I recently asked an AI to create a headshot of Al Wadey.
It gave me a perfectly nice-looking man in a navy blazer. It wasn’t me, and he didn’t look like me. The AI hadn’t executed a search for me or asked who I was. It invented someone who, in its view, looked like an ‘Al Wadey’ and carried on.
This isn’t the AI’s fault. I gave it a name and nothing else.
So, I tried again. This time I gave it a holiday picture from Cape Verde and told it what I needed: a professional headshot for a profile picture or a presentation. The result was me, cleaned up and usable. Then I pushed it further. I said this was boring and asked for something funny that might get people to come to one of my recent conference talks.
It turned me into a lifeguard. I am waste deep in a sea of files called FINAL.jpg, FINAL_final.jpg and FINAL_final_v7.jpg. I’m holding a “DAM Rescue” lifebuoy, and there’s a rubber duck in sunglasses bobbing beside me. The headline reads “Drowning in assets?”
I never told it the conference was about digital asset management. It worked that from months of earlier conversation about my work.
Each step was better than the last. The model was the same every time. The only thing that changed was how much context it had.
Looking back, I needed five things to get it right:
Leave any of these out and the AI fills gaps with a guess. Sometimes the guess is charming, like a lifeguard with a duck. In a DAM, it usually isn’t.
Take a stock image of a woman sitting on the sofa watching TV. On the screen, someone is drilling into a wall.
I asked AI to target as if I were a company that sells drills. It came back with “Drill in action on a living room TV screen”. The primary subject was the power drill, the category was power tools/drills, and the keywords were DIY, drillbit and home improvement.
Then I asked it to tag the same image as if I sold televisions. Now it was “Large television in a furnished living room”. The subject was a television, the category was home entertainment/televisions, and the keywords were living room, screen and viewing environment.
Both sets of metadata are accurate. Neither is right until you know who will be searching for that image and why.
This is the point most AI tagging conversations skip. Metadata is not a description of an asset. It’s a statement of what the asset is for. Every tag is a decision about context, made by someone. If you don’t make that decision, the AI makes it for you, based on whatever it happens to notice.
An untagged asset is invisible. That’s a problem, but at least it is an honest one.
A badly tagged asset is worse. It turns up when nobody wants it and stays hidden when someone does. The search starts returning televisions when your team is looking for drills. Another round of that, people stop trusting the DAM and go back to asking a colleague, or hunting through shared drives.
Rolling out AI tagging without the context to store it doesn’t fix find ability problems. It speeds it up.
Now say your TV company aquires a drill company.
That living room image hasn’t changed, but what it means to your business has. Which context applies now: drill, TV or both? Do you merge the Metadata, create versions of the asset, or tag it differently for each brand? There’s no universal answer. There’s only the answer that fits how your business wants to use that content.
This is why prompts can’t be written once and forgotten. They need updating for two reasons:
The same applies across the whole asset lifecycle.
Ingestion is getting context right the moment the new content lands.
Enrichment is going back and improving content in bulk when the context shifts.
Migration isn’t the one-off project most people imagine. For large Enterprises it happens again and again, every time a brand is acquired or forgotten content turns up.
The anxious version of the story says AI makes the DAM librarian redundant. I think the opposite is true.
Today, in most organisations, the librarian does every step by hand: prepare, check, flag, tag and add metadata, review, publish, and then start again. It’s skilled work spent largely around repetition.
With AI and a humanin the loop, the work divides more sensibly.
The tool does the heavy lifting: it prepares assets andenriches them with metadata,flagging anything it isn’t sure about.
Your librarians bring the context: they review what was flagged, resolve what the AI got wrong and improve prompts so that the next batch needs less correcting.
That last step matters most. It’s what turns a tagging tool into a system that gets better over time. Each round, you can rely on the AI a little more, because people have taught it what your content means.
Everything I showed in my talk, from the images to the Metadata, was generated by AI. None of it would have worked without a person deciding what it is for.
The organisations that get the most from AI in their DAM won’t necessarily have the most advanced model. They’ll be the ones who are clearest about what their content is for, who it serves, and how that changes overtime.
At ICP, we saw these problems over and over across our managed content services clients, so we built Ponseta to allieviate our team from the heavy lifting but keep the context in human hands. If you’re facing the same challenges, we’d love to talk.