You never delegated the claim decision. You delegated a number, and then let the number decide. The vendors agree: their own documentation says the estimator decides. So the gap in most claims operations is not vendor overreach. It is buyer abdication.
This piece is about governing third-party data in claims decisions you already depend on: the estimating database, the repricing engine, the valuation service, the severity score. Reviewing a vendor you have not bought yet is a different exercise, covered in procurement and AI vendor risk. The argument sits inside our wider guide to automating the claims lifecycle.
The insurer owns the decision even when someone else produced the number
Answer
Is an insurer responsible for third-party data used in a claim decision?
Yes. A claims organisation answers for every number behind a decision, including the ones it bought. Washington states it outright: insurers are responsible for the accuracy of their evaluations, including evaluations made on their behalf. The vendor supplies an input. The insurer makes the decision, signs it and defends it.
The cleanest statement of the principle is in force right now. WAC 284-30-380(7) reads: "Insurers are responsible for the accuracy of their evaluations including, but not limited to, evaluations made on their behalf, to determine the amounts owed under the applicable insurance policy." It was re-adopted in 2026 inside a 15-section claim-handling package, effective 18 October 2026.
The vendor publishes an input. The insurer publishes a decision.
Responsibility attaches at the moment the number stops being an input and starts being the answer.
- 01Vendor reportsA published price, allowed amount, value or score leaves the vendor
- 02File receivesThe number lands in the estimate, the reserve or the referral queue
- 03Decision issuesPay, reduce, deny or refer, in the insurer's name and signature
- Unfair-practices exposure
- The claimant's question
- The examiner's file pull
The legal theory is older than any AI rule. Writing for IRMI in 2010, Tim Ryles described the insurer's duty to investigate as nondelegable: "Generally, irrespective of the quality of the vendor's work, the insurer is accountable for it."
A contract can move money between the parties. None of that moves the regulatory duty. When a claimant or a market conduct examiner asks why, the answer cannot be a vendor name. It has to be a record.
Five classes of borrowed input sit inside a P&C claim file
Answer
What kinds of third-party data sit behind a property and casualty claim decision?
Most claim decisions rest on five kinds of input the carrier did not build: estimating and pricing databases, medical bill review and repricing engines, total-loss and severity valuation services, risk and fraud scores, and general model output. Each has a different owner of the method, a different failure mode and a different forum where it has been contested.
A score's provenance is not the same artifact as a record of what was checked, which is the distinction drawn in why a justification record is not an investigation record.
What the vendors themselves say about how far their numbers go
Answer
Do estimating software vendors claim their prices are correct for a specific job?
No. The strongest argument for governing borrowed numbers comes from the vendors. The leading estimating vendor's own pricing white paper says its role is to report a market price, that no published price can be certain to fit a specific contractor, repair or structure, and that the estimator then decides whether the price is accepted or adjusted.
The document to read is not a critic's. It is the leading estimating vendor's own pricing methodology white paper, which describes its number as "a basis from which the estimator can then decide whether the price should be accepted or adjusted".
That clause names the decision-maker, and it is not the software. The vendor calls its output an input requiring a judgment. That judgment is the insurer's work product, and in most claim files it leaves no trace.
The operating model that makes it visible is in which claims decisions stay human, and what they get handed: the machine assembles and cites, the licensed human decides, and the reason is recorded.
The rule that bars sole reliance, and the disclosure duty with it
Answer
Can an insurer rely solely on estimating software to settle a claim?
Not in Washington. A reasonable investigation may not rely solely on a database, estimating software or benchmarks. That rule has been in force since 18 October 2026, paired with a disclosure duty: on a claimant's request, say when the pricing data was collected, where it came from and which businesses supplied it.
WAC 284-30-330(4) says: "A reasonable investigation may not rely solely on the use of a database, which includes, but is not limited to, estimating software and benchmarks gathered from one or multiple databases." Software output is an input to a reasonable investigation. It is not the investigation, and the file has to show the difference.
The companion provision is a records test. WAC 284-30-380(8) says an insurer using "a database, survey, estimating software, or benchmarks" to establish material pricing or labor rates must, on a claimant's request, provide "the date the data was collected, where the data was collected from, and which businesses provided the cost data". Nobody can answer that from a settlement letter.
Washington is one state, but nearly all states have a reasonable-investigation standard. Version-and-date fields cost nothing to capture when the number enters the file.
What the NAIC bulletin actually asks of a vendor relationship
Answer
Does the NAIC AI model bulletin cover vendor models or only models we build?
Both. The bulletin treats a third-party model as the insurer's model. Guideline 1.8 covers AI systems whether developed in-house or by a vendor, 4.1 asks for diligence so vendor-supported decisions meet the legal standards imposed on the insurer itself, and 4.2 asks for audit rights and regulatory cooperation written into the contract.
The bulletin, adopted on 4 December 2023, has no in-house exemption and no vendor exemption. Guideline 1.8 says the program "should address the AI Systems used with respect to regulated insurance practices whether developed by the Insurer or a third-party vendor". The more of your decision you borrowed, the more vendor model governance the bulletin expects to see.
Diligence happens before you sign. Governance happens every year after.
Most installed data and model vendors were bought before either review existed, and the renewal is automatic.
Guideline group 4.0 is the part to read before a renewal, not before a signature. Guideline 4.3 asks for the audit rights to be performed, so holding the right and never exercising it is its own finding. The examination list, item 1.2, says what a department may request: "Information and documentation relating to the Insurer's pre-acquisition/pre-use diligence, monitoring, oversight, and auditing of data or AI Systems developed by a third party."
Scale decides whether this is theoretical. The NAIC's legal adoption map, status 13 November 2026, lists 30 jurisdictions that have adopted the bulletin, plus California, Colorado and New York running their own insurance-specific AI regulation or guidance. A separate NAIC instrument asks a carrier to name its models, covered in why your fraud model now needs a page number.
When the carrier, not the vendor, defends the vendor's method
Answer
Who defends a vendor's valuation method when a policyholder sues?
The insurer. A federal appeals court reviewed a reduction one carrier applied to a vendor's list prices and reversed class certification without holding the reduction unlawful. In Georgia, an appraisal clause did not dispose of allegations about valuation methodology. In every caption the defendant was the carrier, not the vendor.
On 8 July 2025 the Third Circuit reversed class certification in two cases challenging an insurer's use of a Projected Sold Adjustment inside Mitchell International's WorkCenter Total Loss software. The reasoning was about proof, not method. The court did not hold the adjustment unlawful.
The vendor supplied the list price. The insurer defended the deduction.
Across two decades of valuation litigation the caption named the carrier, not the company whose method was in dispute.
- Vendor methodMitchellWorkCenter Total Loss list prices for the comparable vehicle
- Insurer's adjustment$488 to $818Projected Sold Adjustment per vehicle in one 2026 court record
- Named in the captionThe insurerClass certification reversed in 2025 and denied again in 2026
The pattern held in 2026. An Illinois federal court denied class certification on 15 April 2026 in litigation over the same adjustment. No court has held the method unlawful. No court has validated it either. In every caption the defendant was the insurer, not the vendor.
A borrowed number has to be authenticable as well as explainable, which is the subject of why the manual claim file fails the same test.
Four artifacts a claims organisation should be able to produce
Answer
What documents prove a claims organisation governs its third-party data?
Four, and you either have them or you do not: a vendor question set, a claim-file record schema, a contract clause set, and an incident playbook for the week your vendor becomes the story. None of the four requires the vendor's cooperation or a budget line to start writing.
(a) What to ask a data or model vendor before you rely on it
- What exactly does this number represent, and what does it explicitly exclude?
- How is the underlying data collected, from whom, and how often is it refreshed?
- Which version or price-list date does each output carry, and can our adjusters see it?
- What fourth-party data is licensed inside your output, and from whom?
- How are you compensated, and does any part of that depend on the size of the adjustment produced?
- Can you reproduce the exact value you returned for one claim after a model or price-list change?
- If a regulator or a court asks who decided this number was appropriate for this claim, what is your answer?
The last question is the test. A vendor that answers "the adjuster did" is describing its product accurately, exactly as that white paper does. Any other answer tells you what the vendor believes it sold you.
(b) What the claim file must record about every third-party input
A claim file that records a vendor's number has recorded the vendor's work. The insurer is responsible for the evaluation, so the file has to record the insurer's work as well. Seven fields. The sixth is the one that is almost always missing.
Fields one to five are usually recoverable from a system log. Fields six and seven exist only if somebody wrote them. More in what the claim-file rules asked for before AI and on our claim file review page.
The vendor's own white paper says the estimator decides. Most claim files record the number and no decision at all: five fields arrive off the log, and the two that carry the insurer's evaluation are written by a person or not at all.
(c) What belongs in the contract: disclosure, audit rights, indemnity, exit
- Disclosure on request: when the data behind an output was collected, where from, and which businesses provided it. Mirror WAC 284-30-380(8).
- Audit rights, or audit reports from qualified auditing entities, plus a commitment to exercise them.
- Regulatory cooperation, with a named contact and a response window.
- Notice before a method, price-list basis or model version changes in a way that moves outputs.
- Compensation transparency: how the vendor is paid for this product, stated in the contract.
- Indemnity scoped to the real exposure, which is regulatory and claimant-facing, not only a data breach.
- An exit clause that keeps you able to reproduce the vendor's values for the length of your retention obligation.
The exit clause is the one procurement usually trades away, because it costs nothing at signature and everything at renewal. That argument is in build vs buy claims AI.
(d) What to do when the vendor itself becomes the news
In 2026 this stopped being hypothetical for anyone relying on a property risk score. Nebraska's attorney general filed suit against First Street, a climate-risk data company and not a competitor of ours, on 27 August 2026. Florida's attorney general filed the state's complaint on 23 October 2026. Everything in both is an unproven allegation.
Your position does not depend on how either case ends. What a regulator will ask about is your files. So the playbook is an inventory exercise before it is a legal one.
- Inventory which open and closed decisions used the input, and which version and data date each one used.
- Freeze the version and the dated data you relied on, so the file shows what the adjuster actually saw.
- Pull the files where the input carried the most weight and check each for independent verification and a reason.
- Brief the regulator-facing owner, with the inventory attached, before a department or a claimant asks.
Re-validating a borrowed number belongs on a calendar, not in a crisis, which is the argument in claims AI accuracy is a measurement with an expiry date.
Where the evidence has to live
Answer
What should a claim file contain when a decision uses a third-party number?
The borrowed number is not the problem. The missing paragraph beside it is. A file that records what the input was, where it came from, what else was checked and why the decision went the way it did survives the week the vendor becomes the news, whatever the vendor turns out to have done.
Hesper consumes third-party data too, and this governance argument applies to us exactly as it applies to an estimating database. We are not arguing that a claims organisation should borrow fewer numbers. When the other side is automating as well, the file still has to explain itself, which is the argument in subrogation negotiation when both sides automate.
The difference is where the record of the borrowing lives. Every agent action is logged with its source, its reasoning and its timestamp, so "what did this third-party number contribute to this decision, and what else was checked" is a query rather than a reconstruction project.
The reason field is empty today because of capacity. On Hesper's internal benchmarks a manual investigation runs 14+ days per case at 200+ cases per investigator, and roughly 25% of flagged claims get a real workup. On the same benchmarks our agents take coverage of flagged claims to 100%, in minutes, not weeks. The paragraph gets written because writing it is the output.
The work is not to stop buying numbers. It is to write down what you did with them, at the moment you did it, in a field that exists.
Key takeaways
- Washington's re-adopted rules, effective 18 October 2026, make insurers responsible for the accuracy of evaluations made on their behalf.
- The estimating vendor's own white paper says no published price can be certain to fit a specific contractor, repair or structure. The estimator decides whether to accept or adjust it.
- The NAIC model bulletin treats a vendor's model as the insurer's model and asks for audit rights in the contract, across 30 adopting jurisdictions plus three with their own rules.
- In litigation the insurer defends the vendor's method. The Third Circuit reversed class certification in 2025 without holding the Projected Sold Adjustment unlawful.
- The claim file needs seven fields for every borrowed number. The two that go missing are the reason and the independent check.