Construction is the world's largest industry by output and among its least digitized. McKinsey estimates that global construction output could hit $22 trillion in 2040. In order to reach this, it will have to meet a compound annual growth rate of 3.2% – a target that requires an industry-wide productivity transformation.
The gap between potential and reality is stark: McKinsey also found that 98% of megaprojects finish behind schedule and over budget. It’s also a naturally risk-averse industry, as when you’re pouring concrete or hanging steel, there are no “undo” buttons.
Intelligent technological adoption can help resolve some of these issues – and when you zoom out, it becomes clear that retail giants such as McDonald’s, 7-Eleven or Dollar General know this better than most. These are some of the largest builders in the world: for example, McDonald’s operates tens of thousands of locations globally. Supporting that kind of footprint is an army of project managers, coordinators, architects and vendors that ensure every store is built to a brand standard.
Building and renovating stores is a core skill, but not their core business. Their business is serving customers and growing the brand – but they cannot grow without building.
Notably, these companies already have some of the most structured data in the industry: playbooks and design manuals, brand standards and prototypes, checklists and approval flows, typical schedules and budgets, which create repeatable store types.
“If you compare a skyscraper to another skyscraper, there might be only 20% overlap in how they’re built,” says Alim Uderbekov, founder and CEO at Surfaice Corp. “But if you compare a McDonald’s or a 7-Eleven store to the next one, 80% of the process looks the same,”
“That’s why it’s far more practical to train AI to build the next McDonald’s than the next Empire State Building.”
There’s evidence the construction industry is already taking note. Use of AI in the construction market is growing rapidly regardless — from an estimated $11 billion in 2025 to a projected US$27.92 billion by 2031, a compound annual growth rate exceeding 16.6%, according to Mordor Intelligence. Surfaice is one player in an increasingly crowded field.
Uderbekov grew up in Kazakhstan, the son of a construction engineer and an architect.
"Growing up around construction sites, I was always struck by the contrast. I’d see incredibly advanced CAD software, and then I’d walk onto a jobsite with [my] dad where there was far less intelligence built into how work was coordinated. That difference confused me early on, and it stayed with me. I kept thinking how this can be improved,” said Uderbekov.
Aggregating data
While the team Surfaice focus on the retail developer's end-to-end workflow, nPlan is attacking the fundamental planning lie that bedevils every large capital project: the optimistic schedule. The company has assembled what it claims is the world's largest dataset of completed construction programs – more than 750,000 project schedules representing over US$2.5 trillion in capital spend. That archive is the foundation for machine learning models that forecast how individual schedule activities will actually perform, rather than how planners hoped they would.
The distinction matters because traditional schedule risk analysis relies on human estimators applying subjective probability ranges to activities –an approach that academic research has consistently shown to be infected by optimism bias. nPlan's AI-driven Schedule Risk Analysis (AI-SRA) instead learns from how comparable tasks have actually played out across thousands of similar projects, producing probability distributions that are, in the company's assessment, "unattainable by human operators".
The distinction matters because traditional schedule risk analysis relies on human estimators applying subjective probability ranges to activities — an approach that can be affected by optimism bias. Instead, nPlan's approach is to learn from how comparable tasks have actually played out across thousands of similar projects, producing probability distributions that are, in the company's assessment, "unattainable by human operators".
It’s clear that standard large language models (LLMs), to which users send a prompt and wait for a response, aren’t the primary route forward for AI in construction.
When you look closely at retail development, there are some basic demands:
- Context and memory across many months of a project
- Multi-step reasoning and planning
- Operation of third-party tools
- Deep reliance on a brand’s documented best practices and standards
- Coordination across many data sources
Uderbekov explains that he pushed his team to consider how to train AI on construction manuals, playbooks, and real projects of repeatable retail development. The idea was to build an agent that knows how to build a store.
“Most construction software treats the work as a series of isolated problems: one tool for scheduling; another for budget tracking; another for drawings; another for permitting; another for field surveys,” Uderbekov explains.
“Yes, that’s all very useful, but it’s still fundamentally horizontal. No software takes responsibility for the whole outcome. The real shift for me was realizing that AI shouldn’t optimize a task; it should be able to help build the entire store,”.
A project manager understands the entire project vertically: the sequence of tasks, dependencies, brand rules, city constraints, vendor reliability and the risk profile of a site.
This matters because in large retail programs, the engineering challenge grows exponentially with scale. A single store requires hundreds of milestones spanning design, surveying, permitting, procurement, inspections and turnover. Multiply that across 100 stores, and the workload quickly becomes more than 10,000 interdependent tasks, each one a potential source of delay.
Another company in the space is Buildots, which provides customers with 360-degree cameras mounted on their hard hats during routine site walks. These images are then processed by AI models that automatically compare physical site conditions against the BIM model and project schedule.
The commercial results reported by clients are notable. UK contractor Sir Robert McAlpine deployed the platform across more than 260,000 square meters of live projects, using it not just for progress tracking but for subcontractor billing verification and quality audit trails. Intel is now using Buildots to support its global fab expansion, with the company claiming the platform has helped Intel avoid roughly four weeks of construction delays per fabrication facility.
Computer vision models also have clear potential in construction. Cranes are among the most capital-intensive assets on any vertical construction site, and their utilization is notoriously difficult to measure.
Turner Construction deployed CraneView, an Internet of Things (IoT) based tool by construction tech firm Versatile, on two tower cranes at its Manchester Pacific Gateway project in San Diego. The firm was able to use the tool to analyze each crane lift and classify them using machine learning, to build a log of idle periods and surface patterns that superintendents can act on in real time.
Analysis revealed when cranes were sitting idle or handling uneven load distributions, which allowed the project team to rebalance crew schedules and crane assignments. The result was the early demobilization of one of the two tower cranes, generating both time and cost savings that tower crane rental rates – which can run to tens of thousands of dollars per month.
Versatile's approach reflects a growing recognition that the construction site's data problem is not just about integrating existing software systems — it is about creating instrumentation where none existed before. Cranes, excavators, and scaffolding systems that were previously invisible to any digital layer are becoming active data sources, producing the raw material for the next generation of construction AI.
Quantifying results
To quantify the effects of digital transformation on construction sites, Surfaice’s engineers have been evaluating impact using a “total attributable value” model with three components. The first is direct cost reduction, where workflow automation and earlier risk detection have cut delivery costs by as much as 30% in conservative scenarios.
The second component is error-related cost reduction. In many retail projects, design and coordination mistakes can account for 10–15% of total cost; by enforcing guardrails and maintaining a unified source of truth, agentic AI systems have reduced error rates to 2–3%, significantly lowering rework. The third component is redeployed labor value. When AI handles routine administrative work, organizations can redirect roughly 10% of monthly project-management capacity toward higher-leverage engineering and coordination tasks.
As AI evolved, firms must evolve with it. Today, Surfaice's core architecture is built around autonomous agents running on top of state-of-the-art models (including Google’s Agents Development Kit and Gemini family), but the design is model-agnostic by intent. Companies like OpenAI, Anthropic, and Google will keep making their models better. What if every new version of GPT, Claude, or Gemini can make the product better?
Uderbekov added, “No customer has ever asked, ‘Are you using Model X or Model Y?’ They asked this question: ‘Can your agent do this process better and faster than our current way?’ That’s the bar.”
