How AI Construction Analytics Guides Land Decisions

An engineer looked at 70 acres in Whittier and estimated the site was short 40,000 cubic yards of fill. The scan data said 250,000. The grading contractor ran his own numbers and confirmed the scan was right.

AI construction analytics guides land decisions in one specific way: it shortens the distance between an assumption and the moment that assumption gets tested against dirt. The Whittier gap was 210,000 cubic yards between what the land decision was priced on and what the ground actually held. It surfaced during construction, which is an expensive place to find it.

This article covers how accurate that data is, what the analytics produce on a working site, which land decisions they should change, and how to close the loop across your own portfolio.

Why Does The Same Land Decision Get Made Twice?

Every parcel gets underwritten twice, and only one of those two events has a meeting. The first is at acquisition. Someone builds a pro forma from a topographic survey, a preliminary grading concept, and a set of assumptions about how the dirt will behave. Lot count, earthwork budget, schedule, and the price you are willing to pay all come out of that model.

The second is in the dirt, 18 months later, when the excavation crews find out what the site contains. Every yard of unplanned import, every day of schedule slip, and every design revision is the site pricing your assumptions a second time.

On most projects those two events never speak to each other. The second produces detailed evidence about how the first performed. That evidence goes into a project folder and stays there.

Whether any of that evidence can override a licensed engineer depends on how accurate it is.

How Accurate Is The Scan Data Behind An Earthwork Number?

TraceAir states that cut and fill volumes from repeat aerial scans measure to an accuracy of up to 1/10' (3 cm). What makes a number worth acting on, though, is independent confirmation.

On the Whittier project, the check that mattered was not the software. The scan result went to the grading contractor, who ran the quantities independently and confirmed them. That is the verification pattern worth insisting on with any data source: a second party with money on the line arriving at the same number.

Skyline Ranch shows the same pattern with a different instrument. The team there checked platform elevations against GPS rover measurements and found them, on average, within a tenth of a foot.

Brian Mangano, Director of Regional Development at Brookfield, went further on the method: "I am proposing using LiDAR for all our OG topography, regardless of project size, because it is significantly more accurate and faster than other methods."

Capture method matters too, and the two are not interchangeable. Photogrammetry reads the surface it can see. LiDAR, or light detection and ranging, returns ground through vegetation. Either one works on a cleared, actively graded site. On raw land under brush, only one of them is telling you about original ground.

Accuracy is the part that gets solved. The measurement still almost never reaches the people who set the price.

Why Does The Feedback Loop Break?

Three structural things break the loop. Effort is not one of them.

• As-builts arrive at the end. By the time a survey crew documents what was built, the land team has moved on to three other deals. The as-built survey becomes a closeout artifact instead of a planning input.
• The data is not comparable across sites. A topo from one engineer, a takeoff from another, field measurements from a third, and a different consultant on the twelfth project than on the first. Each carries its own shrinkage conventions and takeoff methodology, so nothing lines up and nobody tries to line it up.
• The knowledge is real but invisible. Cost history exists at most builders. It lives in an estimator's spreadsheet and in a VP of Construction's memory of the job that went sideways in 2023, and it reaches land meetings as an anecdote. No version of it exists that a committee can see, sort by market, and price against. The gap is institutional visibility, not institutional experience.

What Does AI Construction Analytics Produce For Land Decisions?

It produces four specific outputs: soil behavior tracking and prediction, automatic quantity detection, automatic progress tracking, and scan to scan comparison. Each one was built for execution. Underwriting is the second use.

Soil behavior tracking and prediction. Shrinkage and bulking are two separate corrections, applied at different points, and they carry more dollar risk than any other line in an earthwork estimate.

The geotech's lab work sets your starting assumption, and nothing airborne replaces it. Repeat scanning adds the other half: a measurement of what the material did once crews moved it, on this site, with this contractor and this handling.

TraceAir's soil behavior tracking and prediction gets more accurate with every newly added job site, so the correction factor available on your fortieth project is not the one you had on your first.

Automatic quantity detection. The platform automatically estimates and detects material quantities: stockpiled materials and SWPPP-related materials such as hydro-mulch, stabilizing fabric, and fiber rolls. Separate tools quantify and plan remedial grading, removals, and street over-excavations, and drone-based remedial as-builts document them.

Automatic progress tracking. Pavement percentage complete updates with every new drone survey, with nobody walking the site with a clipboard.

Scan to scan comparison. Two surfaces from two dates, differenced: volume moved, where it went, and whether it matches the plan.

Across a portfolio those four outputs add up to a record of how sites in your markets behave, captured consistently enough to compare one against another. Land decisions have never had that input.

Which Land Decisions Should Your Site Data Be Changing?

Site data should be changing four of them, starting at the earthwork line in the pro forma and ending at the design itself.

Whether The Dirt Actually Balances

Earthwork balance carries the largest dollar swing in a residential pro forma, and it usually sits on one line built from a preliminary concept.

After a Whittier the useful question is by how much, in which direction, and on what kind of site. One project only gives you a story. Across 20 projects the answer becomes an adjustment factor you can defend in front of a committee.

What A 10% Swing Costs On A Single Community

A soil behavior assumption that is 10% off can move seven figures on one community.

TraceAir puts the arithmetic plainly: 500,000 cubic yards of remedial grading moves the site balance by 50,000 cubic yards if the soil shrinks or bulks by even 10%. At an average of $25 per cubic yard to export, that costs a project over $1 million.

A soil behavior assumption 10% off does not show up as a technical variance. It shows up as the reason the deal you won was the deal you should have lost.

Century Communities saw the other side of that swing on its Glennwood project in Poulsbo, Washington. Cut and fill analysis surfaced both excess dirt and a dirt deficit at the area designated for the playground, early enough to use the excess to put the park up on a plateau. The project balanced itself, and no dirt needed to be imported or exported.

What Is Hiding Under The Canopy

Vegetation carries more acquisition risk than anything else on a raw parcel, because it is the part of the site nobody has measured.

On a 100-acre site in Temecula, dense brush prevented traditional methods from establishing accurate original ground topography. LiDAR returned the ground through the vegetation and captured it before clearing began. For a land decision the sequence matters: the information existed before the site was cleared.

An original ground surprise is usually something nobody could see, priced as though someone had.

When To Revise The Design Instead Of Absorbing The Cost

The decisions worth the most here are the ones the data surfaces early enough to still be cheap.

Skyline Ranch runs at a scale most divisions never touch: 500 acres and 27 million cubic yards of total cut. Early analysis showed the project heading toward roughly 600,000 cubic yards of excess dirt. That finding prompted design revisions. The project finished two months ahead of schedule with 20,000 cubic yards of long haul, under 0.1% of total cut.

The same thing happens at ordinary division scale. D.R. Horton's Memphis division caught a lot layout problem before staking and avoided roughly $50,000 in redesign, then confirmed from scan data that a planned set of retaining walls was unnecessary, taking $300,000 out of the job. Both decisions came earlier than they otherwise would have.

Does This Replace Your Geotech Or Your Civil Engineer?

No, and a land team should be suspicious of anything that claims it does.

The geotech's borings and lab work remain the source of your soil parameters. The civil engineer's design and takeoff are still the design and the takeoff. You still negotiate against the grading contractor's own quantities.

Repeat scanning adds a fourth, independent measurement. It arrives continuously instead of twice, and it compares across projects because every capture used the same method. In Memphis it functions as verification: billed work checked against what the scan shows actually happened in the field, rather than argued weeks later. It sits on top of the professional inputs.

Who Should Own The Site History You Build?

You should, and you should get that in writing before the first flight.

The payoff described here compounds over years. If you spend four years building a variance history that prices land better than your competitors can, that history is an asset, and you should know in advance who holds it.

Get three answers in writing: who owns the underlying scans and derived quantities, what format the historical record exports in, and what happens to your access if you stop renewing. Any vendor should answer those without flinching. A vendor who cannot is asking you to build your negotiating position on someone else's balance sheet.

How Do You Close The Loop On Your Own Portfolio?

Four steps will do it, and none of them requires a new department. The last one decides whether the other three survive.

Start With What You Already Have

Pull the earthwork estimate and the earthwork actual from your last five completed projects and look at the spread. That takes an afternoon and requires nothing new. If you cannot pull those two numbers in that time, that is the finding, and it is worth more than the number would have been.

Pick The Assumptions Worth Tracking Going Forward

Not everything. Shrinkage and bulking against assumption, earthwork quantity variance against the original estimate, and grading duration against schedule cover most of the dollar risk. Three numbers per project. Define the source of truth for "actual" before you start, because final quantities live across pay applications and change orders and sometimes end up negotiated rather than measured.

Capture Them The Same Way Every Time

Cross-project comparison fails on inconsistent measurement, not missing measurement. Pick one capture method, one cadence, and one definition per number, then keep them. Consistency matters more than precision here, because what you are looking for is a pattern.

Introduce It As a Pricing Tool, Not a Scorecard

This is the step that gets skipped, and it is why most variance tracking dies quietly. A history of estimate versus actual is also a performance record on whoever produced the estimates, sometimes including the person presenting it. Nobody volunteers that document. Frame it in market terms instead of project terms. Report a range and a direction, not a list of misses, and lead with what it lets you bid. After 10 or 15 projects it stops reading as a report and starts reading as a negotiating position: you know what a hillside parcel in your market costs to grade, and the seller does not.

What Should You Ask Before Your Next Acquisition?

There are seven worth asking. Anyone can answer the first three today, and the rest get better as the loop closes.

• If this site comes in 20% off our earthwork estimate, does the deal still work, and at what lot count does it stop working?
• Is there a receiving site within economic haul distance, and what does it charge today?
• Is the geotech scope adequate for this parcel, or are we pricing 80 acres of variable ground off six borings?
• What does our grading schedule assumption miss on average, and what does that add in carry cost here?
• What was our earthwork quantity variance on the last five projects in this market, and in which direction?
• What shrinkage and bulking assumptions did we carry, and what did those sites actually deliver?
• How much of this parcel can we not currently see, and what did unseen ground cost us the last three times?

What To Do Next

The gap between a 40,000 cubic yard estimate and a 250,000 cubic yard reality is not an argument for better estimating software. It is an argument for a shorter distance between what you assumed and what you found out, and for that distance being visible to the people setting the price.

Start with the first question on the list. Run the 20% sensitivity on the deal currently on your desk. If the answer is uncomfortable, the loop is worth building.

If you want to see what site data can tell you before a parcel is bought rather than after, TraceAir's land acquisition and pre-development tools are where that side of the work starts.

Frequently Asked Questions

What Is AI Construction Analytics?

Machine learning applied to repeat drone capture of the same site over time. On TraceAir's platform it produces four things: soil behavior tracking and prediction, automatic detection of stockpiled and SWPPP material quantities, automatic pavement progress tracking, and scan to scan comparison. The analytics are the interpretation layer, not the capture.

Should I Use LiDAR Or Photogrammetry On A Raw Parcel?

LiDAR, if the parcel has vegetation. Photogrammetry reads the surface it can see, so under brush it reads canopy, not ground. LiDAR returns multiple hits through canopy gaps, which is what lets the vegetation be classified out and true ground modeled. On a cleared, actively graded site either works.

Does This Only Work On Large Sites?

No. The headline numbers come from big projects, but the recurring value is ordinary division-scale catches. D.R. Horton's Memphis division caught a lot layout problem before staking, avoiding roughly $50,000 in redesign, and confirmed from scan data that a planned set of retaining walls was unnecessary, taking $300,000 out of the job.

When Should A Site Be Scanned For A Land Decision?

Before clearing, if the question is what the ground actually is. On a 100-acre site in Temecula, dense brush blocked traditional methods from establishing original ground topography, and LiDAR captured it before grading and clearing began. Once a site is cleared, the pre-disturbance condition cannot be recovered.