Image Triage Should Not Bury the Fraud Referral

The photograph is not our evidence. It is a pointer to the evidence. We lose that distinction the moment we introduce image triage.

The pitch is persuasive enough.

Read the photographs. Compare an estimate. Clear the ordinary file. Save human attention for exceptions.

On a routine hail claim, sensible.

A roof does not get more damaged because an adjuster waits beside a keyboard.

The model that makes a file look routine is most dangerous when routine handling destroys the evidence that would have made it unusual.

An image classifier ranks a file for us.

It can’t decide the file stopped being desk adjusting and became evidence preservation. That decision sits at the seam between claims and fraud investigation. Design the seam as a queue to minimise, and the system will optimise the referral away.

So here is the limit I would hold.

Photo and estimate automation may accelerate a scope.

It must not overwrite the inspection record.

It must not flatten a referral reason into a score. And it must not let a quick payment path extinguish the materials an SIU would need.

An ordinary property file already has several authorities

We start with the file automation hopes to simplify.

At first notice we take policy number, date and time of loss, cause code, habitability, mitigation, and whether a contractor is already on site.

We set a formula reserve.

Average severity for the peril and region, not an estimate of this loss. We confirm coverage before dispatch: policy in force, deductible, and whether the roof is replacement cost, actual cash value, or a schedule.

Not a generic intake bundle.

A contractor already on site means somebody proposed a scope before we saw the roof. A punctured roof with active leaking goes to same-day mitigation.

An intact envelope is a different path entirely.

A system treating both as a set of photographs threw away the first routing decision.

Catastrophe operations make our boundary clearer.

Staff adjusters saturate within two days of a real event. We assign field work to independent firms by task: inspect, scope, estimate, report.

The staff examiner keeps coverage and cheque-writing authority.

The independent adjuster neither approves nor denies, however complete the photographs look.

The capacity numbers explain why image triage appeals.

A daily-claims adjuster carries forty to eighty files.

A deployed catastrophe adjuster carries several hundred.

First contact is due a day or two after assignment. Physical inspection in a large hail event may be two to six weeks away. Provincial regulators issue temporary certificates in declared catastrophes, so adjusters from elsewhere can work the loss.

That is surge capacity we borrow.

It is not delegated claims authority.

Public adjusters matter for a different reason.

The policyholder pays them, usually a share of the recovery.

Their presence is not an indicator we should act on.

But treat a contractor estimate, a public-adjuster submission, and our own scope as interchangeable uploads, and no later reader knows who produced what, or why.

The examiner’s authority returns with the field report.

The formula reserve becomes a case reserve.

Payments past the examiner’s authority go to a supervisor, a large-loss unit, or reinsurance. A contractor estimate at twice ours produces a supplement, a desk comparison, often a re-inspection. The first payment under replacement-cost coverage is actual cash value: replacement cost less depreciation and deductible.

Nothing there makes a suspicious file proven.

Each creates a version of the file we have to preserve rather than quietly consolidate.

The inspection is not merely training data

On a hail file the adjuster examines each roof slope, because hail is directional. The inspection holds a ten-by-ten test square, a count of fractured mats and bruises, chalk on the hits, and photographs of both the marked square and the identifiable slope.

Repair-versus-replacement thresholds vary by insurer and region.

A visual conclusion cannot replace the record.

Soft-metal damage corroborates. It is not scenery.

Dents in gutter aprons, vents, fascia, screens, condenser fins: they show stones capable of bruising asphalt fell, and from where. The absence of soft-metal damage is the strongest argument against hail on an aged roof. Granule loss and thermal cracking fool an inexperienced observer. The scope becomes a line-item estimate against monthly regional prices. The estimate can hold matching of dissimilar materials, steep-roof charges, by-law work, debris removal, detach-and-reset.

Interior damage carries its own moisture readings.

Scope, photographs, readings, images, estimate versions: a sequence.

Not a bundle to reduce to one total.

The design requirement is concrete.

Preserve the original image, the inspection context, the machine output and version, the human changes, the scope, and the referral indicator, in one sequence.

The corpus prescribes no particular storage architecture.

It makes the consequence plain enough.

The adjuster’s diary entries, photographs, scope, estimate versions, and correspondence are written for somebody who wasn’t there.

An auditor. A market-conduct examiner. Opposing counsel.

Referral is a change of authority, not a bad score

Our fraud flow starts after an ordinary claim opens. Detection arrives through adjuster judgment against a red-flag list, a score at notice, or link analysis across shared attributes. A phone number. A tow operator. A treating clinic. A body shop. A settlement bank account. A witness who keeps reappearing. Image triage is one more source of an indicator.

It is not a conclusion our SIU has reached.

The referral is a form, not a phone call.

“Claimant seemed evasive” gives an investigator nothing.

“Policy bound eleven days before the loss; three contents items lack purchase proof; insured declined to name the retailer” is a scope somebody can check.

Our SIU accepts or declines.

A declined referral returns with a recorded reason, because decline patterns show whether the indicators drifted.

Acceptance changes handling, not the duty to handle.

The adjuster issues a reservation of rights to preserve coverage defences, and keeps adjusting the claim. An SIU file that freezes ordinary handling creates delay and bad-faith exposure.

Which is a useful test for automation.

The system preserves a referral without turning its status into an instruction to stop. The SIU rebuilds the file before any field work. Application and underwriting notes. Payment and endorsement history. Prior claims. Loss date against inception. Last premium payment. Any recent limit increase or rider. CANATICS exposes claims connections across member insurers that one carrier cannot see. Registry, public-body, and open-source searches narrow the next step.

What the image process saw is useful only while its origin survives inside that record.

Privacy is an operating constraint, not legal trim.

Federal privacy law permits collection without consent where consent would compromise an investigation into a breach of agreement.

Quebec adds obligations, and two provinces have their own statutes.

The file records the purpose test before collection.

Open-source material is captured with visible URL and timestamp, preserved, logged with collector and time.

A screenshot in an email is poor material at examination.

Preserve the contradiction before the repair changes it

The field sequence exists to not spend the contradiction early.

A recorded statement usually comes before confrontation.

Loss narrative, whereabouts, ownership, purchase, prior damage, finances: taken while the account is still theirs.

Physical evidence follows, and we hold it.

Every exhibit gets a continuity record from collection.

On a property file that order is blunt.

Where destructive testing is contemplated, every interested party is invited to a joint examination first.

If we proceed alone, that is spoliation.

A repair workflow that discards the original images, substitutes a revised scope, or authorises irreversible work before anybody assesses the referral will get a brisk queue.

It will lose us the case.

The same discipline governs formal demands.

A sworn proof of loss is a statutory condition here.

We may demand documents and signed authorisations.

An examination under oath is contractual and pre-litigation, distinct from discovery.

It locks the claimant into dates, amounts, and names before we disclose contradictions.

An examination that opens with the contradiction gets an explanation instead of an admission.

The final position has an elements test.

The file maps false statement, materiality, knowledge, and timing to named exhibits.

Underwriting guidance in force on the application date says whether the true facts would have changed the decision, the premium, or the payment.

Application misrepresentation supports rescission from inception.

Post-loss fraud engages forfeiture of the claim.

F.H. v. McDougall applies the balance of probabilities while demanding evidence that is clear, convincing, and cogent.

An image score is not an exhibit we can file.

Not every image discrepancy deserves an escalation.

Fraud is a rare-event problem.

Confirmed fraudulent claims are low single digits of most books.

It means we need a referral path that keeps a material discrepancy, says why it mattered, and lets the SIU decline it.

A low review rate isn’t proof of a good threshold.

It may be proof the model learned our capacity limit rather than the risk.

Ask for the loss we did not create

The board opens with cycle time, handling cost, estimates cleared without a visit.

We should bring a second result.

The rate at which viable indicators were preserved, referred, and disposed of on evidence that supports the decision.

That needs separate instruments.

The chain is scored claims, alerts, referrals accepted, findings.

Conversion, acceptance, and substantiation reveal different things.

Substantiation rate alone is not comparable between insurers.

A unit taking only strong files looks excellent by declining work.

On an image model, ROC area flatters under class imbalance.

Precision at k, k being investigator capacity, and lift over random selection answer the useful question.

Did the ranking improve the queue we could staff?

An improvement below that cutoff improved nothing operationally.

The model also faces selective labels.

Its outcomes exist only for claims the SIU investigated, so retraining teaches it to repeat our past choices rather than find fraud.

Our savings number needs its own suspicion.

Demand basis uses the claimed amount. Reserve basis uses the case reserve at referral. Modelled basis uses an expected payout.

Reserve basis is common, and it moves with practice.

Raise reserves at referral and savings rise without a finding.

So whoever pays signs off the savings.

Recoveries stay separate from payments avoided.

A reversed denial reverses the saving that period.

Otherwise the graph rises because we forbade ourselves to learn.

The scorecard also carries the harm.

Report median and ninetieth-percentile days to SIU decision too.

FSRA and the AMF administer rules on handling delay.

Track litigated denials, and how many settle above the denied amount.

A random closed-file audit measures leakage outside the pipeline.

A control holdout is cleaner and harder to defend, because it exempts claims from scoring.

Finally, separate an adverse outcome from a fraud finding.

A claimant who walks away once an examination is scheduled hasn’t necessarily committed fraud.

A denial resolves the contract.

Criminal referral, regulatory referral, FINTRAC, civil recovery: different consequences.

Criminal referral runs to another bar: proof beyond a reasonable doubt, on an offence with its own value threshold.

The SIU should not be asked to make all of those decisions under the label “flag.”

Image triage is valuable when it lightens routine work.

It turns dangerous when routine erases the question of whether the file was routine.