Data debt is not abstract. It is already priced into loss ratios, expense ratios, and the hours of the people who hold the book together. The three costs in the draft — single point of failure, storage tax, and swivel tax — show up with different faces in every line of business in the decision model. AI does not clear the debt; it reprices it. Messy inputs produce confident, polished, and faster wrong answers. That is why evidence readiness, named ownership, and lineage are not optional controls once agents enter the workflow.
Industry figures put the cost of poor data quality in the range of 15–25% of revenue for many organizations, with insurance-specific estimates historically in a similar band relative to operating profits. More recent carrier surveys show 92% of insurers naming data as a barrier to AI success, with silos worse in insurance than in other industries surveyed. Underwriters still spend large fractions of their week on rekeying, reconciliation, and spreadsheets (one recent survey put admin/rekeying near 17% of the week). Completeness is not accuracy: a submission can have every field filled and still be wrong in ways that misprice risk or support a later coverage dispute. NAIC examination standards look for both properly completed applications and file documentation that adequately supports the decision.
The lines below follow the families in the LOB context sheet. Emphasis is on the concrete data objects that create the three taxes and on how the candidate wedges (submission prep, SOV readiness, class/mod reconciliation, cyber evidence, fleet schedule validation, referral packs, bordereau integrity, etc.) interact with them.
Personal lines
Self-reported exposure (mileage, household drivers, roof age/condition, elevation) is the dominant source of debt. Household driver reports and address-linked data frequently surface people who do not actually drive the vehicle or live at the risk; treating the match as proof creates premium leakage or false additions. Credit-based factors have shown post-disruption distortion because missing negative data inflates scores while underlying risk is unchanged. Homeowners and flood carry the classic storage/valuation problem: years of building attributes that are stale relative to actual condition, reconstruction cost, or flood zone.
Swivel tax appears as matching across policy, claims, telematics, and external household files. Single points of failure are often the underwriters or analysts who “know which report to trust” for a given territory or prior loss pattern. AI that auto-adjudicates or prices on incomplete household or renewal status can deny or misprice faster than a human would have noticed the gap.
Small commercial and mid-market standard
Commercial property / SOV. Incomplete or inconsistent COPE data, blank roof/construction fields, missing geocodes, and stale replacement-cost values are routine. Underwriters triage aggressively; low-completeness schedules go to the bottom of the stack or receive punitive defaults (unknown roof = old roof; unknown sprinkler = unsprinklered). Valuation errors (market value or book value instead of rebuild cost) create coinsurance exposure that only appears at claim. CAT models amplify the problem: poor secondary modifiers or missing attributes shift AAL estimates materially. This is why SOV exception rate and time-to-ready are natural first KPIs. The debt is both storage (years of schedules kept “just in case”) and swivel (manual enrichment against geocoders, CAT models, and accumulation systems). A named owner for the SOV version and the geocoder/CAT snapshot is the cheapest control.
Workers’ compensation. Class codes and experience mods are the purest example of lineage debt. NCCI has a few hundred class codes; raw loss-run data produces tens of thousands of variations. Unit statistical reports feed the mod and can carry errors (miscoded lost-time vs medical-only, reserves not updated, payroll left in the wrong class after audit) for multiple rating periods. Misclassification and underreported payroll are major leakage vectors; one industry estimate puts WC premium fraud (with misclassification a large component) in the tens of billions annually. Multi-state and bureau reconciliation create classic swivel. The single point of failure is often the person who knows which bureau report, which audit, and which exception to trust. An agent that reconciles class and mod faster without citing the unit-stat source, the audit version, and the governing-class rationale simply accelerates the wrong mod.
Commercial auto / fleet. VIN decode failures, stale driver schedules, missing or mismatched MVRs, and radius discrepancies (stated vs actual from telematics or IFTA) are measurable and frequent. A single unlisted or poorly qualified driver can generate losses larger than the account premium. MVR cost creates pressure to sample rather than read; sampling is a form of accepted debt. Swivel occurs across the schedule, state MVRs, CSA/SMS data, and any telematics export. Named ownership of the as-of date of the schedule and the MVR pull is required before any bounded agent should clear exceptions.
CGL and umbrella/excess. Classification judgment (ISO codes that “best describe” operations), exposure base accuracy (payroll, sales, or other), venue, and manuscript vs ISO form differences drive both pricing and later coverage disputes. Classifications roll forward year after year even as operations change. Underlying schedule concurrency, clash, and attachment for excess create additional reconciliation debt. Referral quality and quote cycle are the natural KPIs because the preparation work is where most of the swivel occurs. The underwriter who owns the manuscript, venue, and attachment decision remains the named human gate.
Inland marine. Equipment schedules arrive as unstructured PDFs or lists; values, descriptions, and locations are frequently incomplete or inconsistent. Rekeying errors on VINs and values are common. Because many forms are non-filed or judgment-rated, the debt is less standardized and more dependent on specialist knowledge — a classic single point of failure. AI that converts schedules into rating-ready data removes rekeying but still requires a human to own the valuation basis and any condition.
Specialty and long-tail
Cyber. The dominant debt is the gap between self-attested questionnaire answers and actual control evidence. Point-in-time questionnaires miss drift; “yes” to MFA frequently means MFA on some systems but not remote access paths, service accounts, or the specific asset that was later compromised. External scans, evidence hashes, and named attestors are the objects that must be reconcilable. Claims friction and potential rescission risk rise when the attested control is not in place. Evidence completeness and contradiction rate are therefore the correct first KPIs; an agent that polishes a questionnaire without citing the scan timestamp, evidence hash, and attestor simply produces a more confident wrong picture.
Professional liability / E&O, D&O, and related claims-made lines. Temporal continuity is the data object. Retroactive dates, prior-and-pending dates, prior policies, entity/subsidiary charts, and continuous coverage history are easy to lose across carrier changes or entity restructurings. A missing declarations page or an unrecorded gap can erase years of prior-acts coverage. The single point of failure is frequently the person who remembers the continuity chain or holds the prior policy file. Storage of old declarations and entity charts is cheap; the liability is the absence of a named owner and a versioned chain that an agent (or a later examiner) can cite.
Surety. The WIP schedule is the diagnostic document. Cost-to-complete estimates, profit fade versus original estimates, over/underbillings, and reconciliation to the financial statements determine capacity and judgment. Project managers are often the single point of failure for the quality of those estimates. Financial statements and WIP reports are still frequently rekeyed; the swivel between the contractor’s package, the analysis system, and the underwriter’s judgment is expensive. Historical WIP storage is high; the useful control is a named owner for the statement period, the spread version, and the indemnity agreement.
Construction, energy, marine/aviation, and environmental. These lines share high document density (contracts, surveys, engineering reports, known conditions) and specialist judgment. The debt is tribal knowledge plus incomplete or non-standard documents. HITL remains high for a reason; bounded document agents can prepare deficiency lists and completeness packs, but the specialist underwriter must still own terms, capacity, and bind.
Delegated authority, programs, and bordereaux
Data dilutes as it passes through multiple systems and formats. “Clean enough” technical validation (no blanks, arithmetic balances) frequently fails business rules (class not written under the binder, territory, per-location limit, inception date outside the treaty period, commission rate). Bordereaux arrive late or inconsistent; a one-million-line file cannot be manually reviewed. Regulatory accountability remains with the carrier even when authority is delegated. The swivel tax is the reconciliation and audit sampling work; the storage tax is the accumulation of variant bordereaux layouts. An agent that flags exceptions faster is useful only if it cites the binder version, the bordereau period, and the audit sample outcome, and if a named delegated-authority owner approves the exceptions.
What this means for AI
The reusable control pattern in the workbook (intake with source hash and parser version → context with approved sources and version → appetite → assessment with citations and model version → pricing/terms → bind with decision record → operate with drift and incident logs) is the operational answer to data debt once agents are present. Without those evidence objects, the agent simply retrieves and polishes the mess faster.
The four moves map directly:
- Name the single point of failure. For every critical process write the one person whose departure breaks it (the underwriter who knows the true rating logic, the claims handler who knows which report to trust, the specialist who holds the retro-date chain, the project manager whose cost-to-complete estimates feed the WIP).
- Price the swivel. Count the hours spent stitching SOV + geocoder + CAT, class + unit-stat + audit, questionnaire + scan + attestor, schedule + MVR + telematics, or bordereau + binder rules. That is the interest payment already being made.
- Stop the storage tax. If no decision has used a data set in two years, archive it and stop treating it as live. Most carrier storage has never informed a decision.
- Give every critical data set one named owner. Data with no owner has no one to ask, no one to blame, and no one to fix it. The cheapest control in the book is the named owner of the SOV version, the class rationale, the evidence hash, the retro source, or the bordereau period.
Data remains an asset. An asset with no owner, no lineage, and a single point of failure is a liability wearing an asset’s name. The ledger already knows the difference; the agent makes the difference material on a much shorter cycle. The decision model’s insistence on evidence readiness, named human gates, and carrier-specific inputs before any ranking is the practical recognition of that fact.
This is the companion to The Liability Side of the Ledger (Nº 019).
— Krishna