How do we keep humans in the loop for high-risk cases?

As delegated authority organisations bring automation and AI into bordereaux processing and data validation, a governance question follows close behind: how do you make sure the cases that carry real risk still land in front of an experienced person before anything is finalised?

A new coverholder submitting an unusually large claim, a bordereaux entry that doesn't match the binder's terms, or a pattern that breaks from historical norms all deserve scrutiny. Get the framework wrong and you either slow everything down with excessive manual review, or move too fast and miss the cases that genuinely need attention.

This article sets out a practical way to think about human-in-the-loop oversight: what it means, how it differs from other oversight models, and how to design it so that AI supports rather than replaces human judgement on the cases that matter most.

What "human in the loop" means in delegated authority oversight

Human-in-the-loop oversight is a governance model in which a person must review and approve a case before a decision is finalised or an action is taken. It sits alongside two other models worth distinguishing.

In a fully automated model, the system processes data and takes action without any routine human review. This might be appropriate for low-risk, high-volume, well-understood tasks, but it is rarely suitable for cases with material financial, regulatory or reputational consequences.

In a human-on-the-loop model, the system acts automatically but a person monitors outcomes and can intervene after the fact, for example through periodic sampling or dashboards. This provides some assurance but relies on the reviewer noticing a problem, often after the event.

Human-in-the-loop sits between these two. The system does the heavy lifting of processing and analysis, but for cases that meet defined risk criteria, it stops short of finalising anything. Instead, it routes the case to a person who investigates, applies judgement and signs off before the process continues.

For delegated authority, where a managing agent or insurer carries the underwriting and regulatory consequences of decisions made on their behalf by coverholders and MGAs, this distinction matters. Volume and format inconsistency make full manual review of everything impractical, but the material risk in a small proportion of cases means full automation isn't appropriate either. Human-in-the-loop offers a middle path: broad automated coverage, with mandatory human judgement reserved for the cases that need it.

How organisations traditionally identify and escalate high-risk cases

Before AI-assisted processing, most DA organisations relied on manual sampling and rules-based flags to identify cases needing senior review.

A common approach is periodic audit: reviewers select a sample of bordereaux entries, perhaps five or ten percent, and examine them in detail, looking for errors, unusual claims or inconsistencies with binder terms. Sampling is a reasonable way to manage limited reviewer capacity, but it inevitably misses cases outside the sample, including some that would have warranted a closer look.

Another traditional approach uses simple rules built into spreadsheets or policy administration systems: a claim above a certain value triggers a referral, or a coverholder flagged as "new" requires additional scrutiny for a set period. These rules are useful but often static. They are set once and rarely revisited, and they typically operate on a subset of fields rather than the full picture of a bordereaux entry.

Both approaches share the same underlying limitation: they depend on manual effort applied inconsistently across a large and growing volume of data, and are vulnerable to reviewer fatigue, resourcing pressure and simple oversight. Cases that fall just outside a threshold or outside the sampled proportion can pass through unreviewed.

Where AI helps identify and route high-risk cases

AI-assisted processing changes what's achievable here in one specific way: it allows risk criteria to be applied consistently across the entirety of a data set, rather than a sample of it.

Rather than reviewing one bordereaux entry in ten, an AI system can assess every entry against defined criteria, such as claim size, coverholder tenure, deviation from expected values, or inconsistency with binder terms, and flag the ones that meet those criteria for review. This closes the gap left by sampling and reduces reliance on static, manually maintained rules, since AI can also surface anomalies that don't fit a predefined rule but deviate meaningfully from historical patterns.

It is important to be precise about what this changes and what it doesn't. AI's role here is triage: identifying and prioritising cases that meet risk criteria so that a human reviewer's attention goes where it is most needed. It does not extend to resolving those cases. A flagged claim from a newly onboarded coverholder still requires a person to investigate context, apply judgement and decide what happens next. The AI's output is a recommendation for human attention, not a decision.

This distinction should be reflected directly in how workflows are designed: AI surfaces and prioritises, humans review and sign off.

Designing an effective human-in-the-loop workflow

Putting this into practice requires a few deliberate design choices.

First, define explicit criteria for what counts as high-risk, rather than leaving this to informal judgement. Common criteria include claims above a set monetary threshold, entries from coverholders onboarded within a defined recent period, data inconsistent with binder terms, and deviations from historical patterns. These criteria should be documented, agreed with oversight and compliance stakeholders, and reviewed periodically as the business and its coverholder relationships evolve.

Second, assign clear sign-off responsibility. Every escalation should have a named reviewer or role accountable for the final decision, with authority to investigate, request further information and resolve the case.

Third, maintain an audit trail. Regulators, auditors and boards will want to see what was flagged, why, who reviewed it, what they found and what action followed. This record should be a routine output of the workflow, not something reconstructed after the fact.

Finally, monitor and adjust the thresholds themselves. If escalation volumes are consistently overwhelming reviewers, thresholds may be too broad, risking alert fatigue and reviewers rushing through cases. If very few cases are ever flagged, thresholds may be missing genuine risk. Treat the criteria as something to be refined over time based on outcomes, not fixed at launch.

Example

A London managing agent uses an AI-assisted platform to process monthly bordereaux from a portfolio of coverholders writing marine cargo business. The platform is configured to flag any claim above a set monetary threshold, any bordereaux entry from a coverholder onboarded within the last twelve months, and any entry inconsistent with the binder's terms. These flagged cases are routed to the oversight team's queue rather than being processed straight through.

The DA oversight manager reviews a flagged claim from a newly onboarded coverholder that falls just above the monetary threshold. On investigation, the manager finds the claim is legitimate but identifies a wording ambiguity in the binder that should be clarified. The AI's flag did not resolve the case, but it ensured the case reached a human reviewer promptly, with full context, and the resulting decision and reasoning are logged for audit purposes.

FAQs

• What counts as a "high-risk" case in delegated authority oversight?

  High-risk criteria vary by organisation, but commonly include large or unusual claims, new or recently onboarded coverholders, data inconsistent with binder terms, and patterns that deviate from historical norms. These criteria should be explicitly defined and periodically reviewed rather than left to informal judgement.

• Does using AI for bordereaux processing mean less human oversight?

  Not necessarily, and in a well-designed process, often the opposite. AI can review the full data set rather than a manual sample, which increases the consistency and coverage of oversight. The cases that matter most are still routed to a human for judgement and sign-off.

• How do we avoid overwhelming reviewers with too many escalations?

  Set sensible, risk-based thresholds rather than flagging everything, and periodically review escalation volumes and outcomes. If reviewers are consistently overwhelmed, or if very few cases are ever flagged, that's a signal the criteria need refining.