October 12, 2026 · Bindmatic
The certificate dispute starts on page 41 of a subcontract nobody read. Here is a pipeline that finds the insurance article, extracts requirements into a schema with page citations, compares them against the policy as issued, and hands a CSR a gap list while the policy can still be endorsed.
Read moreOctober 5, 2026 · Bindmatic
Prose is not the deliverable; the record is. Here is the schema-first extraction loop we use: constrained decoding, a second validation pass with your own business rules, an explicit way to say "not present", and one repair attempt before the item goes to review.
Read moreSeptember 21, 2026 · Bindmatic
Carrier download tells you a commercial policy changed. Your AMS tells you what it used to be. Here is an agent that parses the AL3 transaction, diffs it against the policy record, and drafts one activity a CSR can act on.
Read moreSeptember 21, 2026 · Bindmatic
Most agent projects that stall do not stall on the model. They stall on blank expiration dates and three records for the same insured. Here is the read-only profile we run in the first week, and the thresholds we use to decide what to build.
Read moreSeptember 21, 2026 · Bindmatic
Half your certificate requests arrive as a link to a compliance portal, and the requirement grid never reaches the mailbox. Here is how to fetch it, normalize it into the request schema you already have, and classify the rejections that come back.
Read moreSeptember 18, 2026 · Bindmatic
The certificate is drafted and the CSR clicked approve. Now the reply has to leave from the agency mailbox, land in the requester's original thread, and file itself in the AMS. Here is the Microsoft Graph path.
Read moreSeptember 17, 2026 · Bindmatic
A producer forwards one 94-page PDF: ACORD 125, two loss runs, a supplemental, and a lease. Here is how to split and label it by page before any extraction runs.
Read moreSeptember 16, 2026 · Bindmatic
Property submissions arrive as a 214-row spreadsheet with merged headers and subtotal rows in the middle. Here is how to parse an SOV deterministically, reconcile the TIV, and hand the producer a gap list.
Read moreSeptember 15, 2026 · Bindmatic
Decision logs prove what the agent decided. Traces tell you where a run spent 42 seconds and which call failed. Here is the span layout, the gen_ai attributes, and what must never leave the agency.
Read moreSeptember 14, 2026 · Bindmatic
Your AMS activity codes already are the taxonomy. Here is how we build the classification step of a service-request agent on them: 300 labeled messages, rules first, a schema-constrained model call, and a margin test that sends the doubtful ones to a person.
Read moreSeptember 11, 2026 · Bindmatic
Token totals per call tell you nothing. Here is how we meter cost per approved draft on an AMS agent, cap it per task, per run, and per day, and cut it in the order that actually works.
Read moreSeptember 10, 2026 · Bindmatic
A scored silent run tells you the agent is safe to turn on. It does not tell you it still works after the model changes underneath it. Here is the pin, the frozen case set, and the gate we run weekly.
Read moreSeptember 9, 2026 · Bindmatic
Vendor accuracy numbers were measured on someone else’s documents. Here is the harness we use to score an agent against 200 closed cases from your own AMS, and the gate we set before it goes live.
Read moreSeptember 9, 2026 · Bindmatic
Every carrier prints loss history differently, and a submission needs five years of it. Here is the classify, extract, validate pipeline that turns a loss run PDF into claim rows a producer can check.
Read moreSeptember 9, 2026 · Bindmatic
Most inbound service requests name no account and no policy. Here is the extract, retrieve, score pipeline we use to match a message to an AMS account, and the margin test that tells it when to abstain.
Read moreSeptember 9, 2026 · Bindmatic
No AMS gives you an idempotency key, and every retry is a second activity on the account. Here is the claim-call-confirm layer we put in front of Epic, HawkSoft, and EZLynx, plus the no-op mode the silent run depends on.
Read moreSeptember 9, 2026 · Bindmatic
MCP is now the default way models call tools, and it is a good fit for an AMS. Here is a read-only server over agency data, the four tools we start with, and why writes stay out of it.
Read moreSeptember 9, 2026 · Bindmatic
Every document your agent reads was written by someone outside your agency. Here are the four controls we build in so a sentence in a certificate request cannot steer the workflow.
Read moreSeptember 9, 2026 · Bindmatic
Issuing the ACORD 25 is the easy half. Here is how we index endorsement forms out of the AMS document store and answer, with a citation, whether the holder's additional insured wording is actually on the policy.
Read moreSeptember 9, 2026 · Bindmatic
Loss runs and renewal status still live behind a carrier login. Here is how to build a Playwright connector that fetches them on a schedule, handles MFA, and tells you the morning it breaks instead of the week after.
Read moreSeptember 9, 2026 · Bindmatic
Six weeks after a certificate goes out, someone will ask who decided the coverage was there. Here is the append-only decision log we build next to every agent, what goes in it, and what stays in tracing instead.
Read moreAugust 27, 2026 · Bindmatic
Every agent send should wait for a licensed person. Here is the callback pattern that lets a workflow pause for three days without losing state or sending twice.
Read moreAugust 19, 2026 · Bindmatic
Textract's Queries feature answers plain-English questions against a document. Here is how to point it at ACORD forms from Python, and what to do with the confidence scores.
Read moreAugust 9, 2026 · Bindmatic
The remarket window on a commercial account closes long before the carrier's terms arrive.
Read moreJuly 29, 2026 · Bindmatic
Issuing a certificate is fast; checking it against the policy is the job that disappears under time pressure.
Read moreJuly 16, 2026 · Bindmatic
A buyer's checklist for principals at Epic, HawkSoft, or EZLynx agencies evaluating AI tools or a custom build.
Read more