UK Construction's AI Problem Starts Before the Estimate

In the 12 months to November 2025, 3,950 construction companies in England and Wales became insolvent, according to the Insolvency Service. That is more than in almost any other sector. Specialist trades account for the largest share, with subcontractors doing electrical, demolition and finishing work. Margins are thin across the industry, and the pressure to price jobs quickly and accurately leaves little room for error.

Artificial intelligence was supposed to help, and the industry largely believes it will. In a 2025 global survey of more than 2,200 construction professionals by the Royal Institution of Chartered Surveyors (RICS), nearly 70% of project managers and quantity surveyors said AI would help them deliver greater value. Actual use tells a different story: 45% of organizations reported no AI use at all, and just 1% had scaled it across projects. Poor data quality was among the barriers respondents cited most often.

In estimating, poor data quality usually starts with measurement. Many of the AI tools now marketed to builders promise fast, priced estimates. But an estimate is only as good as the quantities underneath it, and pulling those quantities from drawings is easy to get wrong and easy to skip.

Prakhar Gupta, co-founder and CEO at Mergro AI, works on that step. Mergro is one of a number of companies building software for takeoffs, which extract measurements from floor plans, tender packs, and structural drawings. These companies also work on quantification, the bill of quantities a contractor needs to price and execute a build.

Guessing First, Measuring Later

The measurement problem starts with how small builders win work. For small residential builders in the UK, estimating has long been closer to improvisation than science. Many contractors price jobs from a handful of photos and general experience rather than detailed measurement.

Speed is the reason. By Gupta’s estimate, more than 60% of jobs go not to the best estimate but to the first decent one. “The speed matters and then the quality of what you write in matters,” he says.

That incentive pushes shortcuts into the measurement stage, and the cost shows up later as waste. Government estimates have long put the share of solid materials delivered to UK construction sites that goes unused at about 13%, with over-ordering among the causes. Gupta says the builders he talks to often consider over-ordering by 20 to 25% acceptable. “Go near a building site, go near their skips where they keep the rubbish,” he says. “You will see brand new things in that skip.”

The Measurement Gap in AI Estimating

The hope was that AI would bring discipline to this process. The difficulty is that measuring from construction drawings is a visual task, not a language task. A system has to recognize doors, windows, walls and rooms from symbols and lines, then measure them correctly.

General-purpose AI models remain weak at exactly this. A 2026 academic benchmark tested leading multimodal models on architectural and engineering drawings. The models handled text extraction well, with accuracy as high as 95%. Reliably counting doors and windows on floor plans remained unsolved, with accuracy often between 40% and 55%.

Gupta sees the same failures in practice. Symbols get misread, as when a toilet is taken for a door because of its curved icon. Rooms are missed entirely, and accuracy degrades as tender packs grow longer. He says some products sold as estimating tools produce priced figures without a measurement step the user can inspect. That leaves no way to check where a number came from.

Takeoff software that uses image recognition has existed for years, and a growing number of companies are adding AI to it. The question for buyers is the same whichever product they look at: can a tool’s quantities be traced back to the drawing?

The talent market makes that gap harder to close. Lack of skilled personnel was the most commonly cited barrier to AI adoption in the RICS survey, at 46%. Gupta sees the shortage from two directions. Technology companies entering construction often lack the domain knowledge to understand the problem, while construction companies building their own tools tend to hire full-stack developers or generalist AI engineers rather than computer vision specialists. The specialists who could close the gap are also in demand from AI labs and major tech companies. “You’re not going to get an A-player who could have worked in some incredible thing, can go into a field and work for Google for an insane salary, now comes to construction,” Gupta says. In practice, much of the responsibility for checking AI output falls back on contractors and surveyors themselves.

Margins Under Pressure at Every Scale

Measurement errors cost companies differently depending on their size, but they cost all of them. For small contractors, thin margins and volatile procurement costs leave almost no room for error; diesel prices alone ripple through material pricing.

For larger contractors, inaccurate takeoffs lead to contract variations, which bring delays and cost overruns. Gupta describes cases where subcontractors invoice for work already paid and main contractors approve the invoices anyway. Across hundreds of subcontractor relationships, there is no single source of truth to catch the error.

The quantity surveyors who sit between those parties have been the slowest to commit. According to Gupta, some surveying practices have spent 18 to 20 months evaluating AI products without finding one reliable enough to build into their workflow. That caution matches the RICS findings: trial use is common, but scaled use across projects is rare.

Inputs Before Outputs

The common thread is a sequencing problem. Much of the market is selling the end product, a full estimate or an automated price, before confirming that the inputs underneath are sound. Gupta’s advice to contractors is to start with an internal check rather than a product evaluation. “Go to your team and ask them, do they have this,” he says. “They realize very quickly they don’t even have the quantities. They don’t have the measurement.”

The principle holds for any AI tool, whatever it does. A model built on unreliable data inherits that data’s errors and delivers them faster. “The industry needs to really go backward a little bit before they can start moving forward,” Gupta says.

For a sector that lost nearly 4,000 companies in a single year, the value of AI in estimating will depend less on how quickly a tool produces a number than on whether anyone can trace that number back to the drawing.

About the Expert: Prakhar Gupta is with Mergro AI, a company building software for construction takeoffs and quantification.

This article is based on information provided by the expert source cited above. It is intended for general informational purposes only and does not constitute legal, financial, or real estate advice. Readers should conduct their own research and consult qualified professionals before making any real estate or financial decisions.