Exclusion Automation Fails When It Loses the Write-Back

Clause extraction has an obvious appeal in claims and policy operations. Give the system a long packet and it finds an exclusion in seconds.

Flood. Faulty workmanship. Pollution. Cyber. Professional services.

Real progress for us, that.

The packet isn’t getting shorter, and neither is our queue. But an exclusion is rarely the end of the coverage analysis. It is usually the start of the part the model is least likely to keep.

The dangerous policy AI isn’t the one that misses an exclusion. It is the one that finds it, sounds certain, and quietly drops the wording that gives coverage back.

A fluent summary of a water exclusion that omits the plumbing escape write-back is not a partial answer.

It is a different answer.

Same when faulty workmanship removes the cost of repairing bad work and resulting damage survives.

The Supreme Court addressed exactly that in Ledcor.

The error is structural.

A clause extractor treats the policy as a collection of clauses.

A policy is an assembled contract.

Declarations, base form, definitions, exclusions, exceptions, endorsements, schedules, conditions, and, here, statutory text that can outrank our own wording.

Finding one sentence is retrieval.

Establishing what survived it is the analysis we owe.

The missing half of the answer

So here is the line I would hold.

Use AI to locate and compare wording.

Don’t use a clause summary as the claims decision layer.

The decision layer needs a versioned dependency map.

It has to show the path from the base wording to the exception, endorsement, schedule, and statute that alter it.

The exclusion does not live alone

The basic sequence is simple enough to be machine-readable.

The grant establishes coverage we might owe.

Definitions can change the meaning of its terms.

An exclusion removes a class of loss.

A carve-back restores a defined part of that class.

Conditions and limits then qualify what we pay.

The burden moves with the structure.

Grant to the insured. Exclusion to us. Carve-back back to the insured.

That burden sequence is an operational requirement.

A system extracting an exclusion should have to ask whether the same form holds an exception.

Whether another attached form modifies the coverage part.

Whether a schedule supplies a limit, location, person, or date that matters. If it can’t answer those, it handed us a lead, not a conclusion.

Consider construction.

The damage-to-your-work exclusion in a CGL leaves the insured’s own defective work as a business cost. But inside the products-completed operations hazard, the subcontractor exception can return coverage where the damaged work, or the work that caused the damage, was done by a subcontractor. A general contractor who subcontracts roofing and one who self-performs it can land in different coverage positions on superficially similar facts.

Not a niche legal flourish.

It is one of the most consequential lines we write in Canadian construction. An automated conclusion saying “your work excluded,” without naming the products-completed operations context and the subcontractor exception, has turned a coverage question into a slogan.

The same pattern shows up in property.

Flood and surface water have been excluded from Canadian personal property forms for decades. Overland water endorsements only entered the market around 2015, after the 2013 Alberta floods and the Toronto storm.

Sewer backup is a separate, older endorsement.

A system merging the two because both are “water damage” will send the file confidently to the wrong form.

Attachment is not an implementation detail

An endorsement is not a footnote, and it is not a universal override. It amends a particular form, coverage part, person, location, or project.

Its preamble and schedule set the scope we are working inside.

It overrides the base form only where there is an actual conflict.

The rest of our base wording survives.

Which is where automation gets careless.

An additional-insured endorsement can add a party to the definition of insured while a separate subcontractor-work exclusion removes the exposure that party expected covered. A certificate on an IBC or ACORD form may report the addition.

It does not amend the policy.

What we need is the attached endorsement.

Not a certificate image suggesting one probably exists.

Ordering is conditional too, not a decorative hierarchy.

We harmonise the contract first.

Where two provisions genuinely conflict, statute governs over contract.

An endorsement governs over the form it amends.

A later conflicting endorsement governs over an earlier one. Specifically negotiated manuscript wording generally governs over printed boilerplate. The declarations and schedules supply the variables those forms operate on. So the first technical object is a dependency record, not an embedding. For every exclusion and endorsement: affected coverage part, base-form number and edition, policy-term attachment, effective date, schedule fields, preamble scope, conflict relationship, and whether it broadens, restricts, or leaves coverage neutral.

We store the source page with it.

A vector similarity score cannot tell an operator whether the endorsement hit the liability section while the claim concerns property.

The identity key matters here.

CG 00 01 12 07 and CG 00 01 04 13 share a number and a broad purpose, and they are different texts. The policy effective date decides which attached edition governs a loss.

An approval or permitted-use date is a different clock entirely.

A system substituting the current edition for a retired one has not refreshed its knowledge.

It has changed our contract.

The write-back has a lineage of its own

The word “exception” makes the problem sound modest.

It is not modest.

The exception is usually where we preserved the fortuitous part of a loss after excluding the certainty, accumulation, or business cost around it.

Faulty workmanship is the cleanest example.

The cost of making good defective work is excluded, and resulting damage can survive.

Fire following earthquake can be written back even though earthquake is excluded.

An earthquake buy-back is a different mechanism again.

An endorsement, commonly with a deductible expressed as a percentage of insured value, and often around a tenth in high-hazard coastal British Columbia.

Exception, buy-back, and deductible are three separate dependencies.

A summary calling all three an “earthquake exclusion” is not usefully compressed.

The market makes the distinction visible elsewhere.

A total pollution exclusion points toward environmental impairment liability, site pollution, or contractors’ pollution liability.

Nuclear energy hazard is generally written through the Nuclear Insurance Association of Canada under the federal statute.

Terrorism has been standardly excluded from commercial property since 2001, and Canada has no federal backstop.

Those aren’t empty gaps.

They are boundaries between our products, pools, and reinsurance arrangements.

Some exclusions arrive through a different chain entirely.

Communicable-disease exclusions propagated across Canadian commercial programmes from 2020.

Lloyd’s required state-backed cyber-war wording on standalone cyber policies from March 2023.

The individual file may show only an attached endorsement.

The actual explanation may start with treaty appetite or a market requirement.

Record that provenance where we have it.

But decide the claim from the wording that attached.

Broadening and restrictive labels therefore need dates.

A new edition can keep the same base form number at renewal and change the result.

The revision circular explains the change. The declarations schedule proves what attached. The footer proves the edition printed.

Three sources, three different jobs.

A library storing PDFs without their relationships is a fast route to false confidence.

Canada is not a localisation field

We cannot build Canadian policy control by pointing a US clause corpus at Canadian PDFs.

Ontario automobile insurance uses the OAP 1 and numbered OPCF forms approved by FSRA.

OPCF 44R provides family protection. OPCF 27 addresses damage to rented or borrowed vehicles. OPCF 43 removes depreciation in its defined new-vehicle window.

Alberta, Nova Scotia, and other common-law provinces use SPF and SEF families.

Quebec uses QPF and QEF forms.

The form population changes again outside automobile.

IBC 2100 is a Canadian reference point for commercial general liability.

But commercial property and liability wordings are generally not publicly filed here.

A proprietary form or a manuscript exclusion may have no public filing trail at all.

The operative library gets built from carrier and broker records.

Not assumed from an American catalogue.

Provincial statutory conditions are another dependency, not a note for legal.

In common-law provinces they are deemed part of relevant policies, and wording reducing them to the insured’s prejudice is ineffective to that extent.

Quebec contracts run under the Civil Code of Québec instead.

French and English texts under one form number need separate control records.

In Quebec the French text ordinarily governs.

None of it is academic when somebody asks a model for the “standard exclusion.”

A standard exclusion belongs to a particular edition, jurisdiction, coverage part, language record, and policy term.

Without those fields we have a plausible generalisation.

We do not have the operative contract.

Build the chain before anybody builds the interface

The control we need is mechanical, which is its virtue.

Start from the exclusion and walk outward, one dependency at a time.

Which definition changes its terms. Which exception sits inside the same form. Which endorsement amends the coverage part. Which schedule supplies the limit, the location, or the date.

Then flag the ones that break: an endorsement whose write-back has no schedule entry, a carve-back conditioned on a form nobody attached, a provincial amendatory endorsement paired with a base edition it was never written against.

For manuscript wording, a file name isn’t enough.

Capture the drafter, the negotiated or library origin, the base form and edition, the redline, the effective dates, the schedule dependencies, and the stated precedence.

Manuscript wording often gives us no bureau footer, no filing trail, no edition discipline.

The later reviewer needs to know what text attached, and what text it was designed to disturb.

AI accelerates every unglamorous part of this.

It extracts form identifiers, detects apparent missing pages, compares editions, identifies candidate write-backs, surfaces a schedule conflict for review.

It makes our policy archaeology less slow.

It should not decide that an exception was immaterial, that an unattached document was operative, or that an American analogue is the Canadian form.

The board hears that our AI programme finds exclusions faster.

The more useful measure is traceability.

For every exclusion conclusion, can we show the attached endorsement and the write-back that could alter it?

Can we reproduce the chain from the issued packet?

The system has understood the exclusion only when it can also explain what survived it.