It started with a fairly simple question:
If we invented Digital Asset Management today, knowing everything we now know about AI and agents, would we still build a DAM?
I'm not suggesting DAMs are about to disappear. Far from it. There are very good reasons they exist and plenty of things they do incredibly well. But most of the fundamental thinking behind them comes from a world where humans were the only real users of the assets they contained.
/01The model we built
We had a problem: organisations were creating thousands, then hundreds of thousands, then millions of digital files. Those files needed somewhere safe to live, but more importantly they needed structure around them. We needed to know what they were, who created them, where they could be used, which version was approved and how somebody could find them again six months later.
So we built DAMs. The basic model made complete sense.
ASSET ↓ DAM ↓ HUMAN SEARCHES ↓ ASSET FOUND ↓ ASSET USED
Over the years we've made every part of that process considerably better. Better metadata, better taxonomies, better search, better previews, better workflows, better rights management, better integrations and better distribution.
But fundamentally, we've still been designing around a human arriving at the DAM and saying: "I need to find something." And that's the assumption I think AI starts to challenge.
/02What AI changes
Imagine someone asks for six strong landscape images of John Deere 6R tractors working in fields that we can legally use for paid social in Germany, but don't give me anything we used in last year's campaign. That's actually quite a sophisticated request.
A good picture editor could understand it immediately because they're combining lots of different types of knowledge without really thinking about it. They understand what's physically in the photograph, which product they're looking at, whether the composition works, whether six images are visually different enough, what campaign they're for, what was used previously and whether the usage rights allow what you're proposing.
Traditionally we've tried to make that knowledge searchable by turning as much of it as possible into metadata. AI gives us another option.
An AI model can suggest useful descriptions of a photograph’s scene, objects and composition, though it can also make mistakes. An agent could combine those suggestions with trusted records about rights, campaigns, products and previous usage, then propose a next step for someone to review.
So instead of the traditional model, we start moving towards something more like this:
HUMAN INTENT
│
▼
AGENT
│
┌──────────┼──────────┐
│ │ │
ASSETS RIGHTS CONTEXT
│ │ │
└──────────┼──────────┘
│
▼
ACTION
/03One level too early
At first glance the obvious answer is simply to put the agent inside the DAM, and that's exactly what we're starting to see happen. Semantic search, automatic tagging, natural-language search, generative AI, agents, APIs and MCP are all appearing across the DAM market. That makes sense, and some of it is genuinely impressive.
But I keep wondering whether we're stopping one level too early. Because there's a difference between making the DAM understandable to AI and making the asset itself understandable to AI. And that's where I think this gets much more interesting.
Take a single photograph inside a large organisation. The DAM might know its title, photographer, keywords and usage rights. The PIM knows exactly which product appears in it. The CMS knows which articles it's appeared in. The ecommerce platform knows which product pages are currently using it. A campaign system knows why it was originally commissioned. Another system may contain the model release, while analytics could tell us how frequently the image has been viewed or whether it has performed particularly well.
The organisation collectively knows an enormous amount about this photograph. The photograph itself knows almost none of it.
That's the bit I've been struggling with.
/04The asset as knowledge
What if instead of thinking about metadata as information belonging to the DAM, we started thinking about an asset as a portable piece of organisational knowledge? Conceptually, that changes the picture quite significantly.
ASSET
│
ASSET IDENTITY
│
┌───────────────┼───────────────┐
│ │ │
CONTENT CONTEXT RIGHTS
│ │ │
what is this? why exists? can I use it?
│ │ │
└───────────────┼───────────────┘
│
RELATIONSHIPS
│
┌──────────────┬───┴────┬──────────────┐
│ │ │ │
PRODUCT CAMPAIGN PEOPLE USAGE
│ │ │ │
└──────────────┴───┬────┴──────────────┘
│
PROVENANCE
The asset starts to become less like a file with some tags attached to it and more like a node in a knowledge graph. It could effectively say: I'm this particular asset. I depict this product, in this location, created during this shoot for this campaign. These are the people involved in creating me. These are the territories and channels where I'm allowed to be used. These are the other assets I'm related to. These are the versions that have been created from me. These are the places I've previously appeared. This information came from a human; this other information was inferred by AI; and this is the provenance showing how I've changed since the original was created.
Now we're getting somewhere beyond better search.
Interestingly, we already have quite a few of the foundations for this. IPTC and XMP have given us ways of describing and transporting metadata for years, while newer standards such as C2PA are beginning to address provenance and authenticity. So I don't think the answer is to invent another giant proprietary metadata specification and stick a .json file beside every JPEG. It's more about the philosophy of where that intelligence belongs.
And oddly enough, this is where Obsidian got me thinking. One of the things I like about Obsidian isn't really its interface. It's the idea that the application doesn't need to own the knowledge. The underlying information remains accessible and portable, relationships can exist between pieces of knowledge, and if you stop using the application tomorrow, you haven't necessarily lost the intelligence you've built up over years.
So I started wondering what the DAM equivalent of that philosophy would look like. Not literally Markdown files for photographs, but an environment where the intelligence surrounding an asset can survive whichever application happens to be managing it today. That starts to produce a very different architecture.
┌─────────────────┐
│ ASSET │
│ + IDENTITY │
│ + CONTEXT │
│ + RIGHTS │
│ + HISTORY │
│ + RELATIONS │
└────────┬────────┘
│
OPEN ASSET LAYER
│
┌──────────────────┼──────────────────┐
│ │ │
DAM CMS PIM
│ │ │
├──────────────┬───┴────┬────────────┤
│ │ │ │
CREATIVE AGENTS ECOMMERCE PUBLISHING
TOOLS
/05Where the DAM sits
The DAM is still there. But it isn't necessarily the centre of the universe anymore. It's one very important way of governing, managing and interacting with the organisation's assets, while the intelligence surrounding those assets becomes available to everything else as well.
That's particularly interesting once agents enter the picture. Instead of asking an agent to search the DAM for me, I could ask it to build the image pack for next week's electric vehicle feature. That's a very different instruction.
An agent could help interpret the editorial brief, find relevant assets, compare alternatives and prepare renditions or draft accessibility descriptions for an editor to review. Rights and expiry checks would still depend on reliable records, appropriate access and human approval.
Or I might simply ask: "We're discontinuing these 14 products next month. Find every asset featuring them and tell me everywhere they're currently being used." That question cuts straight across DAM, PIM, ecommerce and CMS.
Which makes me wonder whether the future architecture is less ASSETS → DAM → EVERYTHING ELSE, and more:
ASSETS
│
▼
ASSET INTELLIGENCE
LAYER
│
┌───────────────┼───────────────┐
│ │ │
HUMANS SYSTEMS AGENTS
│ │ │
DAM/UI CMS / PIM AI / MCP
│
ACTIONS
That doesn't mean the DAM becomes irrelevant. Quite the opposite. Governance, permissions, approvals, transformations, auditability, rights management and controlled distribution aren't suddenly going away because we've got an LLM. But perhaps the role of the DAM changes. Perhaps it becomes the governance and management layer over an organisation's asset intelligence rather than the container where all of that intelligence has to live.
/06Why it might not work
There are plenty of reasons this might not work, and they're important. If the intelligence surrounding an asset becomes portable, somebody still has to decide which system is authoritative. AI-inferred metadata can't quietly become organisational truth simply because a model sounded confident. Rights information may be highly sensitive and certainly shouldn't automatically travel with every publicly distributed version of an asset. Relationships change. Products get renamed. Campaigns end. Licences expire. Embedding models change. Organisations have millions of existing assets and decades of technical debt.
Those problems aren't details to solve later. They're fundamental to whether an architecture like this could actually work. But I think that's precisely why it's worth exploring.
/07The uncomfortable question
There's a tendency at the moment to take existing enterprise software, add semantic search, bolt on a chatbot, expose some MCP tools and call the result AI-native. Sometimes that's absolutely the right thing to do. But every now and again I think we should ask the slightly more uncomfortable question:
If we knew AI agents were going to be first-class users when we originally designed this system, would we have designed the system the same way?
With DAM, I'm increasingly convinced the answer is no.
We'd probably care less about forcing humans to describe every asset perfectly and considerably more about identity, provenance and relationships. We'd distinguish much more clearly between human-verified information and machine inference. We'd assume assets would be consumed by software we hadn't anticipated when they were created. We'd make machine-readable access fundamental rather than something added later through an integration project.
And above all, I think we'd spend much more time considering whether the intelligence surrounding an asset should belong exclusively to the application managing it.
So, if we invented Digital Asset Management today, would we still build a DAM? I think we probably would.
I'm just not sure we'd put the DAM in the middle. And perhaps that's the more interesting question for what comes next.