As the Immediate Past Chair of AIIM, I have watched our industry rename and reinvent itself a couple of times. We went from the Association for Information and Image Management to the Association for Intelligent Information Management. Most recently, we rebranded our events as the AI+IM Global Summit and AI+IM regional events.
I believe we need an additional evolution: from information management to aspect management. This is my last year on the board and I am not sure if the change will happen during my tenure, but I predict this industry change will happen.
That is a big claim, so I won’t ask you to take it on faith. Let me take you through it step by step: how our expectations of search have changed, why our information was never organized for what we now ask of it, why the usual fixes fall short, and what becomes possible once we solve it.
Twenty-five years ago, finding a document at work meant advanced search. You picked the repository, the document type, the department, a date range. You filled in fields. It worked if two things were true: the document had been filed with the right metadata, and you knew which fields to fill in. Searching was a skill, and information managers taught it.
Then Google changed what we expect. One box. Type what you mean, and the best result comes first. Nobody fills in fields anymore. Nobody needs to know how the internet is organized.
Now AI has changed what search can be. It understands the intent of a question, reads what it finds and writes an answer. We followed: we don't type keywords anymore; we ask a full question in plain language, and we expect an answer, not a list.
At home, that works. At work, the same person asks the same kind of question of the company’s AI assistant and gets an answer from last year’s version of the policy. Or waits while the assistant thinks, and thinks again. Same kind of question, same kind of AI. What is different?
Modern AI does understand the intent of a question. But between the question and the answer sits the search, and search keeps very little of that intent. Keyword search sees words. Vector search compresses the whole question into a single numerical representation of its meaning, where a model number or a version barely registers. Hybrid search blends the two using fixed weighting rules. As a result, the AI understands what you mean at the start and tries to recover it at the end by checking the results, but the search in the middle has already decided what it gets to see.
The problem is not the AI model. It is how our information is organized.
Information hoarding or information silos are a significant obstacle to AI success. Without access to relevant data, it’s impossible to train models or execute tasks. Traditionally, we blamed information silos on different systems or expansive technology stacks.
That's half the story. The other half is how we organize information itself. The silo is not only between systems. It sits inside every one of them. We organized information in three separate places:
Content in one place: what a document says.
Metadata in another: what we know about it, such as its type, version, date, status and owner.
Context somewhere else, if it was captured at all: how a document connects, such as the product it belongs to, the case it refers to or the contract it amends.
We never noticed this silo, because people bridged the gap themselves. A colleague who has done the job for ten years knows that an offer may sit in the CRM, the signed contract in the document management system and the drafts in SharePoint. Search didn’t have to know. People did the joining in their heads, and when something could not be found, we blamed the salesperson’s record-keeping.
That is the silo we never saw.
AI doesn’t have ten years of experience. It cannot do the joining in its head. It does exactly what our systems allow.
The result of this is that we search in separate steps. Search ranks the content, by keywords and by meaning. Advanced search filters on metadata: in or out, never weighed. And where there is a knowledge graph, it holds the relationships separately again. The results are merged by rules of thumb and cut off at a shortlist, and then the AI checks what made it through. Users often describe enterprise AI assistants as so slow they feel like they are crashing. They aren’t crashing. They are checking.
Every round of checking is a model call, and every model call costs time and money. Worse, no amount of checking can rescue a document that the search never returned.
There is a second cost. Because three separate results are merged into one score, retrieval has become a black box. Nobody can see why a document was chosen, so nobody can steer it. You can only accept the answer or check it yourself. In practice, the AI is in control of the organization, instead of the other way around.
This is not the problem of any single vendor. It is the structure of how we have organized information. AI is simply the first user that cannot bridge the gap in its head.
A common piece of advice is to clean our content first. Then, the argument goes, AI works and can be safely implemented in the enterprise. In practice, cleanup projects rarely finish. And after cleaning, it also has to stay clean. That requires culture change and training programs on correct metadata entry. And after all of this, it still requires users to write perfect prompts. Twenty-five years in the information management business tells me this is not a very likely scenario.
Cleaning is not wrong. But it should be a choice, not a precondition forced on every AI project. And as AI generates a growing share of new content and new versions, content does not settle the way it used to. A cleanup project is now chasing a moving target. And it treats the wrong problem. The problem was never the quality of our data. It is the silo we never saw: content, metadata and context kept apart, and joined only in people's heads.
A second common fix is to write business rules. If the AI cannot tell which version is right, tell it: prefer the signed contract, prefer the latest procedure, exclude drafts.
The trouble is that relevance depends on the situation. Sometimes latest matters more than signed. Sometimes signed matters more than latest. A rule is written before the question is asked; the right answer depends on the question. And who decides the rule? A programmer does not know how a sales team stores its draft proposals. Knowledge management has long struggled to turn tacit expert knowledge into explicit rules, and AI has not changed that.
So the rules pile up. When an assistant becomes slow, it can be a sign of how many rules have been stacked on top of each other, and they may well have started to conflict.
So here is where we are. Our expectations come from consumer search engines and AI assistants. Our information is organized for advanced search and for colleagues with ten years of experience. And the two fixes on offer either never finish or never fit the situation. Something has to change in how we organize information itself. Not cleaned, but reorganized for the way AI reads.
What is an aspect? An aspect brings a document's content, its metadata and its context together in one fingerprint. As a result, a single ranking takes everything into account at once, and each aspect keeps its own score. Because every aspect is scored on its own, IT can see why something was retrieved and decide what gets out. And the question does not need to be perfect. The question is also translated into aspects that are weighed against the aspects of the answer.
So what would information look like if we organized it for AI instead of for advanced search? I think of it as a ladder. You know the first three rungs:
Data is what a document says. Search finds it.
Metadata is what we know about it. Advanced search filters on it.
A graph is how it connects. Graph search follows it.
Aspects bring all three together in one fingerprint, so a single search weighs them at once.
An aspect is one dimension of what makes a document the right one: its meaning, its type, its version, its status, the product or case it belongs to. Together, a document’s aspects form its fingerprint.
Two things make this different from what we have today. First, aspects are weighed, not filtered. Advanced search says in or out. Aspects say how much each dimension counts and how close each document comes, so latest can matter more than signed for one question, and less for another. And “latest” becomes a distance rather than a rule: yesterday scores higher than last week, instead of a hard cutoff such as “newer than 30 days.” Second, every aspect keeps its own score, so you can see exactly why a document came first.
This is not another search. It is an additional index, and that is the real shift: aspect management is a layer between your content and everything that reads it. Your content stays where it is. Your search, your filters and your graph keep doing what they do best, and can use the aspects at the point where today a reranker or an extra round of model calls tries to sort out the results. A search can also run on the aspects directly. Either way, search and AI run on top of the layer.
Take a service engineer who types: “Printer P7 shows error 49 after the update.”
Ten service bulletins mention error 49. To a search on meaning, all ten look equally good: they are all about the same error. Only one is right: the bulletin for the P7, written for the firmware that was just installed, and actually released.
Here is what aspect search sees. The question is turned into aspects too, so it gets a fingerprint of its own:
The right bulletin matches the question on every aspect. The similar one matches on content and type, but falls away on model, firmware and release status. Aspect search puts the right one first, and the chart shows why.
Notice what the engineer did not have to do. Today, to get this precision, you need advanced search: you have to know there is a model field, a firmware field and a status field, and fill in all three. With aspects, the engineer just typed the question, the way you would in any AI chat. The model number, the firmware and “after the update” were turned into aspects, and they did the selecting. The intent survives because it is carried into the search itself, instead of being lost in it.
That closes the gap described at the beginning of this post: ask the way you ask any AI chat, and get the precision of advanced search.
Finding the right document is only the beginning. Once content, metadata and context are joined, and every aspect keeps its own score, five things become possible.
Validation. Today, when an AI assistant answers, you have two options: believe it, or check it yourself. Most people check, so they double-check every answer, including the right ones, and the time AI was supposed to save is gone. With aspects, every answer comes with its evidence: which document it used, and why that document was chosen, aspect by aspect. The engineer sees in a second that the bulletin is for the P7, for the right firmware, and released. An answer you can validate is an answer you can trust. In my experience, that is the step that turns an AI pilot into something people actually use.
Transparency. Retrieval stops being a black box. Today, retrieval typically produces a single relevance score that is hard to explain. With aspects, IT, compliance and auditors can see why a document was retrieved and explain a decision after the fact. For regulated organizations, that matters.
Steering. Because aspects are weighed, they can be adjusted. For contracts, signed should count more than latest. For finding a similar proposal, latest should count more than signed. Instead of writing yet another business rule, you adjust a weight and see the effect immediately.
A feedback loop. Suppose the engineer looks at the top two bulletins and picks the second one, because for this printer a released bulletin for the previous firmware is more reliable than a brand-new one nobody has checked yet. That choice is not lost. It shows which aspect made the difference (release status over latest firmware), and the weights can be adjusted accordingly. The next engineer with the same kind of question gets that bulletin first. Experts often cannot write down how they judge, but they show their judgment every time they pick an answer. Capture the picks, and you capture the judgment.
Aspects across repositories. Remember the silo between systems? Aspects work across it. A contract in your ECM, its amendment in SharePoint and the related case on a file share can be weighed together in one search, without moving any of them. You don’t have to migrate everything into one place to get one answer.
There is one more benefit. Because aspects are weighed mathematically rather than through another round of model calls, the AI should need fewer LLM calls to reach the right document, which can reduce both response time and cost.
For our profession, this is good news. AI is not making information management obsolete. It is giving it a new object to manage.
Today most organizations have two layers for their content. Collaboration platforms, such as SharePoint, are where people work together. Enterprise content management is where workflows run and records are stored compliantly. Both stay. What is missing is a third layer: aspect management, the layer between your content and your AI, providing quality control and management for what your AI reads.
And who better to own that layer than the people who have always cared about what information means, where it belongs and who may see it? Deciding which aspects matter, keeping them current, setting the weights, governing what gets out: that is information management, applied to AI. It is also how an organization takes back control of its AI, instead of letting the AI control the organization.
At Aspected, the company I founded together with my co-founder Jorn Verhoeven, we are launching the Aspect Management System. It puts a layer between an organization's content and its AI: it builds the aspects from content, metadata and context into one index, keeps them current and lets information managers govern them. Search, Copilot and agents run on top of it, linked through the Model Context Protocol (MCP). Content stays where it is. Xillio, the company I founded in 2004 and still lead, is the launch implementation partner. The Aspect Management System is available on Azure, AWS and as a sovereign deployment. Our short version: AI-ready data isn't clean data. It's aspected data.
No single company can make this shift. It needs the whole profession. So let me end with the questions I would like us to debate:
1. If AI writes more and more of the versions, what is the record?
2. Who in your organization owns relevance: IT, the business, or nobody?
3. Why do we still treat metadata as a gate, in or out, when AI works on probabilities?
4. How many of your AI’s model calls are spent checking a search that could have been right the first time?
5. If your best expert corrects an AI answer today, where does that correction go?
The on-premises era gave us information and image management. The cloud era gave us intelligent information management. AI reads differently. AI needs aspect management.