AI Image Tagging: Turning Your Site Photo Chaos Into Evidence

Every QS has a phone full of site photos with names like IMG_4821.jpg, no idea what date they were taken, and no chance of finding the one you need when a compensation event lands on your desk. AI image tagging fixes that. Here's how it works and where it's genuinely useful.

The problem

You know the drill. Something happens on site, someone takes a photo, and six months later you're scrolling through a camera roll trying to find evidence of ground conditions on a date you can barely remember. Untagged, unsorted photos are one of the biggest gaps in commercial record keeping, and it costs real time when you're building a compensation event or defending a valuation. It's the same problem behind why SharePoint fails as a photo store.

AI image tagging tools automatically scan photos and label what's in them: person, hard hat, excavator, standing water, scaffold, concrete pour. Some go further and read text in the image, spot missing PPE, or estimate percentage complete against your programme.

The result is a searchable, timestamped record instead of an inbox of jpegs. There's now a genuine range of tools doing this, from the generic engines that power a lot of it under the bonnet, through to construction-specific platforms built for exactly this job.

The engines underneath: AWS Rekognition, Google Vision, Azure

Start with the plumbing, because it explains what every construction tool is actually built on. AWS Rekognition, Google Cloud Vision AI and Microsoft Azure AI Vision are general-purpose image recognition engines, sold by the API call, none of them a construction product in their own right.

They identify thousands of common objects and scenes in a photo with a confidence score attached; they can read text caught in a picture (a delivery note, a plant registration plate, a drawing number), and several have a specific function for checking whether people in an image are wearing head cover, face cover and hand cover.

None of them knows what a compensation event is, or what your programme looks like. They just tell you what's in the picture. The construction-specific value comes from what gets built on top.

The reality capture and progress platforms

OpenSpace, Buildots, Disperse and Reconstruct all take a generic engine and wrap it in construction logic. The common pattern is a 360-degree camera, sometimes helmet-mounted, walked through the site on a regular cycle, with the resulting imagery automatically stitched to your floor plans and compared against your BIM model or programme. The output is a percentage complete by element, flagged deviations from plan, and a full visual timeline you can step back through if a dispute lands.

This is the category with the clearest QS relevance: objective, dated, located progress evidence that isn't reliant on someone remembering to take a photo of the right thing.

The safety-focused tools

Newmetrix, formerly Smartvid.io and known for a product called Vinnie, pioneered AI safety photo tagging, scanning site imagery for missing PPE, unsafe conditions and near misses, and building a predictive risk picture over time. Worth checking current availability before relying on it as it appears to have wound down as a standalone product, but the pattern it established, safety photo tagging feeding a hazard log, is now a feature inside several broader platforms rather than a standalone category.

The QS relevance here isn't primarily safety reporting, it's that the same detections often evidence site conditions relevant to a disruption claim.

The platform players

Procore has folded AI image and document analysis into its existing construction management platform rather than selling it as a separate product, which is the direction most of the market is heading. If you're already running photos, diaries and cost records through one platform, this is usually the path of least resistance, because the tagging sits next to the records it needs to support rather than in a separate app you have to remember to check.

Think about what you'd actually search for

Whichever category you're looking at, start the same way. List the things you currently struggle to find. Photos of a specific defect. Evidence of site conditions on a particular date. Proof plant was standing idle. That list tells you whether basic tagging and search is enough, or whether you need the fuller progress monitoring platforms with BIM comparison built in.

Treat every tag as a lead, not a fact

Across every one of these tools, a confidence score of 94% still means the machine could be wrong, particularly in poor light, at a distance, or where PPE is partially obscured. Real site accuracy runs well below the figures quoted in vendor case studies, which are usually benchmarked on clean datasets rather than actual site conditions.

Use tags to find the right photo fast. Don't cite the tag itself as your evidence. Look at the photo, confirm what you're seeing, then use it.

What to watch out for

Site photos usually contain identifiable people, which makes them personal data under UK GDPR. If you're using anything with facial recognition switched on, that's a conversation with your data protection lead before go-live, not after. And none of this is useful unless the tags actually land in your site diary or CDE alongside the cost and programme records. A tagging tool that lives in its own silo just gives you a better organised shoebox.

Worth flagging too: from 9 March 2026, the RICS Responsible Use of AI standard will expect members to be able to show how AI outputs were checked and used in professional decisions. If a tagged photo ends up substantiating a valuation or a CE, you'll want a record of that human check, not just the AI's word for it.

Try this today

Pick your last five compensation events or variations and time how long it actually took you to find the supporting photos for each one. That number is your business case. If it's more than a few minutes per claim, tagging and search alone will pay for itself before you get anywhere near progress monitoring or safety analytics.

How we use this at Gather

Since we're on the subject, worth being straight about our own approach rather than just talking about everyone else's. We use AWS Rekognition to annotate site photos, which further enriches the contemporaneous record sitting alongside each diary entry. It's the tagging and search end of the spectrum described above rather than the progress-monitoring or safety-analytics end, applied directly to the record you're already keeping rather than as a separate tool to check.