The obvious first AI project is a chatbot. We'd skip it. Start with the job everyone agrees matters and nobody gets to.
For one client, that job was their brand library: more than 8,000 photos, logos, videos and decks. Most of it was untagged. Search came up empty, so people reshot photos they already owned.
One person on our team and AI fixed it in a single afternoon. By hand, it would have taken about 245 hours. Here's how we did it, how we kept it accurate, and why this kind of project is the right place to start.
Every brand has a junk drawer
The library held more than 8,000 files. 77% had no tags at all, so search and filters came up empty. About one in four still carried a camera's default name: IMG_4821.jpg, 20230612_101533.jpg, DSC_1182.JPG.
Those files are invisible. When people can't find a photo, they ask around, reshoot it, or use whatever turns up first.
Why it never gets done
The plan already existed. The team had a spreadsheet listing every file and the tags it should carry. Nobody had time to enter them.
At about 90 seconds a file, 6,860 files is roughly 172 hours of tagging. Add matching, checking and reporting, and it's about 245 hours. That's six weeks of one person's time.
Nobody can give up six weeks for data entry. So the project slid to next quarter. Every quarter.
What we did instead
We ran it as one working session. Our person made the decisions. AI did the repetitive work, in four steps:
-
Match. Paired 8,450 spreadsheet rows with the files they describe, using file name and size to tell duplicates apart.
-
Check. Tested all 1,169 planned tags against the tags the system would accept. Where the team had already picked a replacement, we used it.
-
Tag. Applied tags to 6,860 files, folder by folder. A person reviewed a small test batch first.
-
Verify. Re-checked every file after tagging and wrote a status for each one back into the team's spreadsheet.
Four rules that kept it accurate
This is a live library people use every day. So before touching a single file, we set four rules:
- The team's data is the source of truth. AI applied the tags people chose. It never made tags up.
- Only approved tags. Anything the system didn't recognize went to a person, not into the library.
- Add, never overwrite. Existing tags stayed put. Conflicts went to review.
- Log every file. Each file was checked after tagging, with its status written down for anyone to audit.
We also ran a 20-file test and had a person check it before the full run.
The result: 23% to 89% in one afternoon
Tagged coverage went from 23% to 89%. The ten largest folders landed between 96% and 100%. Every tag came from the team's own spreadsheet.
The work took about 2.5 hours, against an estimated 245 by hand. Roughly 100 times faster.
And 245 is conservative. At a more realistic two to three minutes a file, the manual estimate climbs to 300–400 hours.
The surprise: a gap nobody could see
Checking every tag exposed something no one could spot by hand. Of the 1,169 tags the team planned, only 123 existed in the system. About one in ten.
AI grouped the missing tags by where they'd live (asset types, industries, community partners) and flagged likely typos. The team got a short list of decisions instead of a thousand-line error report. What to add stayed their call.
Start with the job nobody wants
AI didn't replace anyone here. It took the part of the job nobody wanted. That's why it worked:
- The decisions were already made. The team knew what the tags should be. AI entered them.
- The rules were clear. Approved tags only, nothing overwritten, every file logged.
- The payoff was visible. A library people can search, and a clear list of what their tags were missing.
Flashy AI projects stall on open questions. Boring ones finish, because the thinking is already done.
So look for a project with a plan, a spreadsheet, and no time. A shared drive nobody can search. CRM fields nobody filled in. A content library with no metadata. That's your first AI project.