Most AI scoping exercises produce some version of the same list. The document management system. The CRM. The policy or practice management platform. SharePoint, email, perhaps a data warehouse. It’s a sensible list. In wealth management, insurance and legal, it’s also an incomplete one, because it rarely includes the place where many of the oldest and most authoritative records still live: the boxes in offsite storage.
That isn’t carelessness. Data discovery tools crawl file shares and databases. They don’t crawl warehouses. Paper only appears on a data map if someone deliberately puts it there. And this isn’t purely a legacy problem either. AIIM’s 2025 survey of 600 enterprises in the US and Europe found that 61% of document processing workflows still include paper[1].
No record you can see is not the same as no record
Anyone who has worked in an established firm for a while knows the phrase “check the archive”. When the system shows nothing before 2006, experienced staff understand that the system starts in 2006, not that nothing happened before it.
AI doesn’t know that unless it’s told. An assistant connected only to digital repositories can’t tell the difference between “there is no record” and “there is no record I can see”. It answers from what it has, and it usually answers with confidence.
For a firm whose digital history begins partway through its actual history, that isn’t an edge case. It’s a blind spot built into every question that reaches back far enough. Your AI doesn’t know your organisation. In this case, it doesn’t even know where its own knowledge stops.
The oldest records often carry the most weight
This would matter less if the boxes held trivia. In these sectors, they often hold the opposite: original, signed, authoritative documents that exist nowhere else.
In legal practice, original wills are the obvious example. The Law Society’s view is that an original will should be stored until after the client’s death or until it can be returned, noting that some firms keep wills indefinitely or for fifty years, and that a prior will might need to be proved if a later one is challenged[2].
In wealth management, the FCA requires suitability records relating to pension transfers, conversions, opt-outs and FSAVCs to be retained indefinitely[3]. A firm that has been advising for decades will hold some of that evidence from before its current systems existed.
In insurance, the long tail is very long. Most mesotheliomas develop more than 35 years after first exposure to asbestos[4], while the central Employers’ Liability Database only collects policies new or renewed from April 2011, plus older policies with claims made since then[5]. The wordings and schedules that establish who was on risk decades ago may survive only on paper.
Historical data is institutional knowledge. In these sectors, it’s also evidence.
A hypothetical: the advice that “didn’t exist”
Here’s a hypothetical. A wealth management firm receives a complaint about pension transfer advice given in the late 1990s. To speed up its response, the complaints team uses an AI assistant connected to the CRM and to client files scanned since the late 2000s. The assistant finds the client’s current holdings and recent reviews, and reports no record of the original advice.
The firm does hold the suitability record. It’s obliged to keep it indefinitely, and it has. The file is in a box in offsite storage, indexed by client surname range and year. Nothing connected that index to the assistant.
If the team takes the answer at face value, it responds on the basis that the firm can’t evidence its advice. The firm wasn’t non-compliant. It was simply uninformed by its own tools.
Compliant, but invisible
Physical archives aren’t outside the rules. UK GDPR covers personal data held manually as part of a filing system[6], and many offsite archives are well run: secure, barcoded and retention-coded at box level. But compliance means the records are kept properly. It doesn’t mean anyone, or anything, knows what’s in them. An archive can pass every audit and still be invisible to every AI initiative in the building.
Compliant does not mean AI-ready.
Scope the index, not just the contents
The instinctive response is to scan everything. It’s usually the wrong one: slow, expensive, and likely to digitise plenty of material that should have been destroyed instead.
A more useful first step is to bring the archive’s index into scope, even while the contents stay on paper. If an AI system knows that a client, policy or matter has records in storage – which boxes, what date range, what kind of documents – it can say so. Instead of “no record found”, it can answer: “Nothing in the digital record before 2008; there are six boxes in storage for this client covering 1994 to 2007.” A false negative becomes a retrieval request.
That depends on the quality of the index, and this is where most archives need attention. A box description such as “Commercial Lines – Renewals 1974–1979” is good enough to manage retention, but not good enough to find a particular policy. Where a use case justifies it, improving indexing to file level is often far cheaper than digitising the contents, and it makes the archive findable straight away.
[Damien: a short practitioner observation would land well here – for example, what box-level indexes in legacy archives typically do and don’t tell you.]
From there, digitisation becomes a series of targeted decisions rather than a single programme. Some files earn their way into the digital estate because a use case needs them regularly. Others are retrieved and scanned only when they’re needed. Retention decisions come first throughout, and originals that must be kept as originals stay where they are.
Most people already accept the principle. In Iron Mountain research with 700 IT and data decision-makers, 96% agreed that a unified strategy for managing digital and physical assets is critical to the success of generative AI[7]. The test is whether the boxes actually appear in the plan.
How Dajon approaches it
At Dajon, we’ve spent nearly thirty years storing, indexing, retrieving, digitising and securely destroying physical records for firms in wealth management, insurance, legal and property. We help organisations put their physical archives back on the data map: improving indexes so that what’s in storage can be found, connecting that knowledge to the systems and AI tools that need it, digitising the parts that earn it, and making defensible retention and destruction decisions about the rest.
The paper still counts. The question is whether your AI strategy is counting it.
Can you really have an AI strategy without a historical data strategy?
References
- Market Momentum Index: IDP Survey 2025 AIIM[↩]
- How Long Should a Client’s Original Will Be Retained? The Law Society[↩]
- COBS 9.5 Record Keeping and Retention Periods for Suitability Records FCA Handbook[↩]
- Past and Current Asbestos Exposure and Future Mesothelioma Risks in Britain International Journal of Epidemiology via PMC[↩]
- The Employers’ Liability Tracing Office (ELTO) AXA Connect[↩]
- What Is Personal Data? ICO[↩]
- Iron Mountain Transforms the Information Experience With New Digital Experience Platform Iron Mountain[↩]