An Underwriting Engine Cannot Replace the Class Plan

An underwriting engine can make a submission move at impressive speed. It reads an application, checks a location, calls a data source, prices it, and issues a document before the broker finishes their coffee. That is useful. It is not underwriting judgment.

The engine is not a smarter underwriter. It is a very fast way to repeat whatever classification argument we actually encoded.

If that argument is filed, credible, permitted, and legible, speed is an advantage. If it is a vendor score with no route back to those things, speed distributes the problem more efficiently.

Clearance and triage are not permission to bind

The distinction gets clearer if we start where the process starts. A commercial submission arrives as an email bundle or a portal package. Application, CSIO or ACORD forms, loss runs, statement of values, whatever else the broker sent. Before pricing, we clear the insured against the book. Existing coverage. Duplicate broker submissions. The Consolidated Canadian Autonomous Sanctions List. The Criminal Code listed-entities register. That clearance is not a decorative pre-check. It stops two of our underwriters quoting one risk through two brokers, which makes us our own competitor. It also puts a broker-of-record change on its own path, with a letter and about ten days’ notice. None of it gets smarter by happening in seconds. It becomes easier to audit, provided the engine keeps the event and its result. Triage comes next.

It sorts a cleared file: write, refer, decline.

A human does it in under twenty minutes.

Commercial triage may use NAICS codes, occupancy and construction class, Fire Underwriters Survey protection data, hydrant distance, licensing, and insured value against treaty capacity.

That is a routing decision.

It is not permission to bind.

I put the line right there.

Clearance and triage are excellent candidates for automation.

Binding stays tied to a written authority, a class plan, and a record of how the file crossed each boundary.

The quote is a stack of arguments, not a score

The class plan is the structure of variables, levels, and relativities turning exposure into price. In Ontario personal auto we file both rates and the risk classification system with FSRA. A variable outside that filing isn’t a clever feature waiting to be switched on. The filed algorithm may combine years licensed, at-fault claims, convictions, territory, vehicle rate group, kilometres, commute distance, and driver assignment.

CLEAR supplies rate groups by make, model, and year, across three coverages.

A long way from a generic probability score.

Generalised linear models estimate frequency and severity, other variables held constant.

The operational output is a set of relativities.

The model supports the argument.

It does not replace the argument.

Commercial pricing makes the same point more visibly.

Property rates against total insured value. Liability rates against receipts or payroll. Fleet rates against power units, radius, and gross vehicle weight.

COPE resolves into construction, occupancy, protection, exposure.

The Fire Underwriters Survey grades commercial and residential protection separately.

A property eight kilometres from a hall with a hydrant nearby is a different rating conversation from one thirteen out with no recognised water supply.

Then the levers an engine folds into one opaque result.

Technical rate is the rate before discretionary adjustment.

Schedule rating is a documented credit or debit for something observed: housekeeping, management quality, loss-control cooperation.

The customary band is a quarter either way, wider on larger accounts under an approved schedule.

One characteristic cannot justify both adjustments.

Deductibles, limits, and subjectivities are separate again.

A small commercial deductible runs in the low thousands, a mid-market one in the tens.

CGL may be written at $1M, $2M, or $5M per occurrence, with an umbrella above it.

A quote can hang on an inspection, a signed application, a no-known-losses declaration, an electrical upgrade.

Those are terms.

They are not unexplained confidence scores.

The engine should show each input, its source, the permitted range, and the rule that moved it.

Otherwise our underwriter can’t tell whether the system priced a risk, moved an authority, or mistook a missing document for a good answer.

Credibility is a constraint on how far automation can go

The most seductive claim in underwriting technology is that more data buys a more individual price.

The mathematics is less obliging.

Pooling works because the coefficient of variation of an average falls with the square root of n.

We classify because a heterogeneous pool priced at its mean invites the better risks to leave.

That lifts the residual mean and starts the ratchet.

Akerlof, then Rothschild and Stiglitz, made the mechanism plain.

A pooling contract is undercut by one aimed lower.

That licenses no infinite number of cells.

Split a cell and the variance of each estimate rises.

The classical full-credibility standard for frequency needs roughly a thousand claims.

A low-frequency line does not obtain that volume because an API made more calls.

Bühlmann credibility gives the practical answer.

Z equals n over n plus k.

Where within-cell noise is large against the gap between cells, that cell’s own experience carries little weight.

The right answer blends toward a broader class mean.

Not a triumphant new micro-segment.

Which is where a black-box engine turns dangerous.

It makes a thin pattern look precise.

It gives two nearly identical risks different prices.

And it leaves no operator able to see that the difference rests on almost no credible experience.

Fine segmentation is not automatically modern.

Sometimes it is an unfiled opinion with good UX.

A variable has to survive four tests, not one

A rating factor isn’t defensible because it predicts.

It has to be stable, on credible volume.

Objectively measurable and hard to manipulate.

Causally connected, or otherwise socially supportable.

And permitted, filed, approved where required.

Credit-based scores show why those tests can’t collapse into accuracy.

They can perform well statistically and operationally while failing the causal and controllability tests.

Ontario settles part of it: credit information is prohibited in auto.

We cannot cure that prohibition by calling the input a feature, or by burying it several layers behind a score.

Proxy testing makes the boundary harder.

Where a prohibited variable correlates with a permitted one, the permitted coefficient absorbs it, even though the prohibited variable never entered our model.

A variable that predicts for its own loss-related reasons, and merely correlates with a protected characteristic, needs a different response from one whose power comes from the correlation.

The first may be residualised or reweighted.

The second may have to go.

Telematics tests whether our team understands the difference.

Distance, time of day, harsh braking, cornering: behaviour, and possibly more causal than a blunt proxy.

But late-night driving tracks shift work, and shift work tracks income.

An opt-in programme selects on driver type before the model starts.

The engine needs a recurring proxy-control stage.

Not a one-time fairness declaration at procurement.

Ontario adds a useful discipline here.

Auto telematics is permitted as a discount only.

The collected data may reduce premium.

It may not surcharge or decline anybody.

That isn’t an awkward limitation to design around.

It is the product rule, and we preserve it.

Referral is where the authority actually lives

Underwriting authority is not seniority rendered in software.

It is a grid of ceilings.

Line size, permitted class, total insured value, limit, catastrophe aggregate by zone, rate deviation, policy term.

A junior underwriter gets about $5M of insured value and a $2M CGL limit.

A senior gets $25M and a $5M limit, and sends bigger risks up.

That same senior may hold only fifteen per cent of pricing discretion either way.

The engine can test those boundaries without hesitating.

It should.

Common triggers: insured value above the treaty line, a catastrophe aggregate above a zone cap, a single loss above $250,000, a class outside appetite, a requested removal of a named-peril exclusion, any need for facultative reinsurance.

A carrier with $1M net retention and a nine-line surplus treaty has $10M of automatic capacity.

Beyond that, facultative cover comes before the quote.

An engine makes those rules visible and consistent.

They are why referral has to stay broad.

Upward referral handles line size or appetite exceptions.

Sideways referral goes to actuarial for a deviation outside the band, catastrophe for an aggregate breach, legal for manuscript wording, reinsurance for capacity.

Routing all of them into one queue called “manual review” is a small administrative mistake with expensive consequences.

The record matters as much as the route.

Referral reason and sanction are the sampling frame for internal audit and for OSFI.

An underwriter who splits a risk into two policies below a line limit has defeated the control.

An engine making that split invisibly has industrialised it.

Keep the class plan outside the model where people can read it

The board hears the throughput number.

We need the better question.

Does straight-through issue preserve the approved rating logic, the referral authority, and an explanation of the action taken?

That question produces a testable file.

For every straight-through bind, keep the class-plan version, inputs and sources, relativities, technical rate, schedule reason, deductible, limit, subjectivities, authority result, referral result, and evidence that capacity was in place.

For each decline, keep the exact knock-out, and make sure the recorded reason matches the broker-facing reason.

A quote-to-decline doesn’t substitute for that.

Auto needs particular care.

Ontario’s take-all-comers rule makes us offer coverage to an applicant meeting our filed criteria.

The Facility Association is the market of last resort.

The Risk Sharing Pool is a different animal.

We issue and service the policy normally, and cede premium and losses to an industry pool, subject to a cap.

A system describing both as generic “risk offload” has not understood the decision it is automating.

So keep the class plan outside the model.

Keep it in a form four departments can read.

And keep referral wider than the sales interface wants.

A vendor model card doesn’t modernise a filed rating rule.

Automation is safe only while the reasoning that prices a risk stays more inspectable than the interface that sells it.