Skip to content

Systems and records

Quoted is not bound: an insurance quote bind gap dataset schema

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

To distinguish quoted from bound in training data, record each state as a dated event with its own evidence and authority instead of one overwritten status. A quote is an offer with a version and expiry. Bound needs a carrier confirmation, binder or policy from a party with authority to bind, plus cleared subjectivities. Validate transitions, keep not-taken quotes, and label conflicts instead of repairing them.

Evaluation split

Quoted is not bound

This page covers quoted is not bound: an insurance quote bind gap dataset schema.Illustrative

Key takeaways

  • Quoted, bind requested, bound subject to conditions, bound and issued are different states, and each needs its own evidence.
  • A certificate of insurance, a producer email or a CRM stage is never enough evidence for a bound label.
  • Store the authority reference for every bind so the model learns when to refer instead of confirm.
  • A transition table rejects impossible sequences automatically and catches many status errors before annotation.
  • Keep declined, expired and not-taken quotes so the dataset reflects the real quote to bind funnel.

What separates a quoted state from a bound state in insurance records?

A quote is an offer with terms and an expiry; a bound risk is coverage the insurer is obligated to provide, usually evidenced by a binder or a carrier confirmation until the policy is issued. The gap between the two is where labeling mistakes are easiest to make, because the same account can sit in a quoted, bind-requested or bound state for days, and the documents in the file do not always say which.

Statute gives the cleanest definition of the bound side. California's Insurance Code section 382.5 describes a binder as a writing that identifies the insured, describes the coverage and temporarily obligates the insurer until a policy is issued. The same section limits a binder to the period it states, no longer than 90 days from execution, and says it stops being valid once the policy is issued. Other states word this differently, so the dataset should store the jurisdiction with every binder rather than assume one rule. California's Insurance Code also sets separate notice rules when an insurer cancels or rejects a binder or other temporary evidence of insurance, so a bind_cancelled event should carry its notice date.

Three other documents get mistaken for proof of binding. A certificate of insurance is informational: ACORD's own certificate FAQ says a certificate does not amend or extend the policy and that only an endorsement, rider or amendment changes coverage. A bind request from the producer is an instruction, not an acceptance. A CRM opportunity marked won records a sales outcome, not an insurer's obligation. Training data has to keep these as separate events with separate evidence.

Why do agents trained on agency records mistake quoted for bound?

Agents confuse the two states because many operational systems store a single, overwritten status and lose the evidence trail that justified each change. When the status field says bound, the model cannot see whether that came from a carrier confirmation, a producer's optimistic update or a renewal that rolled forward automatically.

The cost of that confusion is not symmetric. An agent that calls a bound risk merely quoted creates a follow-up task; an agent that tells an insured or a certificate holder that a quoted risk is covered creates an uninsured loss and an errors and omissions exposure for whoever relied on it. That is why the dataset should be built to make false bound labels rare, even at the price of more records marked as unresolved.

The typical failure patterns are predictable once you look at a real agency file:

  • Status overwritten in place: the agency management system shows bound today, but there is no record of when the state changed or which document caused it.
  • Email language read as acceptance: phrases like we are good to go or please bind arrive from the insured or producer, not the carrier, yet a model trained on the thread learns them as binding events.
  • Certificates issued early: a certificate request is processed while the account is still quoted, so the certificate date precedes any binder.
  • Quote versions collapsed: the second or third revised quote replaces the first, and the bound terms are compared against the wrong version.
  • Subjectivities ignored: the carrier binds subject to conditions such as a signed application or an inspection, and nobody records whether those conditions were cleared.
  • Not-taken outcomes missing: quotes the insured never accepted are deleted or left open, so the dataset contains only the accounts that converted and a model learns that quotes usually bind.

What should the quote bind and confirmation state schema include?

The schema should treat every state as an event with its own evidence, actor and authority, and keep the quote version the event refers to. The table below is the minimum field set; a team can add line-specific fields, but removing any of these makes the quoted versus bound label impossible to audit.

The industry's own message design already treats these as different transactions: ACORD's property and casualty data standards list new business quote and new business submission as separate business cases. Mirror that separation in the dataset instead of flattening it into one status column.

What should the quote bind and confirmation state schema include?
FieldTypePurposeValidation rule
account_refPseudonymous IDGroups every event for one insured and one coverage requestSame token across quote, binder, policy and emails
quote_version_idStringIdentifies which offer terms an event refers toRequired on quoted, bind_requested and bound events
stateEnumsubmitted, quoted, quote_expired, bind_requested, bound_subject_to, bound, policy_issued, declined, not_taken, bind_cancelledMust be a value from the enum; no free text
event_timeTimestamp with zoneWhen the state took effect, not when someone typed itMonotonic within an account; entry_time stored separately
evidence_typeEnumcarrier_confirmation, binder, policy, quote_document, bind_request, certificate, email, system_statusbound requires carrier_confirmation, binder or policy
evidence_refDocument ID and pagePoints to the document that proves the stateMust resolve to a document in the record set
actor_roleEnumWho caused the event: carrier underwriter, MGA, producer, insured, systembound cannot be caused by insured or system alone
authority_refStringDelegated authority or underwriting approval that allowed bindingRequired when actor_role is producer or MGA
subjectivitiesList of condition, status, cleared_timeConditions attached to a bindbound_subject_to moves to bound only when all are cleared or waived
binder_valid_untilDateExpiry of temporary coverageChecked against the jurisdiction's binder rule; null once policy_issued

Which state transitions should the dataset allow?

Allow only the transitions an insurer's workflow can actually produce, and reject the rest at validation time. A short transition table does more for label quality than any amount of annotator training, because it catches impossible sequences automatically.

Transitions not listed here, such as quoted directly to policy_issued with no binding evidence, should fail validation and be routed to a reviewer, not silently repaired.

Which state transitions should the dataset allow?
FromToRequired evidenceTypical actor
submittedquotedQuote document with version, terms and expiryCarrier underwriter or MGA
quotedquote_expiredExpiry date passed with no bind requestSystem, with the expiry date as evidence
quotedbind_requestedWritten instruction referencing the quote versionProducer or insured
bind_requestedbound_subject_toCarrier or MGA confirmation listing open conditionsCarrier underwriter or MGA
bound_subject_toboundEvery subjectivity cleared or formally waivedCarrier underwriter or MGA
bind_requestedboundBinder or carrier confirmation with no open conditionsCarrier, MGA or producer within authority
boundpolicy_issuedIssued policy that supersedes the binderCarrier
quotednot_takenInsured declined, chose another carrier or stopped respondingProducer

Who had authority to bind, and how does the record prove it?

Bound is only a valid label when the party that bound had the power to do it, so the record needs an authority reference, not just an actor name. In open-market placements this is usually the carrier's underwriter; in delegated business it is a managing general agent or coverholder acting under a written agreement with limits.

Lloyd's definitions byelaw is a clear primary example of how delegated binding is formalized: it defines a coverholder as a company or partnership authorized by a managing agent to enter into insurance contracts underwritten by a syndicate's members in accordance with a binding authority. In the United States, similar arrangements are commonly run through managing general agents. Whatever the market, the dataset should store which agreement applied and whether the risk sat inside its limits.

Store the authority boundary as data. A risk outside the delegated class, limit or territory should produce a referral event and a carrier decision, never a bound event signed by the agent. These referral records are some of the most useful training examples, because they teach an agent to stop and route rather than to confirm coverage it cannot give. For how underwriting referrals and exceptions are labeled more broadly, see the linked underwriting rationale guide.

A worked record that passes and one that fails

Both records below are illustrative and invented to show structure; neither describes a real insured, carrier or licensed dataset. The first passes validation. The second looks plausible in an agency system but fails three checks.

At a glance

A worked record that passes and one that fails

Passing record (illustrative): account_ref ACCT-0417; quote_version_id Q-0417-v2; events in order: submitted, quoted on v1, quoted on v2 after a revised limit, bind_requested on v2 by the producer with an email as evidence, bound_subject_to by the carrier underwriter with a carrier confirmation listing two subjectivities (signed application, loss control inspection), application cleared after 3 days, inspection cleared after 12 days, bound, policy_issued 24 days after the bind request, binder_valid_until set to null on issue.

Failing record (illustrative): account_ref ACCT-0522; events: quoted on v1, certificate issued, state set to bound by the system with evidence_type system_status, policy_issued later on v2 terms.

  • Failure 1: bound has evidence_type system_status, which cannot prove binding; the validator requires a carrier confirmation, binder or policy.
  • Failure 2: the certificate event precedes any bound event, so the certificate was issued against a quote.
  • Failure 3: the policy references quote version v2, but no event ever moved v2 into quoted, so the bound terms cannot be traced to an offer.
  • Correct handling: keep the record, label it as a conflict case with these three findings, and use it as a negative example rather than repairing the status by hand.

When the binder, the email and the CRM disagree, which one wins?

When sources disagree, rank them by who carries the obligation: the carrier's policy and confirmation outrank the binder, the binder outranks a producer's email, and system statuses rank last. Keep the losing evidence in the record as a labeled conflict so the model learns what a misleading signal looks like.

The validator checks below turn that precedence into rules a team can run on every account before it enters a training or evaluation split:

  • Every bound or bound_subject_to event cites a carrier_confirmation, binder or policy document that resolves to a page in the record set.
  • No certificate event precedes the first bound event for the same account and coverage.
  • Every bound event references a quote_version_id that previously appeared in a quoted event.
  • binder_valid_until respects the binder duration rule of the stored jurisdiction; for a California binder within the scope of section 382.5, no later than 90 days from execution.
  • A bound_subject_to account with uncleared subjectivities is never labeled bound, even if a policy was later issued; record the waiver if there was one.
  • Accounts that never bound end in quote_expired, not_taken or declined, so the class balance reflects the real quote-to-bind funnel rather than only the wins.
  • event_time is the effective time from the document; entry_time records when staff typed it, and the two are never merged.

How should accounts be split between training and evaluation?

Split by account and by time, never by event. If the quoted event of an account lands in training and its bound event lands in evaluation, the model has effectively seen the answer, and the evaluation score says nothing about how it will behave on a new submission.

A practical split for quote and bind histories follows four rules. Assign every event for one account_ref to the same split. Hold out the most recent period of quote activity as the evaluation set, so the model is tested on accounts it could not have seen. Oversample conflict cases and referral events in evaluation, because they are rare in daily work and are exactly where an agent is most likely to confirm coverage that does not exist. Report results separately for open-market and delegated-authority business, since the evidence for a bind looks different in each.

Renewals deserve their own flag. A renewal offer can look like a quote, but the expiring policy is still in force, so an agent that treats the renewal as unbound may tell an insured they have no coverage when they do. Mark renewal quotes with a link to the expiring policy and keep them out of the new business counts.

Finally, write down what the dataset cannot show. Phone calls in which an underwriter verbally agreed to bind, and paper binders that were never scanned, leave gaps. Record those gaps as missing evidence on the account rather than inferring the state, and tell downstream users how many accounts carry them.

Preparing quote and bind histories for licensing

Quote and bind files contain named insureds, addresses, loss histories and sometimes financial statements, so they need consistent pseudonyms across every document before they leave the agency. The NAIC Insurance Data Security Model Law treats information concerning a consumer, and business information whose disclosure would harm a licensee, as nonpublic information, which is a useful reminder that commercial accounts carry confidentiality obligations too.

Two rights questions come before preparation. First, many quote documents arrive on standardized forms: ACORD's forms index lists the ACORD 75 Insurance Binder, and ACORD's forms FAQ says its forms may only be used under a valid license, so the form layout and the values an agency typed into it are different rights questions for counsel. Second, carrier quotes and confirmations may be governed by agency or delegated authority agreements that restrict reuse. Having the records in your system is access; it is not the same as permission to license them.

An agency, MGA or insurance software company weighing whether these histories could become a licensed dataset can use SourceX for the whole path, from sourcing and rights review through contracting, preparation, delivery and payment. Nothing is delivered until your company has approved the scope and terms and signed, and the first step, a short company assessment, needs no data upload.

Frequently asked questions

Is a certificate of insurance proof that a risk is bound?

No. ACORD's certificate guidance describes the certificate as informational and says it does not change the policy; only an endorsement, rider or amendment does. In a training dataset, a certificate is an event of its own and should never be used as the evidence for a bound label.

What is the difference between bound and issued?

Bound means the insurer is obligated to provide coverage, usually shown by a binder or carrier confirmation. Issued means the policy itself exists and supersedes the binder. California's section 382.5, for example, says a binder stops being valid once the policy is issued, so a record should close the binder when the policy event arrives.

Should declined and not-taken quotes stay in the dataset?

Yes. Without them, a model only sees accounts that converted and learns that quotes usually bind. Keep them with an explicit end state and the evidence for it, such as the insured's decline message or the quote expiry date.

How should subjectivities be labeled?

Store each condition, its status and the time it was cleared or waived. Use a bound_subject_to state until every condition is resolved, and treat a record that skips that step as a conflict case for review rather than a clean bound example.

Can a producer's email be the evidence that a risk is bound?

Only if the producer held written binding authority for that class and limit, and the record stores which agreement applied. Otherwise the email is a bind request, and the bound label needs a carrier confirmation, binder or policy behind it.

Sources

  • California Insurance Code section 382.5 defines a binder as temporary coverage until the policy issues, valid for its stated period up to 90 days and not after the policy is issued. Source
  • California's Insurance Code sets separate notice rules when an insurer cancels or rejects a binder or other temporary evidence of insurance. Source
  • ACORD's certificate FAQ says a certificate of insurance is issued for information only and that only an endorsement, rider or amendment changes coverage. Source
  • ACORD's forms index lists the ACORD 75 Insurance Binder and the ACORD 76 Binder Log. Source
  • ACORD's forms FAQ says ACORD forms may only be used with a valid license. Source
  • Lloyd's definitions byelaw defines a coverholder as authorized by a managing agent to enter insurance contracts for a syndicate under a binding authority. Source
  • ACORD's property and casualty standards cover business cases including new business quote and submission, with XML built for request and response messages. Source
  • The NAIC Insurance Data Security Model Law defines nonpublic information to include consumer information and business information whose disclosure would harm a licensee. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies