Skip to main content
DoneThat

AI Adoption GuideConstructionMobilize

Site Setup Compliance Vision Check

Vision model reviews site setup photos against hoarding, signage, and welfare facility requirements.

Construction processBidAwardPlanMobilizeBuildInspectHandoverClose

By Don, DoneThat’s AI coach · updated

Compare the photo to the setup pack, not to memory

A vision check on mobilize is a quality finding: it names what the photo appears to show, cites the frame, and cites the written requirement. It is not a closed inspection, not a punch signed off, and not a reason to invent hoarding, a missing gate, or a welfare cabin the image never contains.

If the photo is unreadable, the finding stays empty. Blur, night, glare, a thumb over the lens, or a shot that only shows sky is not evidence that the site is compliant, and it is not evidence that an item is missing. Empty is the correct output. Do not fill the gap from yesterday's walk, from the logistics drawing, or from a conversation with the gateman.

The written pack is the source of truth. That pack is usually the construction phase plan extracts, the site induction board specification, the hoarding and public protection details, the traffic and pedestrian signage schedule, and the welfare provision notes covering toilets, drying room, canteen, first-aid point, and drinking water. Hold the image next to those clauses, not next to what you remember from last week's visit.

The model does the looking pass. You still own the drawing. If the pack and the photo disagree, the finding reports the disagreement. If the pack is silent on an item, the model does not invent a requirement.

Site setup photos also sit next to site logistics layout optimisation. Layout work decides where hoarding, gates, and welfare sit. This check only asks whether the installed condition in a photo matches the requirement for that location. Do not mix the two jobs.

Run the check as a three-part cite

Work in this order every time.

Freeze the requirement first. Pull the clause for that location: hoarding height and continuity, public-facing signage (principal contractor, hours, complaints line, PPE, no unauthorised access), internal wayfinding, statutory notices, and welfare (facilities against the headcount the plan assumes, lighting, hot water, drying). Write the clause ID or paragraph number into the finding before you look at pixels. If you cannot name the clause, you are not ready to run the model.

Attach the photo second. Use the frame taken for that location: gate 1 looking out, hoarding bay 4 looking west, welfare compound looking at the cabin doors. Record who took it, when, and from where. Photos already sitting in Procore or Autodesk project folders are a normal class of record. Do not feed a marketing drone still and treat it as a welfare inspection shot.

Ask only for a quality finding third: visible, not visible, or unreadable, plus the two cites. The output should look like this shape. Photo DSC_1842, gate 1, 06:40. Requirement: hoarding continuity and lockable vehicle gate per site logistics note 4.2. Finding: vehicle gate is in frame and appears closed. Pedestrian wicket is in frame. Site board text is not readable at this distance. Welfare is not in frame. No finding on toilets.

If toilets are not in the photo, you do not invent missing toilets. You also do not mark welfare complete. Not in frame is not a defect. Not in frame is a gap in evidence.

Hoarding, board, and welfare in three frames

Use this as a pattern for a single mobilize morning, not as a named project.

A site-setup lead drops three photos into the check: the public hoarding along the footway, the principal-contractor board at the pedestrian gate, and the welfare compound from the compound entrance.

The hoarding photo is sharp. The model cites the photo and the hoarding detail. Panels look continuous in frame. There is no obvious gap at the camera's standing position. Mesh above the solid base is visible. The finding does not claim the whole perimeter is complete. The camera never turned the corner.

The site-board photo is taken from the road. Logos are blobs. Hours and the complaints number cannot be read. The correct finding is unreadable text, not signage compliant. Treating blur as a pass is how a false close gets into the register.

The welfare photo shows cabin doors, a generator, and a stack of stillage. You cannot see inside the toilets or the drying room. The finding cites the photo and the welfare clause, then states that internal provision is not evidenced. It does not close the welfare punch.

HSE walks those three locations with the findings in hand. They confirm the hoarding bay, they walk closer to the board until the text is readable, and they open the welfare doors. Only then does a punch get closed or raised. The model never closes it.

Assume these three mistakes will show up

Flagging a completed gate as missing is common. A photo taken from inside the site, looking along the hoarding, often crops the leaf of the vehicle gate. The model sees an opening and reports gate missing. The gate is in the rebate, fully hung. Do not argue with the model. Reject a finding that cannot cite the gate in frame, retake from the public side so the leaf, posts, and lock are visible, and keep the punch open until HSE has seen the hardware.

Reading a blurry photo as compliant is the quieter failure. Night shots, rain on the lens, and boards shot from a moving van all produce a board is present. Presence of a rectangle is not compliance with the signage schedule. If you cannot read the required lines (contractor name, hours, contact, mandatory PPE, unauthorised-access warning), the finding is unreadable. Empty or unreadable beats a false pass.

Closing the punch from the model is the one that creates a papered-over gap. A quality finding is a draft observation. It is not an inspection record, not a CDM evidence pack, and not a reason to clear a gate, hoarding, or welfare item in Procore, Autodesk, or any other register. If someone treats the flag as closed, the vision pass has replaced the walk. HSE confirms on the ground. The register updates after that confirmation, not before.

Live works checks, such as real-time safety hazard detection, are a different job. Those look for people, plant, and open edges once the site is running. This page is only the mobilize setup condition in still photos. Do not let a hazard model confirm hoarding. Do not let this vision check stand in for a live exclusion-zone watch.

Hand the packet to HSE, then walk it

The useful packet is short: photo ID, location, clause, finding text, and a clear unreadable or not-in-frame note where the camera failed. HSE uses that packet as a punch seed, not as a certificate.

Confirm in this sequence. Walk the cited location with the photo on the phone. If the finding says missing, look for crop, angle, and temporary removal, such as a gate leaf swung out of frame during a delivery. If the finding says present, still read the board yourself and still open the welfare door. If the finding is empty, retake. Do not backfill the clause from memory.

Keep adjacent mobilize work in its own lane. Competency packs for the people who will occupy the welfare and run the gate sit in worker competency document verification. Chasing a subcontractor for a missing RAMS or induction record sits in subcontractor compliance chasing agent. Neither of those jobs should be answered by a photo of a cabin.

When you brief the team, say the rule out loud. The model may point at a photo and a clause. People close the item. If the photo cannot support the clause, the finding stays empty, and that line stays open until someone with a tape, a key, and a pair of eyes has been there.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first