Product
How automated certificate reading is designed to work
COI uploads and human review are available today. Automated AI field extraction, confidence routing, and automatic requirement checks are currently unavailable.
What is available today
A subcontractor can upload a certificate of insurance, and your team can open the original PDF or image, review it, record the relevant values, and make the customer decision. The file and review history stay together so a later reviewer can see the evidence behind that decision. What Sealinn checks is the field-by-field human checklist for that review.
Automated reading is currently unavailable
New uploads are not automatically sent for AI field extraction. Sealinn does not currently assign field-confidence scores, route a new document by confidence, or automatically check extracted values against requirements. Those steps require a future, reviewed activation.
Authorized historical records can still contain values and confidence context created when extraction was previously used. Keeping that history visible does not mean a new upload is processed automatically now.
What the automated reader is designed to do
When automated reading is enabled, it is designed to inspect an ACORD 25 field by field rather than treat the page as one blob. The supported design covers the named insured, the insurer behind general liability, supported policy numbers and dates, selected liability limits, the ADDL INSD and SUBR WVD columns, certificate-holder details, and clearly labeled extra fields.
The planned output attaches a confidence score to each supported value. That matters because one date can be sharp while one limit on the same scan is unreadable. A document-wide average would hide that difference.
Four fields are designed to carry more confidence weight
If automated reading is enabled, the general-liability per-occurrence limit, the general aggregate, the policy expiry date and the named insured are designed to carry the most weight in the displayed extraction-confidence summary. That score would describe the reading, not establish that coverage exists or that a policy is in force.
How confidence routing is designed to work when enabled
- Every returned value at 95% or better, with no requirement violation in the extracted data — the planned view can present the result without a low-confidence warning, but it still needs the authorized reviewer's decision.
- Any returned value from 75% up to 95% — the planned view highlights the uncertain fields for closer review.
- Any returned value below 75%, or a required value that cannot be read — the planned view keeps the document in the review queue and identifies what needs attention.
A person still decides whether to approve it; the score never grants approval by itself. The authorized reviewer should compare the original certificate with the recorded values and any separate endorsement evidence before making a decision.
What automated reading would not prove
- A certificate is evidence, not the policy. Reading a checkbox does not prove that the endorsement exists; ask for the endorsement itself.
- A confidence score is not an accuracy guarantee. Sealinn has not published a measured accuracy percentage against a labeled certificate corpus.
- A clean read is not a coverage opinion. The customer decides what evidence is sufficient for its work and should involve its insurance adviser where appropriate.
- Automation would not remove human review. File-safety, availability, spend, and queue controls can require a manual path even after the feature is enabled.
For the current workflow, use the original certificate as the source of truth, record the human review, and keep any uncertainty visible in the decision note. Reviewing a certificate describes that process.