Experience rating worksheets, read against your loss runs and audited payroll, with every difference listed.

Today an account manager prints the worksheet and goes through it with the loss runs and the audit statements: the policy periods used, each claim's value, each class code and its payroll. Anything that doesn't line up goes in a note, and someone decides whether it deserves a call to the insurer or the rating bureau.

In the proof of concept, AI reads copies of the same papers and lists every difference, each one traced to its page and line, with what the worksheet shows, what your file shows and the date behind each figure. It doesn't say which is right. A person decides, and a person takes it up with the insurer or bureau.

The page isn't laid out to be matched. NCCI's worksheet runs policy by policy: carrier, policy number and dates, then each class code's ELR, D-Ratio, payroll, expected losses and expected primary losses on the left, and on the right the claim data, the IJ injury code, the OF open or final marker, and actual incurred and actual primary losses. NCCI's own guide notes that a claim on a row has no tie to the payroll beside it. So the draft matches claims by policy and claim number, never by row.

One claim, two shapes. The worksheet adds indemnity and medical into a single figure, so a loss run that splits them has to be summed before anything is compared. Small claims may be grouped by injury type, with a count in the claim data column where a claim number would be, which means there is no line to find. A medical-only claim, injury type 6, shows its whole value in the detail, while the reduction most states have approved appears only in the summary. Set the detail against the summary without knowing that and you report a difference that isn't one.

The rating date decides what belongs. The rating effective date sets the experience period, which usually holds three years of data but can run from under 12 months up to 45, depending on short-term policies, shifted effective dates or a new employer. The mod is generally worked out 60 to 90 days before that date. It also carries a status: preliminary until the state's final values arrive, final, or contingent, meaning audited payroll or loss data was still expected when it was produced. Each check opens by stating the status and listing any policy in your file that the worksheet leaves out.

Bureau states write their own rules. Pennsylvania's plan update effective April 1, 2024 refreshed its worksheet to list all individual losses, not only those above the split point or loss limit, and added a status showing whether a mod has been revised; subrogated amounts come off ratable losses before the split point applies. New York's plan manual, in the 2006 edition on the bureau's site, ties revisions to events: payroll corrected within three years of policy expiration reaches the current and up to two preceding mods, a subrogation recovery up to six, and the revised report must carry its identifier or the mod stays put. A reserve that simply moved later isn't one of those events. So a difference is only a starting point, and the draft doesn't say whether a rule applies.

A difference has a route, and a person walks it. In NCCI states the employer and insurer try first, then NCCI's Customer Service Center, then a dispute process with a Regulatory Services Manager assigned, as Connecticut's Insurance Department explains it. New York takes a written request to the Rating Board, a staff reply within 60 days, a conference and then its Underwriting Committee, and every New York policy since March 1, 2015 carries endorsement WC 31 06 18 saying so. Your reviewer will watch for a policy missing from the period, a claim marked open that your run shows final, a claim under a policy your client never held, and payroll on a code the audit statement doesn't show.

With your reviewer's approval, newly published worksheets go through the same check for about 30 days.

A licensed agent or account manager reads every list and decides what, if anything, goes to the insurer or bureau. The AI doesn't dispute a mod, file a correction or contact anyone.

Scoring continues on the live lists under the criteria set for the test.