Evidence
Surveys show AI use is growing. Cases describe reported improvements in specific tasks and assets. Neither can tell you what AI is worth to your organisation; that takes a test on your own work.
This summary draws on 301 public examples and 53 surveys, collected and reviewed by 38thMove.
The findings
Five recurring survey series report growing use. In RIBA’s 2026 survey, 74% of respondents say their practice uses AI, but only 19% report a documented AI policy. Trying AI is different from making it a governed part of everyday work.
Read the survey analysis →Reported benefits include faster work, lower energy use and earlier intervention. GWL reports preliminary 25–29% reductions in heating, ventilation and cooling energy; Yorkshire Water reports 15% fewer reactive site visits.
See the cases in more detail ↓A faster task only creates value if the output is usable once it has been checked and corrected, and the gain outweighs running costs. Many case accounts leave those costs out, so evidence of organisation-wide return is still thin.
How to build your own case →What the cases show
Leak analysis directs investigations, image inspection helps engineers prioritise checks and building controls respond to measured conditions. National Grid uses AI for a first pass at grading corrosion on pylons; Suffolk used it to target schedule recovery. The question is what changed once the information arrived.
Turner builds internal workflow apps, Mott MacDonald’s assistant opens up company knowledge and Arcadis engineers configure reusable technical checks. The sources say how many people use them. They mostly do not show whether the answers are reliable or the firm is more productive.
Engineers verify findings, planners keep statutory decisions and operators investigate alerts. Seattle’s test of automated permit checks agreed well on the checks it was set up for, but missed others. Speed and quality need separate tests.
Victoria’s audited bridge-monitoring venture cost public money without a single accepted installation. Northumbrian Water says AI helped prevent sewer overflows during specific storms, while its company-wide pollution figures worsened. A local gain does not guarantee a better organisation.
Four views of the database
Because every example is classified the same way, I can look across all 301 at once. These four charts are a taste of what the database can answer.
Where AI appears in the work
Inspection and condition assessment (54 cases) and asset operation and maintenance (53) lead the collection. About 7 in 10 cases use predictive AI, which spots, forecasts and optimises (for example image recognition, leak detection and scheduling); about 1 in 5 use generative AI, which writes and creates.
Open the full-size chart →Each of the 301 examples is counted once, under its main workflow. Hatched bars show where 38thMove judged the type of AI because the source did not say. The mix reflects what organisations publish, not how common each use is in the market.
Which assets and stages are covered
Operation and maintenance accounts for 112 cases, against 34 in design. Buildings and property (84), transport (63) and water (45) are the largest asset groups. Find the cells closest to your own work.
Open the full-size chart →Each example is counted once, by its main asset type and lifecycle stage where these are known. “Not assigned” covers company-wide and cross-sector work.
How the collection has grown
The collection has 42 examples first documented in 2024, 82 in 2025 and 87 so far in 2026. Generative AI went from a handful of examples in 2024 to over a third of those first documented in 2026.
Open the full-size chart →Each case is placed in the year it was first publicly documented; 19 undated examples are left out. This is the publication history of a curated collection, not an adoption rate. The type of AI is today’s description; some components may have been added later. 2026 is a part year.
Claimed time savings
There are 35 time-saving claims in 32 cases. Most claim savings of 50% or more on one task. Asset operations report other benefits, such as energy savings, less leakage and fewer reactive visits, which are not time measures and so are not on this chart.
These are claims, not results you should expect. No average is calculated, because tasks, baselines and checks differ.
Open the full-size chart →Each point is one claim about one task. Filled points are reported results, hollow points are estimates and arrows are open-ended claims such as “more than 75%”. Connected points show a range the source gave. Several claims can come from one case.
Workflows follow each case’s maintained primary classification. Claims without a usable task-time denominator, elapsed approval or service time, schedule recovery and algorithm-only runtimes are excluded. A row with no qualifying claim does not establish that the workflow has no benefit.
Your question
I can analyse the database around whatever would help you decide: the work you want to improve, an asset type, a stage of delivery, a technology, a place, a reported benefit or a practical constraint.
You get the relevant comparisons, the reported benefits and how far to trust them, the gaps, and what your own trial would need to prove. The analysis goes further than this site, with more detail and examples not published here. Where the public evidence cannot answer your question, I will tell you so.
Go deeper
The cases
Search by company, place, type of work or technology. Start with a hand-picked set of strong built-environment cases.
Browse the cases →The map
Every example on one map. Filter by type of work, asset, region or type of AI.
Explore the map →The surveys
Who was asked, what “using AI” meant, and what the main survey series found.
Read the survey analysis →Each example records what a source says, who said it, its scope and its limits. Every summary is checked against its source, but the outcome is not audited. Reported results, estimates, targets and research are kept separate. Several cases can share a supplier or source, so they are not independent tests.
The collection reflects what gets published, so it is not a measure of how widely AI is used. Survey results depend on who answered and what was asked, and different surveys are never averaged together.
Cases reviewed 12 October 2026; summary and charts updated 14 October 2026. Survey catalogue covers material published or current from 24 September 2025 to 24 September 2026.