CEO Corner

The answer may need the past and the present at the same time

Return to our illustrative policyholder. The policy was issued in 2008. An endorsement was added later. The book subsequently changed ownership, and the insurer introduced a new servicing procedure this year. A financial quotation must also use the relevant current account position and valuation date.

There is no single “latest version” that answers all of this.

A current servicing instruction cannot simply be assumed to change an older contractual entitlement. Equally, an old contract does not mean employees should use a retired operating procedure. The applicable legal and contractual interpretation belongs with qualified specialists; the technology must preserve and apply the rules they approve.

For this case, the policy administration system or validated calculation service should produce the quotation. AI can assemble the relevant evidence, explain the calculation’s inputs and highlight a conflict. It should not improvise a financial value from retrieved prose.

That distinction is especially important in long-duration insurance. A portfolio can stop accepting new business while its servicing, claims and reporting responsibilities continue. Modernization must carry forward the obligations of that book, including the distinctions that make it different from the insurer’s newer products. A technically successful migration can still leave that business meaning incompletely represented.

The work experienced people do between the documents

When two documents disagree, a capable employee does more than choose the more recent one. They investigate the relationship: is one a general rule and the other an authorized exception? Does one concern a different jurisdiction? Was a change approved but never implemented? Is the inconsistency a genuine defect that needs a business decision?

A system that blends those sources into a smooth paragraph can hide the very disagreement a reviewer needs to see. Fluency can remove the visible signs of uncertainty without resolving their cause.

This is why I would examine the work of experienced staff before designing the retrieval process. Ask a specialist to explain several difficult cases, including a case where the apparently obvious source was wrong. Record the distinctions they used and the evidence that supported them. Challenge informal workarounds too: long-standing practice is not automatically an approved rule.

The resulting asset is more useful than a folder of documents. It is a maintained account of which information applies to which decisions, together with the person authorized to resolve uncertainty. That knowledge can improve staff training, system conversion and operational consistency as well as AI.

The objective is not to encode every professional judgment as a rule. Some cases remain genuinely ambiguous. A good design recognizes those cases early, explains what is unresolved and gives a qualified person the evidence and authority to decide.

A merger makes the problem visible; agentic AI makes it urgent

A bank integrating an acquired business encounters a related challenge. Two customers may appear to hold the same account type while having different agreements, pricing arrangements or servicing exceptions. A consolidated customer record does not establish that their treatment should be identical.

In a hypothetical fee inquiry, an assistant might locate the acquiring bank’s current tariff while overlooking a valid arrangement retained for an acquired customer group. The remedy is to establish the governing agreement and authorized treatment, not merely refresh the tariff more frequently. Where records conflict, the case needs resolution before a response or adjustment is approved.

This has strategic implications for acquisition plans. A business case that budgets for moving records but leaves the reconciliation of meanings to frontline staff can preserve a substantial integration burden. AI investment should expose that burden and help reduce it, rather than make it less visible behind a unified interface.

The Basel Committee’s January 2026 newsletter on BCBS 239 identifies mergers, new products, fragmented data estates and data lineage as continuing implementation challenges. It is an informational discussion of risk-data practices, not a new AI requirement. Its relevance here is the connection between business change, data meaning and sustained management ownership.

Basel Committee, January 2026

Agentic AI raises the stakes because a mistaken interpretation can travel beyond an answer. A retrieval agent supplies evidence, a second agent prepares a recommendation, and a third initiates a permitted workflow. Agreement between those agents is weak reassurance if they all depend on the same inapplicable source. The system needs a common, controlled basis for deciding what evidence is eligible and which actions are authorized.

The architecture must preserve the basis of the decision

Retrieval-augmented generation, or RAG, brings retrieved information into a model’s response. It can make an answer easier to trace, but relevance alone does not establish applicability. A highly relevant paragraph can still concern the wrong contract or customer population.

Our proposed design separates four responsibilities: establishing the case, selecting applicable evidence, producing the explanation, and authorizing any action. The model can contribute within that process; it should not be responsible for granting itself access or deciding the limits of its own authority.

Original Genellipse diagram. Labels are in English; the accompanying explanation is available in your selected language.

Read the explanation and assumptions

Trusted case context, approved sources and authorized financial systems provide distinct inputs.

An independent eligibility check establishes applicable sources, access permissions and effective dates before bounded AI prepares an explanation.

A qualified reviewer handles exceptions. Before execution, the workflow rechecks permissions, source validity and current state. Evidence is retained across every stage.

At the start, trusted systems supply the policy or account identity, product, market and relevant dates. Approved rules determine which sources the user and workflow may access and which versions apply. Only eligible material enters the answer-generation stage. Metadata filters are one available technical mechanism; Microsoft’s search documentation describes filtering content by defined criteria. The difficult business work remains agreeing and maintaining those criteria.

Microsoft: search filters

Current balances and financial calculations follow their own authorized system connections. The evidence record should connect the explanation to the contract version, retrieved material, calculation result, unresolved exceptions and approval. This gives a reviewer something more useful than a transcript: a basis for understanding how the decision was reached.

Before a consequential action, the system checks again. A recommendation may wait in a queue while a source is withdrawn, an account changes or an approval expires. Permission and state must still be valid when the action executes. A conventional API and a Model Context Protocol (MCP) tool both need this discipline; neither connection type creates business authority by itself.

Changing the source means managing the consequences

Knowledge maintenance becomes an operating responsibility once answers influence work. Suppose a procedure is withdrawn at noon. The affected material may already exist in an index, a translated copy, a cached answer and several unfinished cases. Replacing the master file addresses only one part of the problem.

The organization needs to identify those dependencies, prevent inappropriate further use, verify that the change propagated and determine which pending cases need reassessment. If a decision has already been executed, changing the knowledge base does not repair it. An authorized team must assess the consequences and any corrective action.

Historical evidence should remain available under the appropriate retention and access arrangements. There are two different questions to preserve: what should govern a decision now, and what information was available when an earlier decision was made. Recording both effective dates and the dates changes entered the system helps keep those questions separate.

Someone must own the complete process. The business owner approves applicability; engineering verifies propagation; operations handles affected cases; risk, compliance and legal specialists assess consequences within their remit. A service owner must resolve gaps between these responsibilities. NIST’s Manage guidance includes post-deployment monitoring and change management; the operating arrangement described here is our proposed application.

NIST Manage

Oversight should reflect the consequence of error. EIOPA’s August 2025 opinion takes a risk-based and proportionate approach to AI governance in insurance. My practical interpretation is that a general information response and an action affecting a policyholder’s money need different levels of evidence, approval and control. The opinion is addressed to national supervisors; it is not a certification of any platform.

EIOPA opinion

Measure the cost of the whole decision

A productivity claim is incomplete if it counts faster preparation while ignoring the people who reconcile, correct and explain the result. There is also a risk of moving work: an assistant handles more requests, but specialists inherit a growing queue of difficult exceptions.

The business case should therefore measure the complete journey, including the cost of keeping knowledge usable. The following monthly example is hypothetical; its inputs are discussion assumptions, not Genellipse results or forecasts.

The difference is 324 hours of gross capacity. If maintaining sources and performing additional review require 60 incremental hours, 264 hours remain. Those hours become financial savings only if an actual expense changes. They may instead support higher volumes, shorter waiting times or better service. Implementation, platform and other operating costs still belong in the investment calculation.

The target also needs evidence that the intervention caused the improvement. Compare equivalent product groups and case types; account for changes in volumes, staffing and complexity. Check that repeat contact has not simply moved to another channel. An improving average can conceal deterioration in a small inherited portfolio.

Quality cannot be reduced to one accuracy percentage. Errors differ in consequence: a misplaced explanation and an incorrect payment instruction should not carry the same weight. Test difficult cases by contract generation, jurisdiction, exception type and action severity. Track unresolved exceptions and their age, because an assistant that defers everything may be cautious while delivering little operational value. A useful evaluation shows both where the system succeeds and where people still need capacity to intervene.