A project manager opens a folder of bathroom photos to answer a simple question: do we have a clear record of the blocking behind the vanity? An AI assistant can help find candidate photos. Its answer is useful only if the reviewer can confirm which room they show, when they were taken, and what remains outside the frame.

Datum’s proposed starting point is a photo evidence review: retrieve the likely images, carry their recorded context into the review, and identify what the project manager still needs to check. Keep any decision to accept work or proceed with the next activity in the existing project process.

Send the record with the image

Procore’s current photo-viewer documentation separates project context, including location and trade, from technical fields such as Taken On, Uploaded On, and file name. That distinction gives teams a useful starting structure. A photo uploaded today may show work from an earlier visit.

Anthropic’s vision documentation says Claude does not receive or parse image metadata. If your workflow sends only the image, do not assume the model also knows the capture date or recorded location. Supply those fields separately from the source system, with a link or stable file reference the reviewer can open.

Treat those fields as recorded information, not independent proof. A missing capture date stays missing. A room label supplied by the uploader may need confirmation. Keep the original record separate from an AI suggestion so a guessed location cannot silently become the location everyone relies on.

Ask a question the photo can help answer

In a hypothetical remodel, the folder contains a wide view of the primary bathroom, a close-up of blocking, and a similar wall in the guest bathroom. A useful assistant brings back the candidate images and explains which recorded details support each match. If the close-up has no location and lacks enough surrounding detail, it should flag the match for review.

The project manager can then request a specific missing view while the wall is still accessible. A generic instruction to take more photos leaves the field team guessing. A request that names the room, wall, feature, and missing context gives them something to act on.

Anthropic also documents limits in image interpretation, spatial reasoning, and counting. For this workflow, avoid asking the model to certify dimensions or installation quality from a casual photo. A visible component, a field measurement, and an accepted inspection record answer different questions. Link the relevant record when the decision requires it.

Make a short review sheet

  • Question: the project, room, work item, and specific detail the reviewer wants to find.
  • Candidate evidence: original photo reference, recorded capture date, upload date, location, and source of each field.
  • Visible detail: a tentative description tied to the image, with unclear or obscured areas stated plainly.
  • Missing evidence: the view, measurement, or existing project record needed to answer the question.
  • Follow-up: the responsible person and the date supplied by the project team, with the reviewer’s decision recorded separately.

Keep the search scope visible. If the assistant checked only one album, “no matching photo found in this album” is a defensible result. “The work was not documented” reaches beyond what it checked. If an album fails to load, report that failure instead of treating it as an empty folder.

Preserve the originals. An annotated copy can help point out a feature, but it should remain linked to the unedited image. In the first trial, let the assistant prepare the sheet for review without changing photo labels or sending requests to the field.

Test whether the project manager finds the answer faster

Use a small set of resolved photo questions for a practice run. Have the project manager identify the relevant images and missing evidence first. Include similar rooms, duplicate images, a late upload, a missing date, and an inaccessible album. Give the assistant only the records that would have been available at the time.

Count wrong-room matches, unsupported dates, missed candidate images, and claims that go beyond what the photo shows. Measure the time spent finding, checking, and correcting the answer against the current manual process. This is a proposed pilot, with no performance result assumed.

If the review takes too long, narrow it to one recurring question, such as locating pre-cover photos for a named wall. The output should let the project manager open the evidence and decide whether to ask for another view before the opportunity to capture it is gone.

Sources Read

Next step, if this note maps to a problem on your desk: book a conversation with Adam. Thirty minutes, weekday evenings, Eastern time.

Related Field Notes