Mid-term endorsement processing agent

A mid-term endorsement processing agent loads a change request, drafts the endorsement, and stops. It calculates a premium-impact line only when it can cite both the request and the rating table vintage. If either cite is missing, that field stays empty. An underwriter or endorsement ops lead still issues in the policy administration system. The agent does not auto-issue, and it does not invent a premium delta.

Issuance is the legal act. The draft is a work product sitting in a queue. A number that cannot be sourced, or a status that looks bound before the PAS transaction exists, is a servicing error dressed as automation.

Run the desk in this order every time: load the request, draft with cites, leave blanks, then have a human issue. The failure modes (a draft with no table vintage, treating the draft as issued, inventing a premium delta) live inside those steps. Drill them there, not in a separate after-action list.

Start from the inbound artifact, not from the policy you think you remember.

The request may arrive as a broker email, an agency note, a CRM task, or a transcript from ai voice agent for policy servicing. Capture the request identifier, channel, policy number, requested effective date, and the coverage or interest being changed as the first block of the draft. If any of those are missing, do not draft premium and do not draft operative wording. Assemble a stub that lists what is unknown.

Policy administration platforms such as Guidewire or Duck Creek, CRM workbenches such as Salesforce, and agency systems such as Applied Epic hold pieces of the same change. Treat them as a class of systems of record. A closed task in CRM or AMS is not proof that the PAS accepted an endorsement. The agent reads. It does not promote a workflow status into a contract change.

One illustrative pass: a commercial package insured asks, mid-term, to add a newly leased warehouse at a stated building limit, effective the first of next month. The request names the location, the requested limit, and the effective date. It does not name which rating table vintage the book is on. Load that request as written. Copy location and limit into the draft with a cite back to the request. Refuse to compute a premium delta until the table vintage is on the file.

Do not pull last renewal's table because it is nearby. Mid-term rating on the wrong vintage is a different product than the one the insured bought. Freeze the table you will use, or leave premium blank. That is the same discipline as stale-quote repricing lock: no silent substitution of an expired rate file.

A draft is usable work only when it cites the change request and, for any premium movement, the rating table vintage. Both cites are required before a premium-impact line may appear. Non-rating wording (for example, a mortgagor clause with no premium effect) can cite the request alone. The moment a dollar movement appears, the vintage cite is mandatory.

If the vintage is missing, the premium field stays empty. Empty stays empty. Do not interpolate from a similar location, a sibling policy, or last year's endorsement on the same account. Inventing a premium delta is a hard fail. It looks decisive in a queue and is unauditable when a complaint or a market-conduct review asks where the number came from.

If the request is missing a fact the wording needs (occupancy, requested limit, additional-insured interest, square footage), the corresponding clause stays empty rather than filling with a plausible default. Endorsement language drafting is the companion skill for turning a cited request into form-consistent wording. This agent does not replace that work. It refuses to emit a clause that has no cite.

A draft with no table vintage is the usual "almost useful" failure. Location, limit, and effective date are in, and a premium line has a number because the model completed the pattern. Reject that draft. Do not edit around the number. Delete the delta. Leave the field blank. Return it with an explicit note: vintage not cited, premium not calculated.

The same rule applies if the vintage is present but the rating step cannot complete from the request facts. Incomplete exposure is not a license to guess. Cite the vintage, leave premium blank, and list the missing facts by name.

When both cites exist and the calculation completed without substitution, the draft should contain:

• Request cite: identifier, channel, requested change, requested effective date.
• Vintage cite: table or rate-edition identifier and as-of date, copied from the PAS or the rating source of record, not inferred.
• Proposed wording only for clauses the request actually supports.
• A premium-impact line only if both cites are present and no value was substituted.
• Open items: every blank, named.

It must not contain an issued timestamp, a bound indicator, a PAS transaction id that was never written, or a premium that appeared without a vintage cite. Do not label a blank premium "estimated." Estimated is invented.

Issuance is a human action in the policy administration system. The underwriter or endorsement ops lead reviews the cites, fills blanks they can source from the file, or returns the item. They issue through the same controls they use for a fully manual endorsement.

Do not auto-issue. A completed-looking draft is not authority. Treating the draft as issued is the second hard fail. Notices go out, billing cycles the wrong amount, and certificate holders see a change that is not on the contract. Reversal costs more than leaving the item in draft for another hour.

After a human issues, issuance accuracy verification compares the issued form, effective date, and premium to the cited request and vintage. Verification is a separate step. Do not collapse drafting and checking into one agent that "does the endorsement." This agent stops at a cited, or honestly blank, work product.

Keep queue status honest. Distinguish draft, returned for missing cite, ready for issue, and issued. Ready for issue still means a person will press issue. The PAS write happens from the human's session, not from the model's completion callback. Confirmation documents generate after issue, from PAS output, not from the draft file. The agent may assemble a confirmation packet template. It must not date that packet as mailed or delivered, and it must not send it.

If a producer asks whether "the endorsement went through," the answer is whatever the PAS shows as issued, not whatever the draft queue shows as complete. Salesforce or Applied Epic activity can close after the Guidewire or Duck Creek issue, not before.

Three patterns appear as soon as people trust the formatting.

A draft with no table vintage and a filled premium. Kill the number. Do not round it, do not park it in a comment, do not keep it as a whisper in the file. The vintage cite is the permission slip for any delta. Without it, premium stays empty and the item goes back.

Treating the draft as issued because the wording looks final and the premium looks precise. Visual completeness is not issuance. If the workbench can print a confirmation from the draft, watermark it DRAFT until a PAS transaction exists, or disable that path.

Inventing a premium delta when the request is thin or the vintage cannot be read. The model will try to be helpful. The operating rule is the opposite: no cite, no number. Send a blank premium to an underwriter with a named missing vintage. Do not mail an insured a change you cannot defend.

A quieter fourth pattern: mixing systems of record. A CRM task marked done is not a PAS issue. Close CRM and AMS items after the PAS transaction, and only then generate confirmation from PAS output.

Keep the agent on a short leash: load the request, draft with cites, leave blanks, humans issue. That is the whole job. Everything else is how you stop a clean-looking endorsement from becoming the wrong contract.