system. Buildings that were demolished, projects that were abandoned, companies that went bust, all take their data with them. The classic example is Second World War bomber damage analysis. Engineers analysed planes that returned and armoured the parts that had been hit. The planes that did not return had been hit elsewhere. The data only showed survivors.
AI in construction, this means an AI trained on historical project records will learn from completed, documented projects. It will have little to no data on the failures, the cover-ups, the informal decisions that were never written down. When it predicts risk on a new project, it is drawing on an incomplete picture of reality.
The Data Quality Problem:
60% of AI projects abandoned due to poor data quality, per Gartner
69% of companies say poor data blocks reliable AI decisions (Huble survey)
49% of executives cite data inaccuracies as a barrier to AI adoption (IBM IBV)
Biased datasets, biased outputs
The survivorship problem is compounded by demographic and geographic bias in construction data. Most well-documented construction records come from larger firms, public sector projects, and urban developments. Rural construction, small contractors, heritage buildings, and certain building typologies are systematically underrepresented.
If an AI tool for defect prediction was trained primarily on London commercial developments, it will perform poorly on a Victorian terrace conversion in the East Midlands. The model does not know what it does not know. It will still produce an answer. That answer will look authoritative. But its confidence has not been earned.
A 2024 analysis of AI bias in construction found that datasets are routinely biased by historical construction practices, incomplete demographic representation, and inconsistent measurement methods across sites and eras. The result is systematic errors in risk assessment, cost estimation, and defect prediction.
"An AI does not know what it does not know. It will still produce an answer. That answer will look authoritative. But its confidence has not been earned."
The governance framework you need
The UK government's DSIT and the Open Data Institute have both published guidance on AI-ready data in 2025 and 2026. The DSIT four-pillar framework covers technical optimisation, data and metadata quality, organisational governance, and legal and ethical compliance. It was written for public sector datasets but the questions it asks are just as relevant for a construction firm trying to evaluate whether its project history is suitable for AI training.
For built environment firms, auditing AI training data means asking vendors specific questions. What projects were included in the training set? What geographies, building types, and project sizes? What years? What was the data collection methodology? Are the defect logs based on intrusive inspection or visual only? Were failed projects included?
