AI Construction Progress Monitoring From Site Photos in 2026

The camera on the hardhat is now the superintendent’s second set of eyes

Walk a job site on a Friday afternoon and ask three people how far along the second floor is. You’ll get three answers. The drywall sub says 80 percent. The PM’s schedule says 65. The owner’s rep, who last visited two weeks ago, has a number that’s basically a guess. Nobody is lying. The site just moves faster than anyone’s ability to write it down.

That gap is what AI progress monitoring closes. The idea is simple to state and hard to execute: capture the site as images, line those images up against the BIM model and the schedule, and let computer vision count what’s actually installed. No clipboard. No arguing about percentages in the trailer. The software looks at a wall and tells you whether the framing, the rough-in, the insulation, and the board are there — element by element, floor by floor, week over week.

It’s one of the few construction-tech pitches that has graduated from demo to daily habit. Crews walk the floors anyway. Strapping a 360 camera to a hardhat adds nothing to their day. The data just shows up.

How the capture actually happens

There are three common ways to feed these systems, and they’re not interchangeable. The lightest is a 360-degree camera — an Insta360 or a Ricoh Theta — clipped to a hardhat. The person wearing it does their normal walk. The camera shoots continuously, GPS and the building’s known geometry stitch the frames to a floor plan, and you get a navigable walkthrough plus the raw images the AI needs. Buildots built its whole workflow around this. One walk, no dedicated capture crew.

The second is the drone or the tripod-mounted 360, which OpenSpace and others support for exterior shells, large open floor plates, and anything too tall or too dangerous to walk. The third, and the heaviest, is LiDAR. Doxel leans on laser measurement because you can’t count earned value from pixels alone — you need real dimensions to say a slab is poured to the right elevation or a duct run is the length the model says it should be.

Most mature teams mix them. 360 for the weekly interior sweep, drone for the facade and the sitework, a LiDAR pass at milestones that matter for payment. The capture method follows the question you’re trying to answer.

What the AI is doing between the photo and the dashboard

This is where the marketing gets vague, so let’s be concrete. The model has two jobs. First, localization: figure out exactly where in the building each frame was taken. That’s a mix of visual landmarks, the capture device’s position, and the BIM geometry. Get this wrong and every downstream number is garbage.

Second, detection: look at a given location and classify the state of each tracked component. Is there a stud wall? Is the MEP rough-in in? Is the ceiling grid hung? OpenSpace says it tracks more than 700 building components this way. The system isn’t recognizing objects in the abstract — it’s checking the model’s expectations against the photo and marking each element as not started, in progress, or complete.

The output that matters isn’t the pretty 3D walkthrough, though clients love those. It’s the delta. The software knows what the schedule promised for this week and what the camera found. The difference is the report. A trade that’s quietly three weeks behind on the fourth floor shows up as red before it becomes a sequencing disaster on five and six.

Where it earns its keep: delay forecasting

Catching a delay is useful. Predicting one is the actual product. Because these systems log progress rate per trade per area, they can extrapolate. If the ductwork crew has been installing at a certain pace and the model shows how much is left, the finish date isn’t a vibe — it’s arithmetic the software does every week. Buildots markets catching delays about three weeks earlier than a human super typically would, and that lead time is the whole game. Three weeks is enough to add a crew, reshuffle a sequence, or warn the owner before the conversation turns legal.

Doxel pushes the same logic toward money. It ties measured, verified progress to earned value, so the question stops being “does it look done?” and becomes “have we actually banked the work we’re paying for?” On a cost-plus job or a lender-financed project, that distinction is worth more than any render.

The honest limitations

None of this is magic, and anyone who tells you it is hasn’t run it on a messy job. Occlusion is the big one. A camera can’t see the pipe behind the wall that just got boarded, so timing the capture matters — you walk before the cover-up, not after. Miss the window and the AI records “complete” without ever confirming what’s underneath.

Model quality is the other hard dependency. These tools are only as good as the BIM they check against. Feed them a stale model where the real building diverged from design three change-orders ago, and the deltas get noisy. Someone has to keep the model honest, which is labor the sales deck rarely mentions.

And detection isn’t perfect. Shiny surfaces, bad light, a tarp thrown over a stack of materials — the model guesses wrong sometimes. The good teams treat the output as a very fast, very thorough first pass that a human still sanity-checks, not as gospel. The win isn’t zero human judgment. It’s pointing human judgment at the 5 percent that’s actually in dispute instead of the 100 percent nobody had time to inspect.

Who’s actually buying this, and why now

The early adopters were big general contractors on large commercial and data-center work — projects where a week of schedule slip costs real money and the floor plates are big enough that nobody can walk them properly by hand. That’s still the core market. OpenSpace claims deployment on more than 85,000 projects across 70 countries, which, true or rounded, tells you this left the pilot phase a while ago.

What changed to make it stick? Cameras got cheap and good. BIM adoption finally hit the point where most serious projects have a model worth checking against. And the labor math got brutal — experienced superintendents are scarce, and asking the few you have to spend Fridays counting drywall instead of solving problems is a bad trade. The software doesn’t replace the super. It hands back their Friday.

For interior fit-out and design-build teams, the pull is slightly different. It’s documentation. Every captured walk is a timestamped record of what existed when — priceless when a dispute lands, when an owner asks what’s behind a finished wall, or when the facilities team inherits the building and wants to know where things actually are.

Fitting it into a real workflow

The teams that get value do a few unglamorous things. They pick one capture owner per site so the walk actually happens every week — inconsistent capture kills the trend data that makes forecasting work. They keep the BIM current, or at least honest about where it isn’t. And they wire the output into the meeting they already have, so the weekly progress review opens with the dashboard instead of with everyone’s competing memory.

Start narrow. One tower, one trade, one honest model. Prove the delta reports match reality for a month, build trust with the field crews who are understandably skeptical of a camera grading their work, then widen. The failure mode is buying it for the whole portfolio, capturing sporadically, and concluding the AI “doesn’t work” when the real problem was the walk nobody did.

The record outlives the project

Here’s the part owners underrate until they need it. Every weekly walk builds a time machine. Six months after handover, when a leak shows up above a conference room, you don’t open a wall and guess — you scroll back to the week before the ceiling closed and look at exactly how the pipe was run. That capability quietly changes the economics of a building long after the crews leave.

It reshapes closeout too. Instead of chasing subs for as-built markups nobody kept current, the facilities team inherits a navigable visual history tied to the model. Warranty claims get easier to adjudicate because there’s dated proof of what was installed and when. On a portfolio of buildings, that archive compounds into something closer to institutional memory than a pile of PDFs in a folder nobody can find. The capture cost was sunk during construction anyway; the record is the dividend you collect for years.

Common questions about AI progress monitoring

Do I need a complete BIM model to use it?

For the full element-level tracking and delay forecasting, yes — the AI compares the photos against the model, so no model means no benchmark. You can still get value from the plain documentation layer (navigable, timestamped walkthroughs mapped to floor plans) with just a 2D plan, but the automated “percent complete per trade” magic needs the 3D model behind it.

How is this different from scan-to-BIM?

Scan-to-BIM builds a model from reality — you capture an existing building and generate geometry you didn’t have. Progress monitoring assumes the model already exists and checks reality against it over time. One creates the baseline; the other watches you build toward it. Teams often use both on the same project at different stages.

Will it tell me who’s behind schedule, or just that something is?

Both, if the schedule is linked properly. Because tracking is element-level and tied to trades and areas, the report can say the HVAC rough-in on level four is lagging its planned curve by two weeks, not just that “level four is behind.” That specificity is what makes the weekly meeting shorter and the accountability clearer.

Is it worth it on smaller interior projects?

It depends on what you value. For a quick tenant fit-out, the forecasting may be overkill. But the documentation alone — a clean visual record of every condition before it got covered — can justify it on any job where disputes, warranties, or handover records matter. Smaller teams often adopt it for the record-keeping first and grow into the analytics later.

Curious how AI-driven visualization and design tools fit your own projects? See what we build at Pixintellect.

Images: AI-Designed