A site supervisor takes photographs before a ceiling is closed. Weeks later, the commercial team needs to explain which work was complete, which drawing applied and whether a variation was actually accepted. The photographs exist. Finding and interpreting the right ones is the work.
For an Australian contractor buying AI software, this is a more useful starting point than a promise to understand every image. Build a system that retrieves relevant evidence and proposes descriptions, then makes the review decision explicit. A caption saying that installation appears complete must never silently become a certified quantity or an approved payment.
Give the AI a job that a reviewer can check
Australia's National AI Centre adoption tracker examines how SMEs use AI, including applications, expected outcomes and responsible practices. It provides context for evaluating practical adoption; it does not establish that a particular vision model is reliable on your sites.
A useful first capability is to suggest a location, work package and relevant drawing for an uploaded photo. Another is to retrieve earlier images from the same area when a reviewer investigates a change. These are proposed product functions, not claims about a deployed client system. They help an experienced person spend less time searching while retaining responsibility for interpretation.
Define the boundary in the interface. Show suggested tags beside the original image. Distinguish a worker's observation from an AI suggestion and a commercial review. Let the reviewer reject a suggestion without editing the original upload. Avoid one green tick that appears to approve everything from image quality to contractual entitlement.
A folder name is a weak location system
Consider an illustrative project with repeated apartment layouts. An image from unit 410 may look almost identical to unit 510. Visual similarity alone cannot establish where it was captured. If the system confidently assigns the wrong floor, an impressive search experience can produce misleading evidence.
Give each upload an explicit project, area and work package. Capture the source of those identifiers: selected by a person, scanned from a site label, imported from a task, or inferred by software. Preserve capture time separately from upload time because a phone may reconnect later. Make missing or disputed context visible instead of filling it with a plausible value.
The same principle applies to drawing revisions. A photograph should link to the drawing revision used for the review, not whichever file happens to be newest today. If that revision is superseded, keep the original association and explain the later change. This makes the evidence intelligible after the daily site team has moved on.
Provenance helps, but it does not prove completion
The C2PA specification collection describes standards for recording the source and history of media. Its Content Credentials explainer explains how provenance and integrity information can be checked. Those checks support a chain of custody; they do not prove that the photographed work complies with a drawing or was complete at a contractual milestone.
Keep the original file, calculate an integrity hash and record subsequent annotations as separate versions. Where a source includes usable Content Credentials, preserve and validate them. Where it does not, display that limitation without classifying every ordinary phone image as false. A copied screenshot and a first-party capture should not have indistinguishable histories.
Do not add generative cleanup to the evidential original. Cropping for a thumbnail is different from removing an obstruction or creating detail. If a derivative is useful for viewing, label it, preserve the original and let the reviewer open both. The application should make the less convenient source available precisely when an interpretation is disputed.
Design the review around the commercial question
An operations reviewer may ask whether a work package is ready for inspection. A quantity surveyor may ask what evidence supports a claimed change. Those are different decisions. Give each its own status, required context and responsible role.
- ▸An observation records what a person saw and when.
- ▸An AI suggestion proposes classifications or candidate evidence.
- ▸A review accepts, rejects or requests clarification on that evidence.
- ▸A claim assembles reviewed evidence under the relevant commercial process.
The software can link these records without collapsing them. If an observation is corrected, identify every draft claim that uses it. If a claim has already been issued, preserve its exact evidence package and add the correction through the agreed process. The product should help people explain history, not silently rewrite it.
Buy the difficult sample before the whole platform
Ask prospective development partners to work with a permissioned sample containing poor lighting, repeated rooms, old photos uploaded late and pictures unrelated to the selected work package. Include images where the answer should be unknown. A vendor that only demonstrates clear, isolated equipment photographs has not shown the workflow you need.
Measure retrieval usefulness, false location suggestions and reviewer effort separately. Agree what counts as a useful result before selecting a model. Time saved on image tagging is not automatically time saved on commercial approval, especially if reviewers must correct confidently wrong descriptions.
Your pilot should also test access. A subcontractor should see the records relevant to their work without browsing another subcontractor's commercial documents. Export a review package and verify that the original files, revisions, comments and decision history remain understandable outside the application. Data ownership matters when the project finishes.
Connect the evidence to the systems already in use
Start with one source of site photographs, one work-package register and one review destination. If your existing project system already manages contractual approvals well, integrate with it. Custom development is justified where evidence retrieval, context and review repeatedly fall between products, not merely because a newer dashboard looks better.
Put capture reliability, storage growth, permissions, model evaluation and human review into the scope. Include a way to disable AI suggestions while keeping uploads and reviews available. The commercial team should be able to continue working when an external model service is unavailable.
For the purchasing process, use our AI integration vendor evaluation guide and software discovery deliverables. TuniCyberLabs' custom software development can connect the evidence trail to the workflow your team already follows. Discuss a construction evidence pilot with a sample work package, the current photo sources and the decision that remains difficult to substantiate.