A recommendation to accept, counter or escalate a subrogation demand is a claim-file decision, not a strategy call. When both carriers run software that prices the other side, the question becomes evidentiary: if a counter was priced off aggregate precedent, what does the file show a regulator, an arbitrator or a policyholder later?
The upstream half of the problem is in our piece on why missed subrogation is an evidence-timing problem. This one covers what governs the negotiation itself.
The negotiation most carriers are automating ends in a closed room
Answer
What ends a subrogation negotiation between two insurance carriers?
Agreement, or a filing. When two carriers cannot agree, most US intercompany disputes go to Arbitration Forums, where Rule 4-1 bars default judgments and requires decisions based on the evidence submitted, and Rule 3-5 limits an arbitrator to the pleadings, deferments and evidence actually attached to that filing.
The endgame is narrow. Under the Arbitration Forums rules effective October 21, 2024 and revised January 1, 2025, Rule 2-12 allows an appeal in limited circumstances and states that no additional documentation or evidence is allowed. The first submission is the whole case.
Arbitration Forums' own published guidance for members says arbitrators decide "solely on the evidence that is submitted". The counterparty's history is a reason to expect an outcome. It is not a reason the panel can accept.
One more rule matters later. Rule 2-13 makes all matters related to arbitration proceedings confidential, including filings, decisions and awards. The richest corpus of counterparty-behaviour data in the industry sits behind that rule.
Insurance has 1.1 million records of how carriers actually negotiate. Rule 2-13 seals every one.
What Shift is proposing, and the part of it that is already right
Answer
What is adverse carrier intelligence in subrogation?
It is two lenses on a demand: a profile of how the adverse carrier has behaved in past disputes, and a set of comparable claims showing how similar files resolved. Shift Technology's November 2026 paper builds both from the carrier's own historical claims and its arbitration decisions, and leaves the decision with the handler.
The clearest statement of the idea is Shift Technology's November 2026 paper "Stop negotiating in the blind". It is careful about its own claims. Treat what follows as a close reading rather than a rebuttal.
The paper draws the key distinction itself. The comparable claims set is evidence about the claim's likely path, "not about the counterparty". The carrier profile is evidence about the counterparty, "not about the claim". What nobody specifies is what reconciling the two has to leave behind in the file.
It also keeps decision rights human: the agent "recommends whether to accept, counter or escalate. The handler decides".
The one fair criticism is about documentation, not the product. Every number in the paper's two tables is explicitly an illustration, yet figures from those tables appear on Shift's October 20, 2026 end-to-end auto recovery page without the illustration label. More on that in our read of Shift's ARISE autonomy framework.
The institutional clock is 30 days, not 14
Answer
How long does the other carrier have to respond to a subrogation demand in arbitration?
Under Arbitration Forums Rule 3-1 a responding company is notified 30 calendar days before its response is due. Rule 2-9 permits one extension, requested before the current due date and carrying a fee, and states that no additional extensions are allowed. Rule 2-10 allows a one-year deferment the adverse party may challenge.
The response clock does not vary by counterparty, so it does not need a counterparty model. Three rules, not a vendor benchmark, define the real cadence of a subrogation negotiation. Arbitration Forums also advises members to file at least 120 days before the statute of limitations when possible, which puts the operational deadline months inside the legal one.
The clock that governs the other carrier is 30 days, and it is in the rules
Four rules decide when a subrogation negotiation stops being a negotiation. None of them needs a model of the counterparty.
- Rule 3-130 daysNotice to the responding company before its response is due
- Rule 2-91 extensionRequested before the due date, with a fee. No additional extensions allowed
- Rule 2-101 yearA deferment any party may add, and the adverse party may challenge
- Rule 4-1No defaultsDecisions must be based on the evidence submitted to the panel
The clock is institutional. How the other carrier usually behaves is not admissible in the forum that ends the negotiation.
The same rulebook fixes the other structural facts a negotiation runs into, and all of them are knowable in advance.
Now put that next to the illustrative 14-day average response time in Shift's example tables. The institutional number applies to every one of the 5,100-plus member companies. A carrier-specific average may still be useful. Just ask whether it is an institutional fact, a measured customer metric or a product illustration.
Rule 2-13's confidentiality is also the reason a vendor cannot simply pool the industry's arbitration outcomes. That is exactly why Shift scoped its profiles to each carrier's own closed files.
A recommendation to counter is a claim-file decision
Answer
What regulation applies to an AI recommendation to accept or counter a subrogation offer?
Unfair claims settlement practices law. Under the NAIC Model Unfair Claims Settlement Practices Act, Model #900, the violation is conduct committed flagrantly or with such frequency as to indicate a general business practice. One mispriced counter is not a violation. A pattern is, and automation produces patterns by construction.
The trigger in the NAIC Unfair Claims Settlement Practices Act, Model #900 is the pattern. An adjuster who misjudges one counter has made a mistake. A model that applies the same profile-driven range across a book of claims is, by construction, a pattern-generating machine.
Subsection 4(E) is the only one that is measured rather than judged. It covers "compelling insureds or beneficiaries to institute suits to recover amounts due under its policies by offering substantially less than the amounts ultimately recovered in suits brought by them". An agent tuned to a typical counter-offer band is tuned to that variable. The band is a compliance parameter, not just a commercial one.
So every accept-or-counter recommendation has to point at the facts in its own file, which means the file has to carry those facts before the recommendation exists. The decision-rights half is in our piece on which claims decisions stay human and what they get handed.
What the file has to show six years later
Answer
What records does an insurer have to keep of a subrogation settlement decision?
California's 10 CCR 2695.3 requires claim files detailed enough that pertinent events and their dates can be reconstructed and the licensee's actions determined. New York's 11 NYCRR 243.2(b)(4) requires the file for six calendar years after the claim is resolved and closed, or until the examination report is filed, whichever is longer.
Both rules are written in the language of reconstruction. New York's 11 NYCRR 243.2(b)(4) adds that a claim file "shall show clearly the inception, handling and disposition of the claim". Inception, handling, disposition. Not conclusions.
That is the problem with recording a machine-generated recommendation as a settlement rationale. A line saying the model recommended a counter inside the normal range for this carrier records a conclusion, not an event. We have written separately about why a decision log is not the same thing as decision evidence.
When the profile and the file disagree on a single claim, three different readers will later ask the same question about it.
When the profile and the file disagree, the file is what gets read later
The inputs conflict on a single claim, and three different readers will later ask the same question about it.
- 1.A counter arrives on day 26 of the 30-day response window
- 2.The carrier profile says this counterparty's counter sits inside its usual band
- 3.The file says the police report assigns the other driver a left-turn violation
- 4.The reconciliation line records which fact broke the tie and what was attached
One of those readers is a consumer. The insured's deductible rides on the settlement, so the reasoning behind a counter has to be legible to a person who was not party to it.
Human-in-the-loop answers one regulator, not both
Answer
Is keeping a handler in the loop enough to govern an AI subrogation recommendation?
Not by itself. The NAIC model bulletin treats human involvement as one of five calibration factors, and an AI System expressly includes anything generating recommendations. Separately, in Duffy v. Yardi the United States argued that joint delegation to a common algorithm can be unlawful even where the users retain authority to deviate.
Take the insurance regulator first. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers covers an agent that recommends accept, counter or escalate whether or not a handler signs off. Human involvement is a mitigating factor, not a pass. As of the adoption map's November 13, 2026 status, 30 jurisdictions have adopted the bulletin.
On examination a department may request the model's "data source, provenance, data lineage, quality, integrity, bias analysis and minimization, suitability, and Data Currency". For an adverse-carrier profile, that means: where did the behavioural data come from, how old is it, and whose was it.
Two instruments cited in this debate do not reach a claims recommendation at all. The table shows which ones do.
Now the regulator nobody in claims AI is quoting. On March 1, 2024 the Department of Justice and the Federal Trade Commission filed a joint Statement of Interest in McKenna Duffy v. Yardi Systems, Inc. Its load-bearing sentence: "it is per se illegal for competing landlords to jointly delegate key aspects of their pricing to a common algorithm, even if the landlords retain some authority to deviate from the algorithm's recommendations".
That is the exact opposite of how human-in-the-loop is being sold as a catch-all assurance. But be precise about the limits. Duffy and the RealPage case are consumer price-fixing matters. Subrogation is a business-to-business recovery negotiation, and nothing in this piece suggests that any insurer or any vendor is price fixing. Nothing here is legal advice either.
The narrower point is structural. A cross-carrier behavioural-profiling network is close enough to a common algorithm fed by competitors' non-public data that it belongs in front of antitrust counsel before it reaches procurement. Shift's current single-carrier design, built on each carrier's own closed files, sits on the right side of that line.
A regulator asks where the data came from. A decree asks whose it was.
The left column is in force for insurers. The right is a remedy from another industry, offered as a design reference.
Human-in-the-loop answers the insurance regulator. It is not an answer to the competition regulator.
The RealPage remedy is worth reading as a design specification. The Competitive Impact Statement filed November 24, 2025 describes a three-year monitor term extendable by 18 months, and sets out four provisions that work as a procurement standard for anyone buying adverse-carrier intelligence:
- No non-public data from competing properties in runtime operation, subject to limited exceptions.
- Training limited to backward-looking data aged at least 12 months and not drawn from active leases.
- Significant limits on using, sharing or disclosing competitors' non-public data to a customer.
- Customers stay free to reject or override a recommended price, and must set the parameters of any auto-accept feature individually.
The reconciliation is the artefact
Answer
What should a claim file show when an AI-recommended subrogation range and the file's own evidence disagree?
Both inputs, the disagreement itself, and the fact in the file that resolved it, with the evidence that was attached and who decided. That reconciliation is not a decision log. It is the only record that answers an arbitration panel, a market conduct examiner and the insured at the same time.
A carrier profile is evidence about the counterparty and a comparable claims set is evidence about the archetype. Neither is evidence about this loss. The forum reads only this file, and the record has to outlive the negotiation by years. So when the profile-driven range and the file disagree, something has to be written down.
That written-down thing is the reconciliation. It is not a decision log. A reconciliation line on a subrogation file carries five things.
- The profile-driven or comparables-driven expectation, stated as a range with the population it came from and the age of that population.
- The file's own evidence on the same question: the police report, the photographs, the recorded statement, the repair invoice.
- The disagreement, named plainly, rather than averaged away into a single recommended number.
- The fact that broke the tie, cited to its document, and whether that document was attached to the filing.
- Who decided, when, and what they were handed at the moment they decided.
This is why Hesper AI treats recovery as a stage of the claim rather than a product bolted to the end of it. The same evidence file that carried the investigation carries the demand into recovery. Demands and arbitration filings are drafted from that evidence, in minutes, not weeks, and the recovery team makes the pursue, counter or waive call.
Hesper has built-in fraud detection and runs standalone or alongside FRISS and Shift Technology. The product detail sits on our subrogation and recovery platform page, and the line view sits on our auto claims use case.
None of this argues against automating the negotiation. The volume says automate it. The argument is about what the automation is for. A system that produces a recommendation moves the decision. A system that produces a file supports it, and six years later only one of those two things is still in the record.
Key takeaways
- A recommendation to accept, counter or escalate a subrogation demand is a claim-file decision, so the question when both carriers automate is evidentiary rather than tactical.
- The response clock is institutional, not behavioural: Rule 3-1 gives 30 calendar days' notice and Rule 2-9 permits one fee-bearing extension, for every member company.
- Under NAIC Model #900 the violation is a pattern, which is what a model applying one profile-driven range across a book of claims produces by construction.
- Human-in-the-loop answers the insurance regulator. It does not answer the competition regulator.
- When a profile-driven range and the file's evidence disagree, the reconciliation - both inputs, the disagreement and the fact that broke the tie - is the artefact.