An independent magazine about AI at work in construction. Every story starts on a real job site, with the people who used the technology and the results they will answer for.
GRAHAM Ties Its Drone Maps to Survey Control Before the First Flight
GRAHAM Group UK is tying drone maps to survey control before the first flight. A vendor clip on X shows the procedure. Neither company has published a measured result.
On 31 August DroneDeploy posted on X that GRAHAM Group UK, a contractor with about 2,200 employees and 19 offices across the UK and Ireland by its own count, is setting ground control points on a live job and tagging them in DroneDeploy's software before the first mapping flight. The post's own summary carries the whole argument: two drone flights with nothing tying the maps to survey give you two maps that do not line up, and ground control fixes that once, at the start of a job. The vendor also hosts the clip on its Insider channel under the title Setting and tagging ground control points with the Graham Group, described there as the way every map lands on the same coordinates and cut and fill volumes stay comparable across flights.
That is a vendor post about a vendor's customer, and this article treats it as one. The post does not identify the jobsite. Neither company has published a measured result from a GRAHAM job, in volume, hours, or error, as of 2 September. The sequence of work is worth understanding before a contractor buys a drone programme on the promise that its maps will agree.
"Two drone flights with nothing tying the maps to survey give you two maps that do not line up." DroneDeploy, X post, 31 August 2026.
Why two flights disagree
A drone map is built from hundreds or thousands of overlapping photographs, each stamped with the position the drone's own satellite receiver reported when the shutter fired. That receiver is good to a few metres, and its error on Monday differs from its error on Friday. Photogrammetry software fits the photographs to each other very well, and fits the whole assembly to the ground only as well as those stamps allow. The result is a map that is internally consistent and floats. Fly the same site twice and the two maps float by different amounts, and the difference shows up as a change in ground level that no excavator made.
A ground control point is a target on the ground whose position a surveyor has measured in the project's coordinate system before the flight. Give the software a handful of those known positions and the photographs they appear in, and it pins the floating assembly to the ground at those points. Every later flight pinned to the same targets lands on the same grid, and the difference between two maps becomes the difference in the dirt. The post's phrase, once at the start of a job, describes that arrangement: the surveyor works up front, and the targets stay put for as long as the earthworks run.
31 August 2026. Post text, mention of GRAHAM Group UK, link to the civil construction page, and the attached clip. The article embeds the post rather than recapping frames; audio was not transcribed.
Undated video page, read 2 September 2026. Title and description: every map lands on the same coordinates and cut and fill volumes stay comparable across flights.
GRAHAM Ties Its Drone Maps to Survey Control Before the First Flight | ConstructionMagazine.ai
DroneDeploy made the failure case explicit in a LinkedIn post on 4 August, before the GRAHAM clip. In the old way it described, a crew flies Monday morning, cut and fill comes back that afternoon, quantities go to the estimator, and a number lands in a pay application by Thursday. Two weeks later a survey crew checks one section, the elevation is off, and every quantity built on that map is in question. The company's answer in that post is an in-app warning when a map or measurement falls outside an accuracy tolerance, so that a datum mismatch or a bad control point surfaces before the section is built and billed. That warning is a vendor feature described by the vendor. The embed above is the step before it.
How the points get into the map
DroneDeploy's help centre lays out the software side of the workflow. The surveyed coordinates go up as a spreadsheet alongside the images, with a label column that can mark some points as checkpoints, which test the fit rather than force it. After a pre-processing pass, a feature the company calls Ground Control AI looks for standard checkerboard targets in the photographs and places a crosshair on each one. The company says it recommends at least seven tagged photographs per point, and that most maps finish that automated step in under 90 minutes, with maps over 2,500 images taking longer. Point files can be saved at project level and reused across later uploads, which is the software's version of the once-at-the-start idea.
The human step is the one the clip is there to show: a person opens each thumbnail and confirms, or nudges, the crosshair onto the exact centre of the painted target before the map processes. The help centre says the automated tagger works on high-contrast checkerboard targets and that other target styles fall back to manual tagging. The help pages sat behind a bot check on 2 September, so the figures above come from indexed excerpts of them.
The ground control workflow as DroneDeploy's help centre and the GRAHAM post describe it, read 2 September 2026
1Surveyor sets targets on site and records each position in the project coordinate system
→
2Crew flies the site
→
3Images upload with a spreadsheet of point coordinates, some marked as checkpoints
→
4Software pre-processes and auto-tags checkerboard targets
→
5Reviewer confirms the crosshair on the centre of each target
→
6Map processes against the confirmed points
→
7Later flights reuse the saved project points
What the vendor claims, and who will own it
DroneDeploy's civil construction page, the one the X post links to, carries the company's own accounting for what accurate mapping is worth: a claimed $10,000 saved for every $1 million spent on construction, split into roughly $4,000 from documentation and disputes, $3,000 from remote visibility, and $3,000 from schedule and planning. The page gives no method and no customer behind those averages, and they should be read as marketing arithmetic. The same page says the software is used on 3 million sites in 180 countries and promises alerts when maps and measurements miss accuracy tolerances. One customer quote on the page, from a construction technology manager identified only as Scott Brown, puts the sales case in a sentence: trade partners argue counting trucks; we're counting pixels. The page gives neither his employer nor a date.
DroneDeploy's claimed average saving per $1 million of construction spend, US dollars, vendor solutions page, read 2 September 2026
The vendor is about to change hands. On 29 July Procore announced a definitive agreement to buy DroneDeploy for approximately $845 million in cash, with closing expected later this year subject to regulatory approval. Procore's chief executive, Ajei Gopal, said in the release that the plan is to cross-sell DroneDeploy into Procore's customer base as soon as the deal closes and, over the longer term, to fold its capture into digital coworkers that track a project and act on it, to paraphrase his statement. A contractor evaluating the GRAHAM workflow today is evaluating a product whose roadmap and pricing a different company will set within months.
GRAHAM's relationship with the vendor predates the 31 August post. On 12 August DroneDeploy said on LinkedIn that it had been shortlisted for the GRAHAM Supply Chain Awards 2026, that the entry rested on work across live GRAHAM projects, and that the ceremony takes place at the Crowne Plaza in Belfast on 12 November. A supplier shortlisted for a customer's award and a customer in the supplier's marketing clip are one relationship seen from two sides. Neither post says how many GRAHAM jobs fly drones, how often, or what the company has measured.
What to ask before the first flight
Ask who sets the control. The people on the ground with the pole and the receiver are the survey function, and the drone is a second instrument working off their marks. A drone programme that arrives without a surveyor's time budgeted for control is the two-maps problem waiting to happen. Ask how many points, where, and how they will survive the job. A painted cross on a haul road lasts until the next load of stone.
Ask what the checkpoints say. Control points force the fit and checkpoints test it. A vendor report that shows the map hitting its own control points proves little, since those are the points it was told to hit. The residual at the checkpoints tells you whether a cut and fill comparison between two flights means anything, and the help centre confirms the software lets a user mark points as checkpoints. Ask for that residual on every flight, in the project units, before anyone puts a volume in a pay application.
Ask what the vendor's tolerance alert is set to and who sees it. DroneDeploy's 4 August post describes a warning when a map falls outside tolerance. A warning that goes to the pilot's inbox and never reaches the quantity surveyor is a warning that gets billed against. Ask who at GRAHAM sees it, if you can get GRAHAM on the phone, because the post does not say. Then ask the vendor for the one thing neither company has published about this work: a before-and-after on a volume dispute, with dates, on a site they will identify. Until that exists, the embed is a demonstration of the procedure as the vendor presents it, and nobody has claimed more for it.
· DroneDeploy
Undated vendor page, read 2 September 2026. $10,000 per $1 million claim and its split, 3 million sites in 180 countries, accuracy alerts, GIS help, Scott Brown quote. All vendor figures.
Undated. Ground Control AI, checkerboard targets, at least seven tags per point, under 90 minutes for most maps, over 2,500 images take longer. Read through indexed excerpts on 2 September 2026; the page sat behind a bot check.
Undated. CSV upload, checkpoint label, project-level point files reused across uploads. Read through indexed excerpts on 2 September 2026; the page sat behind a bot check.
Undated company page, read 2 September 2026. 2,200 employees, 19 offices across the UK and Ireland, construction and civil engineering among its divisions. Company's own numbers.
29 July 2026. $845 million cash, close expected later this year subject to regulatory approval, Ajei Gopal statement on cross-sell and digital coworkers.