Home Institute The Institute Eva M. Erpenbach Research The Research Architecture Institute Essays Research Dialogues Guest Essays External Contributions Fellowship The Fellowship Fellows Fellowship Contributions Partners Project Partnerships Project Partners Young Research Contact Legal Notice
DEEN
← Institute Essays
Institute Essays / 05

Provision 29 and AI: Why Separate Assurance Sources May Share an Analytical Path

Separate assurance functions may share an AI-mediated selection or prioritisation step. Separate judgments do not necessarily provide independent detection paths.

Eva M. Erpenbach
Founder & Scientific Director, Erpenbach Institute

Provision 29 requires boards to monitor and review material controls and to report on their effectiveness. Under paragraph 295 of the FRC guidance, the board should form its own view of effectiveness from the evidence it obtains, while the financial statement auditor’s opinion does not cover the Provision 29 statement. Where several assurance functions contribute to that evidence, separate judgments can still rest on a shared analytical path if the same AI system influences what reaches each function for review.

Companies rely on separate assurance functions because one review should not simply reproduce another. Separate functions in compliance, risk and internal audit may examine the same area from different positions while remaining organisationally distinct. AI can leave that structure intact while giving several assurance sources the same analytical dependency.

Consider a company in which compliance and internal audit both review a large population of transactions. Each function conducts its own assessment and reaches its own conclusions. To focus attention, however, both use the same AI system to identify transactions that warrant closer review.

Each function can remain independent and independently responsible for its conclusions. Both reviews nevertheless begin after the same analytical step has influenced what reaches them. If that step omits, classifies or deprioritises a matter, the effect occurs before either function begins its own assessment.

Shared underlying data presents a different case. Two assurance functions can examine the same transactions and still determine independently which cases require attention and what conclusions follow. A shared analytical step can perform part of that determination for both. Shared infrastructure or formatting tools do not by themselves create the same dependency.

A similar issue can arise in third-party risk. A risk function may monitor suppliers while internal audit periodically reviews the same exposure. If both use the same AI-generated prioritisation to determine which suppliers receive closer examination, their later assessments can remain independent while the route through which suppliers reach those assessments is shared.

A shared analytical dependency exists where the same analytical step can affect whether, or in what form, a matter reaches more than one assurance source for independent assessment. It can arise even when the system performs as intended. Several assurance sources can use the same analytical step to surface, classify or prioritise matters before reviewing them.

The governance relevance is clearest when a board considers the assurance on which it will rely. Paragraph 273 of the FRC’s Corporate Governance Code Guidance asks the board to consider the level of assurance it receives on the risk management and internal control framework and whether this is enough to help it satisfy itself that the framework is operating effectively. Paragraph 188 says the audit committee should discuss what information and assurance it requires and satisfy itself that those sources are sufficient and objective. Multiple assurance sources may contribute to the board’s assessment while still sharing an upstream analytical step.

Audit committees should identify shared analytical dependencies where they rely on multiple assurance sources for the same significant risk or control. Management or the relevant assurance functions can identify whether a common AI system performs the same selection, classification or prioritisation step across those sources. The practical question is whether each source determines what requires attention through its own route, or whether the same analytical step affects what more than one source receives for review.

The FRC guidance also expects audit committees to pay particular attention where the work of risk, compliance, finance, internal audit and external audit may be aligned or overlapping and to oversee those relationships so that they are coordinated and operate effectively to avoid duplication. Coordination and shared technology can improve coverage and make assurance work more efficient. A separate question arises when the same analytical step influences what several assurance sources receive for review.

The appropriate response depends on the significance of the risk and the role of the shared step. For some significant risks, the committee may conclude that at least one route should allow relevant matters to reach independent assessment without passing through that step.

That route can be proportionate to the risk. One assurance source might use an independently defined sample for part of its work. A function might retain direct access to the underlying population and periodically test matters outside the AI-selected set. For some risks, a separate exception or escalation route may be sufficient. These measures preserve an additional detection path where the shared analytical dependency matters most.

The dependency should also be visible when the committee and board interpret assurance reports together. Separate reports may represent separate judgments without representing separate opportunities to surface a matter that the shared analytical step could affect.

As boards develop the evidence and assurance supporting their Provision 29 reporting, formal assurance structures may remain unchanged even as AI is used within them. That alone does not establish whether a significant matter has more than one analytical route into the assurance process.

Separate assurance sources do not necessarily provide separate detection paths.