The five-year commercial loss run is the single most-touched artifact in a renewal cycle and the worst-formatted document in the entire P&C ecosystem - typically a 200dpi PDF scan of a Guidewire ClaimCenter export, or a Duck Creek bordereau dumped to Excel with merged headers, or a Sapiens output with three different date formats inside one column. The underwriter at Travelers, Chubb, Hartford, CNA, or a Lloyd's syndicate at One Lime Street will not quote without a clean version, and the treaty broker at Guy Carpenter or Aon Reinsurance Solutions will not bring the account into a $25M single-risk treaty submission without a frequency/severity narrative that ties to Schedule F (assumed and ceded) and Schedule P (loss development triangles by accident year, line, and net/gross). AI does the structural lift - OCR, normalization, outlier identification, IBNR signaling - in 35-45 minutes instead of the 6-8 hours a junior analyst used to spend on the same file. The producer, the underwriter, and the appointed actuary still own the judgment calls: which losses are reserve-redundant, which open files are likely to develop, which subrogation flags are worth chasing, and which trends the carrier needs to see before the 22% rate ask becomes a 12% rate ask. This lesson is the structured workflow for taking a messy GL/WC commercial loss run through Convr or Indico Data OCR, into a normalized table, into a frequency/severity narrative, into a treaty broker note that reconciles to the carrier's Schedule F and Schedule P - and what every adjuster, producer, and underwriter on the file owns when AI is the first pass.

The Document You Actually Receive

The commercial loss run that lands in the producer's inbox at 4:47 p.m. on a Thursday with a Friday-end-of-day market-pack deadline is almost never the clean export the carrier's policy admin system can produce. It is the version somebody printed, rescanned at 200dpi for "compliance," and emailed as a 14-page PDF where claim numbers wrapped across lines and the "Cause of Loss" column truncated to 12 characters. A typical five-year GL/WC loss run for a mid-market manufacturer has 60-180 line items across CGL, WC, and inland marine, with 8-15% of those carrying recovery type or subrogation status in a free-text comment field, and roughly 6-10% of the rows showing data quality flags - missing date of loss, missing claimant, paid-versus-incurred mismatch, or a status of "Closed-Reopened" that the carrier's PAS has not reconciled.

The structured extract the AI produces in the first pass has nine canonical columns: date of loss, claim number, policy term, cause of loss (ISO 17 cause-code mapped), status (Open / Closed / Closed-Reopened / Litigation), paid indemnity, paid ALAE, outstanding (case reserve), total incurred, recovery type (subro / salvage / SIR-eroded / deductible-applied), and subrogation status (Open / Pursued-Recovered / Pursued-No-Recovery / Closed-Waived / Not Pursued). The Convr appetite-and-data platform, Indico Data's intelligent intake, Hyperscience's document automation, and Send's submission intake module all produce a structured JSON or CSV in roughly the same shape - the differences are at the system-prompt layer (how aggressively the model attempts cause-code mapping versus surfacing as unknown) and the integration layer (which of those bookings drop directly into Applied Epic, AMS360, Vertafore Q360, or the carrier's PolicyCenter submission record). The output the underwriter actually wants is a single canonical table the carrier can load into Guidewire DataHub or push into a Schedule P-style triangle without a second cleanup pass.

The OCR and Normalization Pipeline

OCR on a 200dpi scanned loss run in 2026 is no longer the bottleneck - Indico, Convr, Hyperscience, and the LLM-native intake modules in Send and Cytora all read scanned PDFs at 96-99% character-level accuracy on standard typewritten outputs and 88-94% on noisy fax-of-a-fax inputs. The real lift is in normalization: standardizing date formats (12/4/2023, 04-DEC-23, and 2023-12-04 all need to collapse to ISO-8601), reconciling claim-number formats across multiple carriers when the producer pulls runs from prior placements, and applying the ISO 17 cause-code mapping that lets the underwriter compare frequency by cause across accounts. A messy "slip and fall - parking lot" comment maps to ISO cause-code 02 (Slip/Trip/Fall) with sub-code 02-A (Premises). A "rear-ended at red light - LX" maps to commercial auto code 11 (Vehicle Collision) with sub-code 11-C (Struck by Other Vehicle). The model handles 75-85% of cause-code mapping cleanly; the remaining 15-25% goes to a human-in-the-loop queue the underwriter or producer's analyst clears in 10-15 minutes.

The reconciliation pass is where the workflow earns its keep. The loss run from the prior carrier almost never reconciles to the agency management system on the first pass. Total paid on the loss run will disagree with total paid in Applied Epic by 3-8% on a typical mid-market account - sometimes because the loss run shows gross paid before the deductible reimbursement, sometimes because the carrier issued a payment after the run was generated, sometimes because a recovery was credited to a different policy term. The AI surfaces the deltas; the analyst resolves them. The output is a reconciled paid/outstanding/incurred table that ties to the carrier's last invoice or Schedule F line item within $1,500 on a $400K loss-incurred book - tight enough to walk into a renewal market call without an asterisk on every number.

The Three-Claim GL Loss Run Worked Example

Consider the canonical mid-market manufacturer with five policy terms 2021-2025, total premium-of-record $1.2M annually, mix of CGL ($420K), WC ($580K), commercial auto ($140K), and umbrella ($60K). The clean output from the AI pass shows three CGL losses worth narrating, one WC, and a flat commercial auto book.

CGL claim 1: 2022-03-14 slip-and-fall on premises, claimant store visitor, paid indemnity $48,000, paid ALAE $11,500, outstanding $0, total incurred $59,500, closed 2023-08-22, no subrogation. CGL claim 2: 2023-09-08 products liability - alleged ingredient contamination in a private-label product line - paid indemnity $0, paid ALAE $73,400, outstanding $375,000 (case reserve set per defense counsel May 2025 estimate of likely jury verdict range), status Open-Litigation, no subrogation but supplier indemnification claim pending. CGL claim 3: 2024-11-02 premises slip with alleged TBI, paid indemnity $0, paid ALAE $14,200, outstanding $85,000, status Open, defense reports liability is contested. WC claim: 2024-06-17 lifting injury L4-L5 herniated disc, claimant warehouse picker, paid indemnity $94,000 (TTD), paid medical $186,000, outstanding $145,000 (MMI projected Q3 2026 with possible PPD rating), status Open, no subrogation, no recovery type. Commercial auto: zero reported losses across all five terms.

The AI's first-pass narrative on this account reads: "Frequency is in line with industry baseline for the SIC (manufacturing, food and beverage co-packer): 0.6 claims per $100K of CGL premium versus the 0.7-0.9 industry expected. Severity is the concern. The 2023 products-liability case at $375K outstanding is the single largest exposure and the open coverage question is whether the supplier indemnification will reduce net incurred. The 2024 premises-TBI case at $85K outstanding has high variance - TBI cases at this defense-reported severity range typically develop to $180K-$425K total incurred in this venue (Cook County, IL). The WC claim is on the expected curve for a lifting injury at MMI projection 14 months out. The recommendation to the treaty broker is to flag the products-liability case as a single-risk reserve-watch item for the umbrella tower and to confirm whether the supplier indemnification recovery has been booked to ceded recoverable on Schedule F."

Frequency, Severity, and the IBNR Signals

The frequency narrative the underwriter actually reads compares the account's claims-per-million-of-premium and claims-per-thousand-FTE against the carrier's internal book benchmarks for the SIC and class code. The severity narrative compares average paid-and-incurred per claim, plus the tail (the 90th and 95th percentile claim) against the same benchmark. AI does the math; the underwriter judges the trends. A 5-year severity trend showing average GL claim incurred rising from $14K (2021) to $48K (2024) is a signal of either bad facts or bad luck or both, and the renewal underwriting note has to articulate which. A WC claim count rising from 4 to 11 across five terms is either a hiring growth story (FTE up 60%, claim count up 175% - concerning) or a class-code shift (added a higher-hazard operation - quote-with-restriction territory).

IBNR signals are where the AI pass adds the most analytical value to a producer or underwriter without an actuary on the team. The model flags every line item where paid-to-incurred ratio looks anomalous (sub-30% paid on a litigated GL file aged 18+ months suggests development pressure to come), every open file aged beyond the carrier's average closure curve for that cause code (the 2024 TBI case at 18 months open with $14K ALAE paid is below the typical $25-40K ALAE curve at this age, signaling either defense-counsel underbilling or upcoming spend acceleration), and every cluster of small-paid frequency claims that may indicate emerging trend (three sub-$10K WC claims in 90 days at the warehouse picking station is a leading indicator the underwriter wants flagged). The model produces the signals; the human reading the file decides what they mean. The CCC IX claims platform's litigation-prediction model and the Shift Technology fraud-screening output feed in as additional severity signals when the carrier subscribes to those data sources.

Reconciling to Schedule F and Schedule P

The treaty broker's submission to the reinsurer - whether for the cat-XOL treaty at Munich Re, the casualty quota-share at Swiss Re, or the working-layer per-risk excess at Hannover Re - requires reconciliation between the account-level loss runs the producer carries and the cedent carrier's NAIC annual statement Schedule F (Part 1 assumed, Part 3 ceded) and Schedule P (loss development by line of business and accident year). The mechanical reality is that the account-level paid-and-incurred numbers on the loss run roll up to a line on Schedule P that includes hundreds or thousands of other accounts, and the reconciliation works in reverse: take the carrier's Schedule P loss-development triangle for the line (CGL, WC, commercial auto), and confirm that the development pattern shown at the carrier book level is consistent with the development the account-level loss run is exhibiting.

An account showing 35% paid-to-incurred at 36 months on its open GL file when the carrier's Schedule P shows 62% paid-to-incurred at 36 months on the CGL line as a whole is a flag - either the account's reserves are heavier than book average (more conservative), or the account's claims are developing slower (longer tails), or the carrier's case-reserve practices on this account are different. The AI pulls the comparison; the appointed actuary or the treaty broker's analyst signs off on the interpretation. The same reconciliation runs against Schedule F when the account has assumed reinsurance (a captive cell, a fronting arrangement) or ceded reinsurance (the producer's prior-carrier excess tower) that needs to be netted before the renewal market call. The output is a reinsurance-review note the treaty broker hands to the reinsurance underwriter with the Schedule P triangle, the account-level loss triangle, and a one-paragraph narrative on where the two diverge and why.

The Reinsurance-Review Note for the Treaty Broker

The treaty broker at Guy Carpenter, Aon Re, Howden Re, or Gallagher Re does not have time to read a 14-page raw loss run. The reinsurance-review note is a one-to-two-page artifact that distills the account into the four things the reinsurance underwriter actually needs: the loss-pick at the working layer ($1M xs $1M, or $4M xs $1M), the loss-pick at the upper layers ($5M xs $5M, $15M xs $10M), the single-risk exposure relative to the $25M single-risk treaty limit (does any individual claim breach the limit and need facultative), and the qualitative narrative on trend (rate adequacy of the original premium, loss-trend assumption embedded, any large open files with material development risk). The AI drafts the note structure and populates the loss-pick math; the producer or treaty broker's analyst writes the qualitative paragraphs.

For the manufacturer above, the working-layer loss-pick over the five years comes to roughly $73K (the closed 2022 slip-and-fall above the $25K self-insured retention plus the WC indemnity penetrating the deductible buyback), but the open products-liability case at $375K outstanding plus the contested TBI at $85K outstanding sits squarely in the working layer and could push the layer loss to $400-$525K if both develop adversely. The single-risk exposure on the products-liability case, if it develops to a $2M settlement (within the defense-counsel range), is well below the $25M single-risk treaty limit and does not require facultative. The note flags the case for reserve-watch and recommends the umbrella tower be reviewed for breach risk in the renewal cycle.

What the Producer, the Underwriter, and the Actuary Each Own

The AI produces the structured table, the frequency/severity narrative, the IBNR signals, the Schedule P reconciliation, and the draft treaty-broker note in 35-45 minutes. That output is not an underwriting decision, not a reserve recommendation, and not a reinsurance commitment. The producer owns the renewal-market strategy: which carriers see this account, what supplemental data goes in the cover memo, whether the products-liability case gets disclosed before the underwriter asks or surfaced when asked. The underwriter at Travelers, Chubb, Hartford, or CNA owns the appetite call: do they bind in light of the open severity, what restrictions or sub-limits or AOP-deductible changes are required, what rate adequacy do they need on the renewal premium. The carrier's appointed actuary, if the account is large enough to warrant individual reserve review, owns the case-reserve interpretation: is the $375K outstanding adequate, redundant, or deficient against the carrier's reserve practice for products-liability cases at this development stage. The treaty broker owns the cession recommendation: which layers, how much, on which terms. AI does the lift; the four humans on the file do the judgment.

The file-note discipline is non-negotiable. NAIC Model Bulletin §4 and the Colorado Reg 10-1-1 algorithm-inventory requirement both require that any AI-assisted underwriting or reserving artifact be traceable: which model produced the output (GPT-4o, Claude 3.5 Sonnet, the internal Convr or Indico extraction model), which version, which prompt template, which input documents, and what human reviewer signed off and on what date. The producer's Applied Epic file note, the underwriter's PolicyCenter or Duck Creek submission note, and the actuary's reserve memo all carry the same model-version stamp and the same prompt-log archive location. Skip the discipline and the next DOI market-conduct exam or the next bad-faith deposition turns the AI-touched artifact into a discovery hazard.

Key Takeaways

• The 9-column canonical loss-run table - date, claim number, policy term, ISO cause code, status, paid indemnity, paid ALAE, outstanding, total incurred, recovery type, subrogation status - is the structural output every downstream artifact (renewal pack, treaty note, reserve memo, Schedule P reconciliation) builds on. Convr, Indico, Hyperscience, Send, and Cytora all produce this shape; differences are in cause-code mapping aggressiveness and integration depth into Applied Epic, AMS360, or the carrier's PAS.
• OCR on 200dpi scanned loss runs hits 96-99% character accuracy on clean inputs and 88-94% on noisy fax-of-a-fax inputs. The bottleneck has moved from extraction to normalization: ISO-8601 date standardization, ISO 17 cause-code mapping (75-85% clean, 15-25% to human queue), and reconciliation against Applied Epic or AMS360 within $1,500 on a $400K incurred book.
• The frequency narrative compares claims-per-million-of-premium and claims-per-thousand-FTE against the carrier's internal SIC/class-code benchmarks; the severity narrative compares average paid-and-incurred and the 90th/95th percentile tail. A 5-year severity trend showing GL average incurred climbing from $14K to $48K is a flag the renewal pack must address before the rate ask lands.
• IBNR signals the AI flags: paid-to-incurred anomalies on aged litigated files, open files past the cause-code closure curve, ALAE-spend below the development curve at the file's age, and clusters of small-paid frequency claims signaling emerging trend. CCC IX litigation-prediction and Shift Technology fraud-screening feed as additional severity signals when subscribed.
• Schedule P reconciliation runs in reverse: the account-level paid-to-incurred development at 12/24/36/48 months gets compared to the carrier's Schedule P CGL/WC/auto triangle, and divergence triggers a flag - heavier reserves, longer tails, or different case-reserve practices on this account. Schedule F handles assumed/ceded netting when captives or fronting arrangements sit in the structure.
• The treaty-broker reinsurance-review note is one-to-two pages and answers four questions: working-layer loss-pick, upper-layer loss-pick, single-risk exposure relative to the $25M treaty limit, and qualitative trend narrative. AI drafts structure and loss-pick math; the producer or treaty broker analyst writes the qualitative paragraphs that Guy Carpenter, Aon Re, Howden Re, or Gallagher Re actually read.
• The producer owns market strategy; the underwriter owns the appetite call; the appointed actuary owns case-reserve interpretation; the treaty broker owns the cession recommendation. AI does the 35-45 minute structural lift instead of 6-8 hours; the four humans on the file own every judgment call.
• NAIC Model Bulletin §4 and Colorado Reg 10-1-1 require every AI-touched loss-run artifact to carry the model name, version, prompt template, input-document list, and human-reviewer sign-off in the Applied Epic / PolicyCenter / Duck Creek file note. Skip the discipline and the DOI market-conduct exam or bad-faith deposition turns the AI output into a discovery hazard.
• The 35-45 minute structured AI pass replaces 6-8 hours of junior analyst time but adds the IBNR signaling, the Schedule P reconciliation, and the cross-account benchmarking the analyst pass historically did not include. The role evolution is not displacement - it is upgrading the producer's, underwriter's, and actuary's read on every account they touch.