For the past few years, most enterprise AI has been in the business of answering. A member of staff asks a question, the AI finds some information and drafts a response, and a person decides what to do with it.
That’s changing. The current wave of AI is designed to act: to work through multi-step tasks, use other systems, update records and send communications with less and less human involvement along the way. Gartner predicts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from none in 2024[1].
Most of the conversation about that shift is about the AI itself: what it can be trusted to do, how it should be supervised, who is accountable. Far less attention goes to what it is acting on. When AI starts acting, the information it draws from stops being reference material. It becomes the basis of decisions. In a very practical sense, your archive starts deciding.
The human in the middle was doing more than we realised
When AI only answers questions, there is a person between the answer and its consequences. That person does more than approve or reject. They notice when something looks wrong.
A claims handler reading an AI summary knows that the policy wording it quotes was replaced two years ago. An underwriter knows that the risk survey it cites relates to premises the client has since sold. A complaints handler knows that the letter it drafted refers to a product that was renamed. None of this is written down as a control. It’s simply what experienced people do when they read.
In effect, that person has been quietly acting as a data quality check on the archive. Every time they caught an outdated document, a misfiled record or a contradiction between two files, they corrected the AI’s output before it reached a customer.
Agentic AI removes some of that checkpoint, by design. That’s where the efficiency comes from. But it means the weaknesses in historical information that people used to absorb now flow straight through into actions.
A hypothetical: the endorsement that was never cancelled
Here’s a hypothetical example of how that can play out.
An insurer deploys an AI agent to handle routine renewals on a commercial lines book. For each case, the agent reviews the policy history, checks the current terms, prepares the renewal and issues the documents. A sample is reviewed by underwriters, but most renewals go out without a person reading them.
In one client’s file, an endorsement adding an exclusion was cancelled several years ago. The original endorsement is a clean, well-indexed document. The cancellation is a scanned letter, filed separately, with poor text extraction and no link to the endorsement it reversed. To the agent, the endorsement is clearly present and nothing clearly cancels it. The renewal goes out with the exclusion reinstated.
No one notices until the client makes a claim that the exclusion would affect.
Every step the agent took was reasonable given the information it had. Nothing about the model failed. The error was in the archive, and it was carried into a binding customer document because no one was reading in between.
An experienced underwriter might have remembered the cancellation, or questioned an exclusion that didn’t fit the risk. The agent had no way of knowing it was missing anything.
Acting raises the bar for historical information
An AI that answers questions needs to find relevant information. An AI that acts needs considerably more from the records it relies on:
- Currency: Is this document still in force, or has it been superseded, cancelled or replaced?
- Authority: Is this the binding, approved version, or a draft, copy or working note?
- Linkage: Is the record correctly tied to the right customer, policy or matter, and to the documents that amend or reverse it?
- Permission: Is this information appropriate to use for this purpose, and on whose behalf is the agent acting?
- Traceability: If the action is later challenged, can the firm show exactly which records drove it?
Many historical estates were never built to answer these questions explicitly, because people answered them implicitly. For an AI that only advises, gaps here produce poor answers that someone may catch. For an AI that acts, they produce poor outcomes that may not be caught until a customer, an auditor or a regulator finds them.
The regulatory view
For UK financial services, the regulatory position reinforces the point rather than changing it. The FCA has said it will not introduce new regulations for AI, relying instead on existing frameworks including the Consumer Duty, the Senior Managers and Certification Regime, and its expectations on governance and controls[2]. Accountability for outcomes therefore stays exactly where it was, with the firm and its named senior managers, whether a decision was made by a person or an agent. For firms within the scope of the EU AI Act, AI used to assess creditworthiness or for risk assessment and pricing in life and health insurance is classified as high-risk[3], which brings explicit data governance requirements. Neither regime treats “the agent did it” as an answer.
Autonomy should be earned by the records
None of this is an argument against agentic AI. It’s an argument for matching the autonomy you give an agent to the reliability of the information it will act on.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls[1]. Some of those risk controls will concern the AI itself. I’d suggest many more should concern the records underneath it.
In practice, that means a few things. Deciding deliberately which information an agent may act on, rather than connecting it to everything it can reach. Starting with processes where the underlying records are already current, linked and well governed. Keeping a person in the loop wherever the history is known to be messy, and treating that as a signal to fix the records rather than a permanent cost. And making sure every action can be traced back to the documents that informed it.
The question to ask before giving an agent more autonomy isn’t only “is the AI good enough?” It’s “are the records good enough to be acted on without anyone reading them first?”
How Dajon approaches it
At Dajon, we’ve spent nearly thirty years managing the records that financial services and insurance firms rely on: storing them, digitising them, governing their retention and destroying them securely at the end of their life. We know how much of a typical policy, claims or client file only makes sense to someone who already knows its history.
That’s the work we’re focused on now: helping firms understand which parts of their historical information are reliable enough for AI to act on, and bringing the rest up to that standard. That means capturing documents so that what’s extracted can be trusted, linking amendments and cancellations to the records they change, marking what is current and what has been superseded, and keeping out of an agent’s reach the information it shouldn’t be using.
Agentic AI will change how financial services operate. The firms that benefit most will be the ones that recognised early that when AI starts acting, the archive starts deciding.
Can you really have an AI strategy without a historical data strategy?
References
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 BigDATAwire[↩][↩]
- AI in Financial Services: Shaping Our Approach Through Industry Engagement FCA[↩]
- Annex III: High-Risk AI Systems Referred to in Article 6(2) EU Artificial Intelligence Act[↩]