Last week I was in Baltimore for the AIIM AI+IM Global Summit. Three days. Hundreds of information leaders from across the world. Box, IBM, M-Files, OpenText, Hyland, and the rest of the global IM community in one room.
I came back with more notes than I can process in two weeks.
But one conclusion came back in every session, in every hallway, in every coffee break:
The pilots are over.
AI is in production.
And the gap between the winners and the strugglers has almost nothing to do with which model they picked.
It comes down to foundation.
In this article I bundle what I heard, saw and learned in Baltimore. Not as a summary of the full programme. But as a reflection on what it means for the BeNeLux SMEs we work with. And on a shift that was carefully introduced from the AIIM main stage, which in my view is the most fundamental movement in our field in ten years.
From trial to deployment
Until last year the AI conversation was mostly about models. Which one is the best? Which one is the fastest? Which one costs the least per token? Who wins the benchmark?
In Baltimore that conversation was gone.
It was no longer about models. It was about operational reliability. About traceability. About governance. About integration of inference into real workflows. About what happens when an AI agent makes the wrong call, and who is accountable for it.
That is a fundamental shift. And it changes the nature of the conversation a CEO needs to have about AI in their organisation.
The core: structure, not intelligence
The line that came back across dozens of sessions in Baltimore, in different formulations, was this:
Structure, not intelligence, defines reliability.
AI is only as good as your data. Put AI on top of bad data, and you start making really bad decisions really fast.
The number that came back everywhere: 80 to 90 percent of all enterprise data is unstructured. PDFs. Word documents. Emails. Spreadsheets. SharePoint sites nobody owns anymore. Shared drives from 2014 with permissions nobody can explain.
That is the feedstock for any AI worth running. And in most BeNeLux SMEs, that feedstock is a mess.
You can put the most sophisticated model on the planet on top of it. The output stays unreliable. Not because the model is bad, but because the input is.
A new term, a bigger shift
In Baltimore, a new term was carefully introduced from the main stage. A term that on first hearing sounds like nothing more than a rebranding, but which on closer inspection is a fundamental redefinition of our field.
No longer information management.
Context management.
The shift is not a play on words. It is structural.
Information management was about the document. Where it lives. Who owns it. How long we keep it. What metadata it carries. What the retention period is.
Context management is about something else. It is about what the document means at a specific moment, for a specific decision, for a specific system, human or agent. It is about the rules that apply at that moment. About who is authorised to act on the document. About why the AI is using this document right now and not another.
Because here is the critical point that came back in Baltimore again and again: AI agents do not read documents. They consume context. The right document in the wrong context produces a confidently wrong answer. The wrong document in the right context produces a quiet compliance breach. The document on its own, without context, produces nothing useful at all.
That changes what the information manager does. Or more precisely: it changes who the information manager has to be.
The old question was: is this document classified, retained, accessible? The new question is: does the AI know what this document means, when it can be trusted, who is allowed to act on it, and under which rules?
Same data. Different job. Higher stakes.
The information manager managed records. The context manager manages meaning.
That is not a title change. That is a seat at a different table. An architectural role. A strategic role. A role that sits at board level, not in the archive room.
What this means for BeNeLux companies
At this point all of this might sound like a story for multinationals. For Shell, for Prudential, for the organisations on the AIIM stage.
The opposite is true. For BeNeLux SMEs this is more urgent, not less.
A few reasons:
- An SME of 80 people does not have the luxury of a three-person AI Governance Office. There is usually nobody who explicitly owns the AI being used in the organisation. Meanwhile Copilot licences are running, employees are using ChatGPT on personal accounts, AI is embedded in the SaaS applications HR and finance work with every day. Nobody has it mapped. Nobody has the mandate.
- The compliance reality is coming relentlessly. The EU AI Act is in force. GDPR has not gone anywhere. ISO/IEC 42001 is increasingly being asked for by corporate clients in their vendor reviews. The first question an auditor asks is always the same: show me your AI inventory.
- Most BeNeLux SMEs cannot produce one. Most do not yet know they need one.
And at the same time: the productivity pressure is real. Employees look for tools that work. They find them. They use them. Whether management likes it or not. According to recent research, 59 percent of employees use AI tools their employer has not approved. 75 percent of those have already pasted sensitive company data into those tools.
This is not a security problem. This is a context problem. And it grows every day.
What to do
Three moves every BeNeLux SME could consider over the next three months. Based on what I heard in Baltimore and on fifteen years of work in this field.
- One. Stop treating AI as a project. Start treating it as architecture. A Copilot rollout is not an AI strategy. An agent pilot is not governance. The question is not which model you pick. The question is which layer you build so that all the models coming over the next five years can run reliably on top of it.
- Two. Appoint a context manager. Or admit you do not have one. The role probably does not exist on your org chart. Not the IT lead. Not the DPO alone. Not the records officer. Someone has to hold the map of what AI is in use, what it touches, what it decides, and under which rules. In an SME this can be a half-time function within the management team. But it has to sit explicitly with someone.
- Three. Start with one process, not with everything. Pick one business process. A process that matters. Contract management, customer onboarding, HR onboarding, take your pick. For that one process, map out what AI is involved, which documents are flowing, which decisions are being made, and which rules should apply. Fix the foundation there. Then let the agents loose. Then scale.
In closing
The work we have been doing at SoftAdvice for fifteen years under the banner of Process & Information Architecture was reformulated on the AIIM main stage as the strategic layer under every serious AI implementation.
That is not a coincidence. The professional community has grown into the same conclusion. Not because AI invented new disciplines, but because AI is finally making visible what we have known for a long time: without structure, without context, without governance, no intelligent technology will run reliably.
We are ready for it.
More than that. We have been quietly building something specific for the BeNeLux SME market. An approach that is fixed in scope, limited in duration, and built on the regulatory reality our clients operate in.
More on that soon.