In this piece
How much of your company photo archive can AI sort?
Think of a construction photo forwarded through WhatsApp. It’s the example I use to explain our work for Burgenwelt Ehrenberg in Tyrol. The copy has no coordinates and there is no recognizable landmark in the frame. A model can tell you there is construction work in the picture. Which attraction was under repair? A colleague who remembers the work may be the only person who can file it where someone else will look for it later.
Ehrenberg had accumulated years of media in cloud folders. The folder names made some photos hard to find, and moving files individually in the interface available to the team made a manual cleanup impractical. We worked with staff on the categories and built a way for them to correct our first pass. Our Ehrenberg archive case study describes what we delivered.
Which photos should you give a supplier for a trial?
Ask the people who use the archive to pick a few real searches. They might need photos of a particular attraction, a renovation or an event. Agree what they would call the corresponding folders. Give the supplier a working copy that includes the old folder paths, along with examples from different years, cameras and contributors. Include pictures that staff struggle to place themselves. A trial made up entirely of well-labelled camera originals will tell you little about those.
| Include in the sample | What you want to learn |
|---|---|
| A camera photo with a date, coordinates and elevation | Can the supplier explain which attraction the location points to and recognize when another one is close enough to be plausible? |
| A recognizable attraction with no usable coordinates | Can a reviewer see which feature led the model to suggest a particular folder and check the suggestion against the image? |
| A construction photo forwarded through a messenger, with no landmark or usable location data | Does the file stay open for someone familiar with the work, even if the model can accurately describe the scene? |
What can the supplier infer from the files?
At Ehrenberg we started with the old folders and the available capture data. The date helped put photos in sequence. The EXIF GPS fields can include the camera’s coordinates and elevation. Some attractions were only a few hundred metres apart, so an elevation reading could help distinguish two otherwise plausible locations. Camera position and the place shown in the picture can differ. The IPTC Photo Metadata User Guide gives each its own field.
When those clues left a question open, we used AI to inspect what was visible in the image and suggest a group. That helped when the subject was recognizable. For the construction example, recognizing the subject still leaves the project unknown. The team also had to decide whether it wanted to retrieve such a photo under the attraction or under the work carried out there. Only the people using the archive could settle that question.
The Library of Congress describes a similar problem with uncaptioned photographs. It invited readers to look for clues in the images. A name or sign that someone could identify gave the archive a way to describe the subject more precisely.
How did staff correct a folder without moving every file?
We hosted a separate Nextcloud instance in our office and gave Ehrenberg access to the working copy. Staff looked through successive arrangements and showed us which groupings made sense and where we had misunderstood a category. They could try a different arrangement while their own archive stayed available.
For the remaining decisions, we built a local browser tool. Staff could correct a destination or mark an item for deletion as they looked through the photos. They sent us exports of those decisions. We used them to apply the moves and deletions to our Nextcloud copy. The agreed next step was to carry the changes into Ehrenberg’s own Nextcloud once the team gave final approval.
For your own trial, ask for a review record you can read alongside the image. It should show the original file path, proposed folder, reason for the suggestion, reviewer’s decision and approved action. For an unidentified building site, a useful entry could read “Construction visible. Project unknown. Ask the site team. Leave the file in place.” That entry gives the next reviewer an actual question to answer.
When is it worth commissioning the whole archive?
Have someone who uses the archive try the original searches in the trial folders. Record which photos they find, which assignments they correct and which files still need a colleague with local knowledge. Time the review as well as the initial sort. If staff repeatedly move photos between the same two categories, revise those categories and try that part of the sample again.
Use the trial to agree which folders can be processed together, who can resolve exceptions and who signs off on changes to the live collection. If staff can find the material, handle the remaining reviews in the time available and reproduce their edits on a copy, you have a basis for pricing the rest. A folder of unresolved images can stay open while well-understood groups move ahead.
If the archive trial is part of a wider AI programme, our guide to deciding whether to scale an AI pilot covers the investment decision. Our guide to scoping a pilot explains how to agree on the evidence before commissioning a larger project. To discuss an archive pilot with our computer vision team, bring a few representative photos and the folder names your colleagues already use.