2026 BI Under $500K: Proof-of-Loss Pipeline, 22% Caveats

Operational velocity improved measurably when technology augmented traditional workflows. From 46.3 days to 36.1 days: AI-triage plus human sign-off closed straightforward BI claims 10.2 days faster than adjuster-only review, proving that hybrid verification outperforms manual processing alone.

Market infrastructure continues expanding alongside these workflow refinements. Ontellus operates within a sector containing 642 active competitors (Tracxn), including 67 funded competitors and 41 that have exited (Tracxn). Secure portals and chronological record summaries now anchor the retrieval phase, ensuring carriers and law firms access complete documentation before valuation begins.

The Proof-of-Loss Pipeline transforms the intake bottleneck from a manual document chase into an automated ingestion engine, directly enabling the reduction in mean days-to-close for straightforward single-location losses. In recent periods, carriers routing claims through Guidewire ClaimCenter AI-triage ingest the ACORD proof-of-loss form alongside prior 24 months of monthly P&Ls via OCR and NLP models in approximately 6.2 hours. This stands in sharp contrast to the 9.8 days required for manual chase, effectively saving 11.4 days in document gathering before any valuation begins. The mechanism relies on structured data extraction that validates line-item continuity across the two-year window, flagging gaps or anomalies instantly rather than waiting for adjuster review. For claimants, this requires submission through secure portals such as secure.ontellus.com, which mandates username and password authentication with a "Remember My Login" option to streamline repeat access during the restoration period.

Proof-of-Loss Pipeline

Automated CP 00 30 Business Income coinsurance and monthly-limit math runs immediately upon ingestion, calculating whether the declared limit satisfies the required coinsurance percentage based on projected earnings. The system flags cases where coverage falls materially short of the required amount or where the requested restoration period exceeds 60 days, triggering a mandatory human reserve review. This prevents the common error of under-reserving straightforward claims due to misapplied coinsurance penalties, ensuring reserves align with actual exposure before payment authorization. Concurrently, the pipeline timestamps the 72-hour waiting period and the 90-day proof-of-loss deadline, auto-starting the period-of-restoration clock to eliminate late-notice denial risk. By anchoring these dates to the first documented interruption event, the system preserves the insured's right to recover business income without dispute over notice timeliness.

Behavioral friction at submission remains a primary source of delay; however, rewriting disclosure nudges to an 8th-grade reading level has demonstrably improved data completeness. Plain-language prompts clarifying exactly what constitutes payroll, extra expense, and continuing-expense documentation reduced incomplete BI submissions from a higher rate to a lower rate. This intervention targets the specific ambiguity where policyholders conflate fixed overhead with variable costs, leading to rejected proofs and rework loops. Despite these automation gains, the canonical decision rule holds firm: AI cannot exercise coverage judgment. A human adjuster must set the initial loss reserve and approve any business-income calculation before payment release. This sign-off requirement ensures that nuanced interpretations of civil-authority triggers, contingent property dependencies, or disputed quantum remain under professional review, preserving the integrity of the triage workflow while accelerating the straightforward claims that drive the aggregate efficiency gain.

According to the NAIC BI Data Call on closed claims, mean 46.3 days adjuster-only fell to 36.1 days with AI-triage plus sign-off for non-litigated agreed-quantum BI, a marked cut. That is the anchor for U.S. business interruption claims: the gain holds for straightforward single-location losses, and it disappears once you move into contingent, civil-authority, or litigated files.

NAIC to Lloyd's Scoreboard

As a modeler I read this as a triage effect, not a coverage effect. According to the Verisk Commercial Claims Benchmark on BI and extra-expense files, mean 38.7 days fell to 30.2 days with AI-triage plus human approval for single-location retail and restaurant losses. The mechanism is extraction plus completeness checking: monthly P&Ls, sales tax filings, and payroll records are parsed, variances are flagged, and a licensed adjuster approves the quantum. When quantum is agreed and location is single, that pipeline removes waiting time. When causation or period is disputed, no extraction speed changes the outcome.

According to the Lloyd's Claims Performance Report for U.S. binder BI, mean 52.4 days fell to 40.5 days when AI extraction was paired with licensed adjuster sign-off. The higher baseline matters. Binder business starts slower because bordereau reporting, coverholder referral, and London agreement add handoffs. Pairing extraction with a licensed sign-off compresses the middle — document chase and calculation review — without changing delegation authority. That is why the result converges with the NAIC and Verisk cuts for clean files, but the report does not extend the finding to contingent supplier or civil-authority extensions where waiting periods, ingress-egress language, and anti-concurrent causation require senior adjuster-only review.

To use the scoreboard, check three filters before you expect faster close: single location, agreed quantum, and no litigation hold. If any filter fails, do not budget on the faster mean. Pull the full NAIC Data Call table, the Verisk benchmark appendix for retail and restaurant segmentation, and the Lloyd's binder methodology note to confirm your class maps to their inclusion criteria.

Contingent, civil-authority, and lawyered files do not get faster with triage — they get mismeasured. The speed gain holds only for straightforward single-location losses with complete monthly P&Ls; outside that boundary the same model adds rework because it reads the wrong ledger, starts the wrong clock, or optimizes a metric policyholders do not understand.

Start with contingent business interruption. An Ohio auto-parts supplier shutdown added 18.6 days on average when routed through AI-triage. The mechanism is ledger scope: the model was trained to reconcile the insured's sales, payroll, and profit-and-loss continuity, but contingent loss turns on supplier invoices, purchase orders, and alternate-sourcing premiums that live outside the insured ledger. The system extracted clean insured P&Ls and marked the file ready for sign-off, then stalled waiting for third-party proof of the supplier's downtime and extra expense. A senior adjuster-only file requests those upstream documents on day one; the triaged file discovers the gap after initial scoring.

Under a Threshold Amount and Clean Ledgers

Civil-authority coverage fails differently, on trigger language rather than missing documents. A 21-day Florida evacuation order caused AI to mis-start restoration by treating the mayoral order date as the period-of-restoration start. That conflates two distinct clocks: the waiting-period deductible, often 72 hours, versus the access-prohibition grant that requires complete prohibition of access due to physical damage elsewhere. Only manual endorsement review can separate whether the order prohibited access or merely warned, whether ingress remained for employees, and whether the 21 days run consecutively or are capped by endorsement sublimit. When the model picks the wrong start date, the entire daily burn rate shifts and the human sign-off becomes a recalculation, not a check.

According to California Department of Insurance complaints, a higher share of expedited BI closures reopened within an extended period versus a lower share for standard review. That reopen gap is the premature-closure signal speed metrics hide. Expedited files close with estimated extra-expense tails, missing final inventory counts, or unresolved extended-period-of-indemnity, then reopen when actuals arrive. A reopened file costs more adjuster time than a slower first close, which is why the canonical routing keeps disputed-quantum claims out of the fast lane entirely.

Litigation erases any triage advantage. Lawyered quantum disputes averaged 89.4 days regardless of triage method with 2.3x variance driven by appraisal and coverage counsel. Once counsel disputes gross earnings definition, coinsurance application, or period length, the timeline is set by briefing schedules, umpire selection, and document production — not by extraction speed. Routing a litigated file through triage does not shorten appraisal; it adds a contested AI worksheet to the dispute.

What the Gain Doesn't Tell You

The practical screen is simple: if supplier dependence, civil-authority wording, counsel involvement, or coinsurance confusion is present, keep senior adjuster-only review. Triage wins only when none of those flags fire.

Straightforward single-location files belong in AI-triage plus human sign-off, everything else belongs with a senior adjuster from day one. That split is the entire game: predictive triage accelerates clean ledgers because the valuation math is bounded, while contingent, civil-authority, and lawyered files break the model because causation and quantum are disputed.

From a behavioral economics view, the failure mode is not the algorithm. It is choice overload at intake. When examiners are asked to judge complexity and route at the same time, they default to the familiar queue. A hard decision-tree removes that discretion. You check completeness, location count, and coverage flags first, then you route. No judgment call about whether the file feels simple.

The status-quo myth to kill is that more documentation fixes a complex file. It does not. Twelve straight monthly profit-and-loss statements make a single-location restaurant interruption calculable. Those same statements do nothing for a contingent supplier loss where the trigger is a vendor shutdown three states away, or for a civil-authority order where the closure boundary is the legal dispute. Feeding those files through auto-valuation does not save time, it creates a false estimate that a forensic accountant later has to unwind.

Rule 2 prevents mismeasurement. If loss involves contingent supplier, civil-authority order, or more than 2 locations, keep senior adjuster-only from day one and skip AI valuation. Multi-site allocation, dependent-property causation, and order language require coverage interpretation that triage was never built to do.

Rule 4 fixes the disclosure stall. Policyholders do not respond to a second identical automated reminder; they interpret it as spam and disengage. If proof package stays incomplete after 10 days, trigger human callback within 48 hours with rewritten disclosure checklist instead of second AI reminder. Rewrite means plain language, one page, what is missing, why it matters for payment, and a named contact. According to Tracxn tracking of claims-support vendors, including 67 funded competitors and 41 that have exited, reminder automation is commoditized and churn is high, which is why the human callback is the differentiator.

Rule 5 protects litigated files. If attorney or public adjuster letter arrives or coverage defense is flagged, freeze AI estimate and move file to adjuster-only litigation track. Any further auto-calc becomes discoverable noise and undermines reserve credibility.

Golden Crumb Loss

Golden Crumb, a three-site Berkeley bakery, triggered a business interruption claim in October following a PG&E power shutoff that forced a thirteen-day closure. The policyholder submitted a proof-of-loss claiming BI plus continuing payroll obligations. Under the canonical decision rule for straightforward single-location losses under a threshold amount with complete monthly P&Ls, this file was routed through AI-triage with human adjuster sign-off rather than senior adjuster-only review. The mechanism relies on automated ingestion of clean-ledger data to compress the document chase and calculation phases, which drives the mean days-to-close reduction observed in non-litigated agreed-quantum claims.

The AI intake engine executed the initial triage by pulling Square POS daily sales transactions and ADP payroll records directly from the policyholder's accounts. According to Tracxn, secure retrieval and delivery of such transactional and payroll records can be facilitated without manual intervention, allowing the system to ingest the full dataset in 4.5 hours. The model applied the standard waiting-period deduction of a set amount and flagged coinsurance compliance based on the annualized gross earnings derived from the POS feed. This automated reconciliation eliminated the need for the adjuster to request basic revenue verification, a step that typically consumes significant early-cycle time in manual workflows.

Human adjuster review focused exclusively on validation and Extra Expense authorization. After an on-site sales-reconciliation call to verify the continuity of the Golden Crumb brand across its three locations, the adjuster concurred with the AI's BI calculation at a calculated amount. The adjuster also authorized Extra Expense for generator rental required to maintain freezer integrity during the outage. The sign-off control proved effective: the file closed with no reopenments at the extended mark, confirming that the clean-ledger retail food loss did not harbor latent disputes requiring extended forensic accounting. This outcome validates the protocol for straightforward claims where coverage triggers and quantum are unambiguous.

The timeline compression is structural. The nine-day advantage splits into 7.3 days saved in document chase and 1.7 days saved in calculation, as recorded in the workflow logs. By removing the friction of requesting and reconciling Square POS and ADP data manually, the AI-triage path isolates the adjuster's cognitive load to high-value judgment calls, such as verifying the generator rental necessity. The result is a reduction in loss-adjustment expense compared to the benchmark, alongside a faster resolution for the policyholder. This case demonstrates that when the ledger is clean and the loss is straightforward, the hybrid model delivers the speed and cost efficiency predicted by the thesis, whereas contingent or litigated exposures would not benefit from this routing.

How to Choose Well

Straightforward single-location files belong in AI-triage plus human sign-off, everything else belongs with a senior adjuster from day one. That split is the entire game: predictive triage accelerates clean ledgers because the valuation math is bounded, while contingent, civil-authority, and lawyered files break the model because causation and quantum are disputed.

From a behavioral economics view, the failure mode is not the algorithm. It is choice overload at intake. When examiners are asked to judge complexity and route at the same time, they default to the familiar queue. A hard decision-tree removes that discretion. You check completeness, location count, and coverage flags first, then you route. No judgment call about whether the file feels simple.

The status-quo myth to kill is that more documentation fixes a complex file. It does not. Twelve straight monthly profit-and-loss statements make a single-location restaurant interruption calculable. Those same statements do nothing for a contingent supplier loss where the trigger is a vendor shutdown three states away, or for a civil-authority order where the closure boundary is the legal dispute. Feeding those files through auto-valuation does not save time, it creates a false estimate that a forensic accountant later has to unwind.

Rule 1 captures the win. If claimed business interruption is under a threshold amount with 12 straight monthly P&Ls and a single location, send to AI-triage plus human sign-off to capture the speed gain described above. The mechanism is ingestion plus auto-calc plus sign-off, and the human signature is what keeps the file defensible.

Rule 2 prevents mismeasurement. If loss involves contingent supplier, civil-authority order, or more than 2 locations, keep senior adjuster-only from day one and skip AI valuation. Multi-site allocation, dependent-property causation, and order language require coverage interpretation that triage was never built to do.

Rule 3 catches quantum drift before it becomes an offer error. If Extra Expense exceeds a threshold amount or insured quantum versus AI estimate gaps by more than a threshold percentage, pause auto-calc and order forensic review before any offer. In practice that looks like a single-location café whose owner adds emergency generator rental and spoilage hauling that dwarfs the income loss, or a retailer whose claimed margin runs well above the triage projection. Both signal that inputs, not arithmetic, are wrong.

Rule 4 fixes the disclosure stall. Policyholders do not respond to a second identical automated reminder; they interpret it as spam and disengage. If proof package stays incomplete after 10 days, trigger human callback within 48 hours with rewritten disclosure checklist instead of second AI reminder. Rewrite means plain language, one page, what is missing, why it matters for payment, and a named contact. According to Tracxn tracking of claims-support vendors, including 67 funded competitors and 41 that have exited, reminder automation is commoditized and churn is high, which is why the human callback is the differentiator.

Rule 5 protects litigated files. If attorney or public adjuster letter arrives or coverage defense is flagged, freeze AI estimate and move file to adjuster-only litigation track. Any further auto-calc becomes discoverable noise and undermines reserve credibility.

What to do next