PolicyROS™Planned product · runs on the Reasonex™ engineDeciding what a policy covers
Today’s insurance AI reads the documents. Someone still has to work out what they mean when read together.
PolicyROS is hallucination-free on the decided path, and deductive. It derives the answer by applying the governing provisions together, and no generative model writes any step of it.
A different class of question
Insurers already use AI to extract, summarise, predict and route. Those systems handle the tasks that surround a decision. The determination itself needs several rules, facts, exceptions, priorities, dates and amounts considered together. PolicyROS is built for this class of problem.
- Is this claim covered?
- Which exclusion applies?
- Does an endorsement restore what an exclusion removed?
- Has a condition been satisfied?
- Which sublimit applies?
- Is the underwriter permitted to bind this risk?
- Can a decision be made at all on the facts available?
A policy stack is a set of provisions that grant, remove, restore, cap and gate cover. The outcome depends on the order and priority in which they apply. PolicyROS compiles the stack into a governed rule set, then resolves how those provisions interact against the established facts. Illustrative wording and figures.
Retrieval ranks passages by resemblance. Resemblance is not entailment.
It does not hallucinate on the decided path
No generative model produces a reasoning step or the final determination. The same facts against the same rules return the same answer on every run.
It reasons deductively
It applies several provisions together and derives what follows, including where they compete for priority. The clause that decides the outcome is often not the clause that most resembles the question.
Stated precisely, so it can be tested: the claim covers the governed decision path. We do not claim that OCR cannot misread a number, that a language model cannot misread free text, or that a person cannot enter a wrong fact.
Three layers. PolicyROS is the third.
The first two layers answer what the information is and what it suggests. The third answers what the rules require given that information.
| Layer | Primary function | Typical systems |
|---|---|---|
| Language intelligence | Understand insurance terminology, extract facts, classify documents, summarise claim files | General and insurance-specialised language models, OCR and extraction pipelines |
| Predictive and causal intelligence | Estimate risk and identify drivers: loss prediction, pricing, fraud scoring, lapse and retention | Machine learning models, actuarial and causal analytics |
| Deterministic reasoning | Determine what follows from governed rules and established facts, and what is missing | PolicyROS |
One water-damage claim, step by step
The five-layer example above, run through the engine. Each entry is one step in the determination trace.
- Escape of water from a burst internal inlet pipe, 3 March; discovered the same day
- Notice given to the insurer 12 March
- Remediation quotation S$310,000
- Stack compiled: base form, two endorsements, schedule
Both branches computed. On either reading of cl. 7.4 the sublimit caps recovery at S$250,000, so the flagged question changes the route to the answer rather than the amount. Where a flagged reading would change the amount, the engine states the dependency and stops there.
Direct physical loss. Engaged on the established facts.
Flood or surface water. Whether an escape from an internal pipe falls within “surface water” turns on how the clause is read. Named as a question of construction and returned to a person, with the question stated.
Sudden and accidental discharge of water restored. If cl. 7.4 reaches the facts, the endorsement prevails to the extent of the restoration; the priority and its basis are recorded.
Notice within 14 days of discovery. 3 March to 12 March computed as 9 days against the calendar; condition satisfied; the counting is shown.
Water-damage sublimit S$250,000 against a quantum of S$310,000. Recoverable amount capped at S$250,000; the arithmetic is shown.
Illustrative scenario: simplified wording, figures and dates that do not reproduce any insurer’s policy.
What you get back
Your documents
Base form, endorsements, riders and schedule compiled into a governed rule set. Anything that cannot be read is reported before any question is asked.
The facts
An unestablished fact is never treated as a false one. Where a required fact is absent, the engine asks for it by name.
One answer
The same inputs produce the same result on every run, and the result does not vary with phrasing.
The numbers
Dates, limits, sublimits and aggregates are computed from the clause, the figures and the calendar, with the working recorded.
What it will not decide
Questions of construction, matters of discretion and missing facts are named and routed to a person. No gap is filled with an assumption.
The record
Every step carries the clause that authorises it and a re-checkable record, so a determination can be verified before reliance and reproduced afterwards.
In claims, and before you pay one
The applications below describe how the same reasoning applies to insurance decisions. They are designs for discussion, not shipped features.
Is this claim covered
Insuring agreements, exclusions, endorsements, conditions, sublimits and aggregates applied together, not clause by clause.
Which policy year pays
Claims-made questions of dates and related claims resolved to the period that responds.
How much is payable
Deductibles, sublimits, aggregates and limits computed from the clause and the figures, working shown.
Checking an earlier decision
A determination reconstructed independently: the facts relied on, the clauses applied, what was missed.
Answering a complaint
The explanation produced from the decision trace, so it reflects the actual basis of the decision.
What a fraud score lets you do
A model produces a score; PolicyROS determines what the insurer is permitted to do with it.
Checking a risk before you bind
Appetite, thresholds, referral rules, mandatory endorsements and delegated authority applied to a submission.
Who is allowed to bind it
Whether a given underwriter may bind, and if not, which threshold was exceeded and whose approval is needed.
What the final wording really covers
The compiled stack shows what the combination grants, removes, restores and caps, and which endorsement takes priority.
Did the policy match the quote
Not whether the wording changed, but whether the coverage effect changed.
Will the treaty pay
Attachment, retention, layer limits, aggregation and reinstatements worked through to a recoverable amount.
What a wording change affects
Which products, endorsements, claim outcomes and underwriting rules depend on an amended clause.
What each can promise
| Capability | Retrieval / generative | PolicyROS |
|---|---|---|
| Read, search and summarise policy wording | Yes | Yes |
| Apply several dependent rules exactly | Not guaranteed | Yes |
| Resolve governed endorsement priority | Not guaranteed | Yes |
| Identify a missing fact blocking the result | Not guaranteed | Yes |
| Compute dates, deductibles and limits deterministically | Not guaranteed | Yes |
| Produce the same result from the same inputs | Not guaranteed | Yes |
| Abstain where a governed decision cannot be made | Model-dependent | Governed |
| Hallucinate within the governed decision path | Possible | No |
Prediction stays with your models. Interpretation stays with people.
It does not replace catastrophe, pricing, fraud, image-damage, mortality or propensity models, or document extraction. Those solve probabilistic prediction problems; PolicyROS determines what follows from governed rules and established facts. The two are used together.
Some questions require interpretation and should not be decided mechanically: whether wording is ambiguous, whether damage falls within a contested definition, whether conduct was reasonable, whether evidence is credible. PolicyROS returns those with the issue stated, so human attention goes to the genuinely interpretive part of the file.
How a pilot would work
Every claim on this page is stated as behaviour you can test on your own documents, not as an architecture you have to take on trust.
Rule preparation
Select one product line and compile its base policy, schedule, common endorsements and claims-handling rules. Compilation reports what could not be parsed before any question is asked.
Historical claims
Run closed claims through the engine and compare the insurer’s actual decision with the PolicyROS result, examining each difference and its cause.
Operational test
Run alongside adjusters without affecting the live decision, measuring agreement, processing time, missing-fact identification, escalation rate and consistency.
Controlled production use
Use for defined decision classes with human oversight.
- One answer per question. Put the same determination through repeatedly and vary the phrasing. The answer does not change.
- Citation on every step. Each step names the provision it rests on. Open the clause and check it says what the step says it says.
- Computation, not retrieval. Alter a date or a figure by one unit and confirm the computed deadline or amount moves correspondingly.
- Abstention. Remove an established fact from the file and confirm the engine stops and names it.
- Scope. Ask a question outside the compiled corpus and confirm it declines rather than answering.
We have no insurance pilot data yet and do not offer a number in place of one.
