● Rottawhite — AI Systems Studio

Submissions arrive as email. They become data.

A submission arrives as a forwarded email with six attachments in four formats. Someone keys it in. That someone is usually expensive.

Test it in a week — $2,500 →How we prove accuracy →

Today

What the desk looks like now

  • Submissions land in a shared inbox and are triaged by hand, often by an underwriter rather than an assistant.
  • ACORD forms, schedules of values, and loss summaries arrive as scans, spreadsheets, and PDFs that do not agree with each other.
  • Clearance against existing business happens by memory or by searching the agency management system one name at a time.
  • Turnaround slips, and the brokers who feel it stop sending you their better risks.

The system

What the agent does

Reads the whole submission

Email body, ACORD 125/126/140, schedules of values, loss summaries, and the attachments that arrive as photographs of a page. Extraction is field-level, not a summary.

Reconciles what disagrees

When the ACORD says one TIV and the SOV totals another, the discrepancy is surfaced as an exception rather than silently resolved in favour of whichever was read last.

Runs clearance

Checks the named insured and address against existing submissions and bound business, catching the duplicate that would otherwise be worked twice by two underwriters.

Writes into your system

Structured records pushed into the agency management or policy admin system, with the uncertain fields flagged for review instead of guessed.

Evidence

What gets logged

The audit trail is the part that makes this usable in regulated work. Every output traces back to a document.

  • Every extracted field linked to the source document and page it came from
  • Every reconciliation decision, with both conflicting values retained
  • Every clearance check and what it matched against
  • Confidence score per field, and which fields were routed to a human
How the audit trail works →

How accuracy is measured

Built from your own documents, including the bad ones. An average that hides the hard cases is not a number worth having.

  • Field-level extraction accuracy against a labelled set of your own submissions
  • Exception rate — the proportion routed to a person — tracked over time
  • Accuracy by document type, so a weak format shows up rather than averaging out
How the eval harness works →

Questions

Common objections

Our submissions are messy and inconsistent. Does that break it?

Messy is the normal case and it is why the eval harness exists. We build the test set from your worst documents as well as your cleanest, so the accuracy number reflects what you actually receive rather than a tidy sample.

What happens when it is not sure?

It says so. Fields below a confidence threshold are routed to a person with the source page attached, which is far faster to check than re-keying from scratch.

Can it write directly into our agency management system?

Where there is an API, yes. Where there is not, we work through the import formats the system already accepts rather than claiming an integration that does not exist.

Also

Other back-office workflows

Workflow

Policy checking

The carrier sends the policy. Somebody is supposed to read all sixty pages and confirm it matches what was agreed. Often nobody does.

Workflow

Certificates of insurance

Low value per unit, enormous volume, zero tolerance for error, and it never stops. Certificates are the definition of work a machine should do.

Workflow

Loss runs

Every carrier formats loss runs differently, none of them formats them well, and an underwriter needs five years of them in one view.

Next step

Start with one workflow.

A week, a fixed fee, and a measured answer on your own documents. If it will not work, you find out for $2,500.

Book a 30-min call →The $2,500 sprint →