In a widely-shared ARMA webinar, Kaan Volkan made this clear warning: AI large language models, applied without adequate information architecture, fall dangerously short for records programs - and what that can cost organisations that fail to heed the warning, including falsely destroyed records. His presentation title said it plainly: “LLMs Destroyed My Records.”
I am not surprised. Forty years in this work, and I've watched this pattern repeat too many times to count - but never at this speed, and never this final. The stakes have never been higher. These records could be wrongfully destroyed, and they are not merely operational assets. Often they are the evidence trail of decisions made by public institutions - institutions accountable to the people they serve. Destroy those records, and you destroy the ability to hold power to account. That is not a records management problem. It is a democratic one.
This is the argument I am making in my essay, Records, Accountability and Democracy. My central contention is this: records management is not an administrative discipline. It is the cornerstone of democracy. Right now, we are failing it.
The frictionless trap
Think about what democratic accountability actually requires. A citizen exercises their right to information. A parliamentary committee calls for the papers. A royal commission follows the thread of a decision back to its origins. An auditor reconstructs how a contract was awarded. Every one of these processes depends - entirely and without exception - on records having been made, kept, and preserved in a form that is authentic, reliable, complete and accessible. When records fail, accountability does not merely become difficult. It becomes impossible.
The paradox of our era: we have never created more information, and never had a harder time finding it. I call this the frictionless trap. When creating a document costs nothing, when saving or sharing a file takes one click, and when every message is automatically retained somewhere, then the natural response is to create more, save more, share more, and organise nothing. A single procurement decision may now be scattered across Teams messages, Outlook threads, SharePoint documents, OneDrive folders, Salesforce records, SAP entries, and mobile phone photos - with no authoritative record of what was decided or why.
“An un-auditable system of government is not a democracy in any meaningful sense. It is a managed fiction.”
We have seen this play out with disturbing regularity: ministers who cannot produce documents showing how decisions were made, procurement processes whose evidentiary trail has been scattered across personal email accounts and encrypted messaging apps, public health deliberations that are simply unrecoverable. These are not administrative failures. They are democratic failures - and they are being produced, in large part, by the collapse of records management discipline in the digital age.
Compounding the technical fragmentation is a cultural one that I think we underestimate. Staff in contemporary digital workplaces rarely think about evidentiary completeness - they think about getting work done. As AI is increasingly involved in decision-making, this gap will widen. Records that would explain how an AI-assisted decision was reached may simply never be created, leaving organisations unable to account for choices that may have profound consequences for the people they serve.
Architecture first - always
The profession’s response to fragmentation has been AI autoclassification and manage-in-place governance. Both are genuine advances, and I am not dismissing them. However, when adopted as complete solutions rather than carefully bounded components of a broader strategy, they can make the democratic accountability problem significantly worse than the one they set out to solve.
Research from Cornell University found that AI systems hallucinate at rates of 69 to 88 percent on complex tasks, and up to 50 percent on relatively simple ones. In a records context, this is not mere operational inefficiency. An AI that mis-classifies a document - applying a short-term retention schedule to a high-value record, or failing to recognise that a cluster of emails constitutes the evidential record of a significant decision - is potentially destroying the very evidence that democratic accountability will one day demand. The AI does not know it is dismantling an audit trail. But that is precisely what it does.
The root cause is almost always the same: AI was applied before the information architecture was in place. I experienced this directly while developing an architecture for a government agency. I persuaded them to build information categories around how people actually worked, rather than using inherited records classification schemes that staff neither recognised nor used.
AI could then be training to work with the grain of the business - and, critically, reflect the actual shape of the decisions the organisation was making and needed to be held accountable for. The organisation’s IT team subsequently declared the approach ‘too complex and old school’ and instructed their AI to classify existing unstructured content instead. Very little progress followed.
“A government that claims its records are managed by AI - but whose information architecture has never been analysed - has not improved its accountability. It has substituted the appearance of control for the reality of it.”
The same failure mode applies to manage-in-place governance more broadly. Tools like Microsoft Purview offer real advantages; they scale, reduce friction, and integrate with the platforms that people actually use. On the other hand, they were not designed to maintain the contextual integrity that high-value, accountability-significant records require.
The context of the aggregation of related items across systems, defensible proof of disposal, and the full evidentiary record of a complex decision. A global IG World survey found Purview adoption for governance had actually declined - from 50 to 40 percent of organisations - with 94 percent reporting significant challenges.
That is not a failure of education. It is practitioners arriving, through experience, at the same conclusion: manage-in-place, taken on its own, is not adequate for the kind of defensible record-keeping that public accountability demands.
The hybrid answer
My answer is not the centralised EDRMS model or outsourcing records management to applications. Nor is it pure manage-in-place. It is a deliberate hybrid: the right model for the right record, matched to the accountability that record carries. Years of implementation and reading brought me here - not theory alone.
High-value records - those that document decisions, transactions, and the exercise of power - are centralised in a purpose-built EDRMS with defensible lifecycle management. Repeating transactional processes can embed records management at the point of creation. Routine, low-risk content is governed in place, with the limitations of in-place tools understood and mitigated.
In all cases, information architecture must come first. Not as preliminary work before the records management strategy but as the core strategy itself. Let me be emphatic about that.
Get it right, and AI classification becomes a powerful, scalable tool: governance applied at the volume modern organisations demand. Get it wrong, or skip it, and AI embeds the records problem at speed and scale, under a false veneer of automated compliance.
The accountability gap doesn't close. It widens, invisibly, behind a dashboard that says everything is in order.
Why I’m writing this now
I am reflecting on this at a moment when the world is producing governments and institutions that are increasingly comfortable with opacity. They are content to operate in environments where the records of how decisions were made is inaccessible, incomplete, or simply non-existent. AI will accelerate that tendency if we let it. It will not, on its own, reverse it.
In my full essay, I set out six principles for practitioners and make the case, with the conviction of forty years, that records management is one of the most important things that a democratic society can invest in.
It isn't glamorous and it rarely gets its due. Records managers are the custodians of institutional memory. Institutional memory is the currency of accountability. Without it, the promise that power can be seen, examined, and held to account is nothing but mere words.
Records, Accountability and Democracy: Reflections After Forty Years is published in full on my blog.
Nigel Carruthers-Taylor is the retired founder and Principal of iCognition, acquired by Kapish in April 2026. He is the author of a series of practitioner articles on hybrid records management, AI in records practice, and the use of Microsoft Purview and EDRMS in the Microsoft 365 environment, published in IDM and elsewhere.