AuditGap Premium audit review

For workers' comp premium audit vendors

We review submitted audit documents and flag what's missing before your team opens the file.

A policyholder sends in a payroll register, three 941s, a stack of DE 9Cs, a check register and six certificates. Something is always missing — and finding out which thing costs you a chase cycle you can't spare inside a 60-day window. AuditGap reads the bundle the hour it lands and tells you exactly what to ask for.

Three design partner slots, no cost and no contract. I'm the founder and the only engineer; you'd be talking to me directly.

01

The analysis was never the bottleneck. The chase is.

When documents arrive clean and complete, a desk auditor can close a simple file in an afternoon. That almost never happens. What actually happens is a partial bundle, a wrong date range, a certificate that expired in month seven, and four more weeks of voicemail.

The elapsed time in a premium audit is spent before the analysis, waiting on a small business for whom your audit is unpaid homework. Extraction software helps after the documents land. It does nothing about the part that actually consumes the sixty days.

A specific request gets answered faster than a generic one. That is the entire thesis, and it's why the output here is a drafted ask, not a dashboard.

"Please send your remaining documents" starts another loop. "Please send your 941 for April–June 2023, and a workers' comp certificate for Wilson Painting Services, who you paid $10,259" ends one. The second message takes a trained auditor twenty minutes of cross-referencing to write. This writes it while the email is still open.

02

The five cross-references

Each one compares documents against each other and against the policy period — the gaps a checklist can't see, because they only exist in the relationship between two files. Open any rule to see the language it actually emits.

  1. R1 Quarters that don't line up with the policy period

    Policies run any month to any month; tax filings are locked to calendar quarters. A June-to-June policy needs Q3 and Q4 of one year and Q1 and Q2 of the next. We derive the required set from the declarations page and check both the 941s and the DE 9Cs against it — missing quarters and quarters that fall outside the period.

    Verbatim output

    MediumR1_QUARTER_ALIGNMENT
    Missing Form 941 for 2023Q2

    Policy period requires quarters 2023Q1, 2023Q2, 2023Q3, 2023Q4; no Form 941 was submitted for 2023Q2.

    policy declarations · Form 941 · DE 9C

  2. R2 Wage totals that don't reconcile

    The payroll register is the claim; the 941 and the DE 9C are what the business already swore to the IRS and the state. We bucket register rows by pay date and compare each quarter against line 2 and against subject wages. Small deltas are definitional noise — 401(k), Section 125, officer treatment — so a finding needs to clear both a percentage and a dollar floor before it fires.

    Verbatim output

    HighR2_TOTAL_RECONCILIATION
    Payroll register total does not match Form 941 for 2023Q1

    Payroll register shows $216,997.50 for 2023Q1 but Form 941 line 2 shows $184,447.88 (difference $32,549.62, 17.6%).

    payroll register · Form 941 line 2 · DE 9C subject wages

  3. R3 Subcontractors paid with no certificate on file

    Every payee in the check register whose payments clear the reporting threshold inside the policy period needs a certificate whose named insured matches them. Matching is fuzzy on business suffixes — "Rodriguez Concrete LLC" and "Rodriguez Concrete" are one entity — and strict everywhere else. This is usually the most expensive line in the audit and always the most fixable, if it's found on day one instead of day fifty-five.

    Verbatim output

    HighR3_SUB_MISSING_COI
    No workers' comp certificate on file for Wilson Painting Services

    Wilson Painting Services was paid $10,259.00 during the policy period but no certificate of insurance is on file. Uninsured subs are added to chargeable payroll.

    check register · 1099-NEC · ACORD 25

  4. R4 Certificates that don't actually cover the payments

    A certificate on file is not the same as coverage. We check that a workers' comp row exists at all — general-liability-only is the classic small-business mistake and counts as no certificate — and then run the date-window math payment by payment. Coverage is unioned across every certificate matching that sub, so an expired cert plus its renewal reads as continuous and only a genuine lapse is exposed, with the uncovered dates and the dollars named.

    Verbatim output

    HighR4_COI_VALIDITY
    Workers' comp certificate for Herrera Drywall Inc does not cover all payments

    The COI on file for Herrera Drywall Inc covers 2025-06-01..2025-10-12, but payments fall outside that window on 2026-01-29, 2026-02-26, 2026-04-15, 2026-05-07 (total $32,620.11 exposed).

    ACORD 25 coverage rows · effective/expiration dates · check register dates

  5. R5 People and documents the other documents imply

    A name on the payroll register that appears on no state wage report — or the reverse — is either a filing error or a wage gap, and either way it's a question worth asking before sign-off. Person names are matched at a stricter threshold than business names, because "Jose Garcia" and "Jose Garza" must never merge. The same rule catches a bundle that's simply missing a document type the others depend on.

    Verbatim output

    MediumR5_IMPLIED_MISSING
    Employee Eric Garcia is missing from the SUTA reports

    Eric Garcia appears in the payroll register but on no state wage report (SUTA) covering the policy period.

    payroll register ↔ DE 9C · full-bundle completeness

What it does not do: assign class codes, make officer determinations, or decide anything that carries dispute exposure. Those are judgment calls and they stay with your auditor.

03

What comes out the other end

Two files per audit. A structured gap report your system can ingest, and a follow-up request in plain language your auditor can edit and send. Both of the exhibits below are unedited output from a single run.

gap_report.json — Johnson Concrete, Inc.
MediumR1_QUARTER_ALIGNMENT
Missing Form 941 for 2023Q2

Policy period requires quarters 2023Q1, 2023Q2, 2023Q3, 2023Q4; no Form 941 was submitted for 2023Q2.

HighR3_SUB_MISSING_COI
No workers' comp certificate on file for Wilson Painting Services

Wilson Painting Services was paid $10,259.00 during the policy period but no certificate of insurance is on file. Uninsured subs are added to chargeable payroll.

followup_request.md

Follow-up request — Johnson Concrete, Inc.

Policy number: WC929872815

Thank you for sending your audit documents. To finish the review we need a few things from you:

Documents to send

  • Please send your IRS Form 941 for April–June 2023 — the quarterly wage report your accountant files.
  • Please send a workers' comp certificate of insurance for Wilson Painting Services, who you paid $10,259.00 during the policy period. Without it, those payments get charged to your premium.

Every finding carries its rule, severity, the entities and quarters involved, the dollar exposure where it's computable, and the source file each number came from. Extraction is the only model-driven step — every decision after it is deterministic code, so the same folder always produces the same report, and a wrong answer is a bug someone can point at rather than a reroll.

04

Where accuracy stands, stated honestly

50synthetic companies in the test corpus
46 / 46planted defects detected
0false flags across 15 clean files
115automated tests on the pipeline

The caveat, because you'd find it anyway: those files are synthetic. I generate the companies, plant the defects, and score the pipeline against its own ground truth. That proves the rules are correct and deterministic. It does not prove they survive your scans, your carriers' formats, or a bookkeeper's photograph of a payroll report taken at an angle. That is exactly what a pilot on your closed files is for, and the number that matters is the one we measure there — against what your auditors actually found.

05

How it runs

01

You send the folder

The documents as they arrived — PDFs, scans, whatever the bookkeeper exported. For pilots, redacted: strip employee names and SSNs. We need structure and numbers, not identities.

02

It reads and cross-references

Each document is classified, parsed into typed records, then checked against the others and against the policy period. Five rules, run in a fixed order, with every threshold in one config file.

03

You get two files

A gap report with severities, dollar exposure and document references, and a drafted follow-up request. Your auditor edits and sends it, or your system ingests the JSON.

06

What this is not, and doesn't have yet

  • No customers yet. You would be among the first three. I'd rather write that here than have you discover it on a call.
  • It does not assign class codes and does not make officer determinations. Those are judgment calls with real dispute exposure, and they stay with your auditor.
  • One state today. California, DE 9C-style per-employee wage reporting. Additional states are rulebook work, not a configuration flag, and I'll quote them honestly.
  • No SOC 2 yet — a real answer, not a dodge. What exists today: pilots run on redacted files, encrypted in transit and at rest, deleted on a written schedule after analysis, and a DPA I'll sign. SOC 2 comes when a carrier-adjacent contract requires it.
  • No payroll-provider integrations and no carrier system connections. It reads submitted documents. That is the entire surface today.
  • It will not replace your auditors. It removes the first pass — the read-everything-and-list-what's-missing pass — so a file reaches a human already triaged.
07

The pilot, in full

Three design partner slots · no cost · no contract

  1. Pick 5–10 closed audit files. Closed means zero risk to a live audit, and it means ground truth already exists.
  2. Redact them. Employee names and SSNs out. We need document structure and numbers, nothing that identifies a person.
  3. You get gap reports within a week. One per file, plus the drafted follow-up requests.
  4. You compare against what your auditors actually found. Both directions count — what we caught that they caught, and what either of us missed.
  5. If the accuracy is there, we talk about a paid per-document pilot. If it isn't, you've spent an hour of somebody's time and I've learned where the rules break, which is worth the week to me either way.
08

Who you'd be dealing with

I'm Shishir. I built AuditGap myself — the synthetic document generator I test against, the extraction layer, the rule engine, and the scoring harness that tells me when I've broken something. There is no team behind me and no sales department. When you email, you get the person who wrote the code.

I'm an engineering physics student at IIT Bombay, building this alongside coursework. I'm telling you that in the second paragraph rather than letting you find it later, because in a business whose entire job is verifying what people claim, being caught softening a detail would cost me more than the detail ever could.

What I can offer that a larger vendor can't: you talk directly to the engineer, a rule you need changed ships in days rather than a quarter, and you'd be shaping what this becomes rather than filing tickets against it. What I can't offer: a support organisation, a compliance binder, or a reference customer. That's the trade, stated plainly — decide whether it's a good one for a first pilot on closed files.

— Shishir

09

Getting in touch

Email
shishirvsh285@gmail.com
If you'd rather talk
Say so and name a window that suits you. I'm in India, so I'll be calling you at an hour that's inconvenient for me and convenient for you — which is the correct way round.
What happens next
A reply from me, usually within a day. Not a sequence, not a demo request form, not a calendar wall.

If you're only curious what the competition is doing

Ask. I'll tell you what I know about how carrier-side audit automation is currently sold and where I think the vendor-side gap is. I'm trying to understand this industry properly, and the fastest way to do that is talking to people who run the desks — even if you never buy anything.

The questions I keep asking: how many files a desk auditor closes in a week, what a completed audit costs you all-in, and which audits you lose money on.