Guardians of the Record: What Happens When AI Enters the Archive

A few weeks ago, I stood in a hotel meeting room in Manchester and told a room full of records managers that the job they have been doing for decades is about to change in ways most of them have not fully processed yet.

That is a strange thing to say to a group of professionals who have spent their careers quietly ensuring that the pharmaceutical industry does not lose its memory. Archivists and records managers are not typically on the receiving end of disruption talks. They are the ones who clean up after disruptions. They file the lessons learned. They preserve the audit trails. They are, in a very real sense, the guardians of the record.

But here is what I could not stop thinking about on the flight home: the archive itself is broken, and AI is both the most promising fix and the most dangerous accelerant of that brokenness.

The dark data problem nobody talks about

Somewhere between 60 and 73 percent of enterprise data goes entirely unused. That number comes from Forrester, and it has held roughly steady for years. In pharma, the situation is arguably worse because the data is not just unused, it is locked in formats, systems, and storage conditions that make it structurally inaccessible.

We are not talking about obscure legacy files. We are talking about stability studies, deviation reports, batch records, regulatory submissions, clinical trial documentation spanning decades. Information that was generated at enormous cost, validated under strict regulatory frameworks, and then effectively buried.

The archive was supposed to be the institutional memory. In practice, it became a warehouse or even worse a cemetery.

What AI actually promises

The pitch for AI in records management is compelling because it addresses real pain. Automated classification and metadata enrichment. Intelligent search across unstructured document sets. OCR that actually works. Anomaly detection that can flag inconsistencies across thousands of documents faster than any human team.

One speaker at the conference described a case where a 30,000-document SharePoint archive had grown to 300,000 documents through successive migrations. Nobody could find anything. Nobody trusted the system. AI-driven cleanup, classification, and deduplication turned it from a liability into a functioning resource.

That is not hype. That is a legitimate transformation from document graveyard to living intelligence source.

The problem is that none of this comes without new risks that the industry has barely begun to think through.

Three risks that do not make it into vendor demos

The first is hallucination in high-stakes contexts. When an AI system classifies a document incorrectly, enriches metadata with inferred information that is wrong, or surfaces a “related” record that is not actually related, the consequences in a GxP environment are not a minor inconvenience. They are a data integrity failure. ALCOA++ does not have a carve-out for “the AI was confident.”

The second is what I call the deletion agent problem. AI systems tasked with archive cleanup, deduplication, and migration will inevitably make decisions about what to keep and what to discard. In a regulated environment, every deletion decision is a potential audit finding. The question is not whether AI can identify duplicates. The question is whether the governance framework around that AI is robust enough to ensure that what gets deleted was genuinely redundant and that the decision is fully traceable.

The third is context collapse. Archives preserve documents, but they also preserve context: why a deviation was handled a certain way, what the regulatory landscape looked like at the time a decision was made, what constraints the team was operating under. AI systems that surface documents without preserving that contextual layer risk turning rich institutional knowledge into decontextualised data points. That is not intelligence. That is noise with a confidence score.

The oversight trap

During my keynote, I made a point that visibly unsettled a few people in the room. Human-in-the-loop is the industry’s favourite answer to AI governance, but it is an incomplete one.

Consider the scale problem. If an AI system processes 10,000 documents per day and flags 200 for human review, the human reviewer is not actually reviewing the other 9,800. They are trusting the AI’s judgment on those. That is not oversight. That is delegation with a governance label.

It gets worse. When AI highlights the issues it finds, the reviewer’s attention narrows to those highlights. If the AI misses something, the odds of the human catching it drop dramatically. I call this AI blinders: the paradox where AI-assisted review can actually reduce the quality of human oversight by changing how humans allocate attention.

The lane-assist analogy works here. When your car keeps you in the lane, you stop paying the same attention to steering. When the system fails, your response time is worse than if you had been driving manually. In regulated environments, that gap between nominal oversight and actual oversight is where audit findings live.

The intermediate layer

So what do we do? Stop using AI in archives? Obviously not.

The architecture that makes sense is what I described as a curated buffer zone. The principle is straightforward: never let AI operate directly on your source records. Instead, create an intermediate knowledge layer, a curated, governed copy where AI can classify, enrich, search, and flag. The source archive stays air-gapped. The intermediate layer is where intelligence happens. Changes flow back to the source only through validated, human-governed processes.

This is not a novel concept. It is how any well-designed system handles the tension between access and integrity. But in the rush to deploy AI in records management, this architectural discipline is often the first thing that gets compromised.

The question nobody has asked yet

Here is where I ended the keynote, and where I will end this piece.

We have spent decades building frameworks for archiving human decisions. ALCOA tells us how to ensure that data is attributable, legible, contemporaneous, original, and accurate. Every regulatory inspection assumes that somewhere in the archive, there is a traceable chain of human judgment.

But AI does not make decisions the way humans do. A large language model does not have a “rationale” in the way a deviation investigator has a rationale. A classification algorithm does not have “intent.” A recommendation engine does not have “judgment.” These systems produce outputs that are probabilistic, context-dependent, and impossible to reproduce exactly.

So when AI starts making decisions that affect regulated records, how do we archive that? What does ALCOA++ look like for a system that gives a slightly different answer every time you ask? What does “contemporaneous” mean when a model’s parameters change between Tuesday and Wednesday?

The archive has always been the place where we preserve the evidence of how decisions were made. If we cannot archive AI’s decision-making process with the same rigour we apply to human decisions, then we have a governance gap that no amount of human-in-the-loop theatre will close.

The regulators have not asked this question yet. When they do, the organisations that have already thought it through will be the ones setting the standard. The rest will be scrambling.

That is not a prediction. That is a pattern. And if you work in records management, quality, or regulatory affairs, this is the conversation worth having before it becomes an inspection finding.

---

This piece is adapted from a keynote delivered at the HSRAA 2026 Annual Conference in Manchester, “Guardians of the Record: How AI Can Help (and Hurt) GxP Archives.” The views are my own.