What is the escalation path when AI and humans disagree?

AI tools are increasingly used to validate, classify and extract data from bordereaux and other delegated authority documentation. As their role grows, so does the likelihood that an AI output will conflict with the judgement of the person reviewing it.

This is not a sign that the AI or the reviewer has failed. It is an expected part of introducing AI into interpretive work, and it needs a defined process rather than case-by-case handling.

Without that process, some staff will defer to the AI output by default, others will override it without recording why, and neither approach stands up well under audit or regulatory scrutiny.

This article sets out how delegated authority organisations should design and operate an escalation path for exactly this situation.

Why AI and human disagreement needs a defined process

When an AI system flags a classification, extraction or validation result that a human reviewer disagrees with, the organisation faces a choice point that recurs constantly across a bordereaux processing cycle.

If that choice point is handled inconsistently, three problems follow.

First, outcomes vary depending on who happens to be reviewing the exception that day. One analyst might accept the AI's flag without question, while another overrides it based on personal experience, with no record of either decision.

Second, accountability becomes unclear. If a decision later proves wrong, it may be impossible to establish who made the call, on what basis, and whether it was reviewed by anyone with the authority to make it.

Third, the organisation loses evidence. Lloyd's oversight expectations and regulatory review both depend on being able to show that exceptions were identified, considered and resolved by an accountable person. Ad hoc handling leaves nothing to show.

A defined escalation path solves all three problems by making disagreement handling predictable, accountable and evidenced.

How DA organisations traditionally handle exceptions and disagreements

Escalation is not a new concept in delegated authority oversight. Long before AI entered bordereaux processing, DA organisations already had mechanisms for handling situations where something did not look right.

Underwriters have long operated referral processes, where a case outside agreed authority or outside expected parameters is referred upward for a decision. Query logs have long served as a record of issues raised against a bordereaux submission, tracking what was queried, who answered, and what was agreed. Manual override processes have existed wherever a system-generated figure or flag needed a human correction, typically requiring a reason code and a named approver.

These mechanisms share a common structure: something is flagged, it is assessed against a threshold of significance, it is referred to someone with the authority to decide, and the decision is recorded.

Extending escalation to cover AI-generated outputs is therefore not a new discipline. It is the application of an existing discipline to a new source of exceptions.

Where AI changes the escalation picture

AI introduces a few genuine differences that traditional exception handling did not need to account for.

AI outputs often carry a confidence score or probability rather than a simple pass or fail flag. This means triage thresholds need to reflect not just the type of discrepancy but how confident the AI was in its own output, and how material the discrepancy is to the underlying business decision.

AI can also generate a much higher volume of flags than manual review previously produced, since it is reviewing every record rather than a sample. Without sensible thresholds, this volume can overwhelm reviewers and push all disagreements, however trivial, into the same escalation queue as materially significant ones.

Finally, because AI outputs can appear authoritative, there is a risk that reviewers defer to them without genuine scrutiny, or conversely dismiss them without genuine consideration. Both outcomes undermine the purpose of having a human in the loop at all.

None of this changes who should hold decision authority. It changes how disagreements are surfaced and how they should be prioritised for human attention.

Designing and documenting the escalation path

A workable escalation path has four stages.

Detection is the point at which a disagreement is identified, either because the AI flags something the human reviewer disputes, or because the human reviewer identifies something the AI has not flagged that they believe it should have.

Triage assesses the disagreement against predefined risk or materiality thresholds. Low-value, low-risk discrepancies may be resolved by the reviewer directly, following documented guidance, while higher-risk or higher-value discrepancies are referred onward.

Referral sends the disagreement to a named resolution authority, such as a DA oversight manager, rather than a committee or shared inbox. A specific named role avoids diffusion of accountability and ensures someone can always be identified as having made the final call.

Logging records the outcome: what was flagged, who reviewed it, what decision was made, and the rationale behind it. This log is what makes the escalation defensible under audit or regulatory review.

Throughout all four stages, the AI system's role is to surface the disagreement and provide supporting information. It does not resolve the disagreement itself. That authority always sits with the named human role.

Example

A Lloyd's managing agent uses an AI tool to validate monthly bordereaux submitted by an overseas MGA writing agricultural risk. The AI flags a batch of premium figures as inconsistent with the binder's rating basis, but the DA oversight analyst reviewing the flag believes the figures are correct given a recently agreed endorsement.

Under the defined escalation path, the analyst logs the disagreement, which is triaged as low materiality and referred to the DA oversight manager for a same-day decision. The manager reviews the endorsement documentation, confirms the analyst's reading is correct, and records the rationale in the escalation log.

The disagreement is resolved within the same reporting cycle, the bordereaux is accepted without delay, and the escalation record provides clear evidence for internal audit that the exception was reviewed and resolved by an accountable human decision-maker.

FAQs

• Who should have final say when AI and a human reviewer disagree?

  Final decision authority should always rest with a named, accountable human role, never the AI system, regardless of how confident the AI output appears. The AI's role is to surface information and flag discrepancies, not to resolve them.

• Should every AI and human disagreement be escalated?

  No. Escalation paths should apply risk-based thresholds, so only disagreements above a defined materiality level require formal referral. Applying the same process to every minor discrepancy would overwhelm reviewers and undermine the purpose of triage.

• What needs to be recorded when an escalation occurs?

  At minimum, the record should show what was flagged, who reviewed it, what decision was made, and the rationale behind it. This is what makes the resolution defensible under audit or regulatory review.