Don't Just Quote the Fleet.
Design the Program.

Part one

The question nobody owns

We wanted to price the question rather than argue about it. If a motor carrier's own loss experience were examined properly across a full decade, how far would the right structure sit from the one it had been renewing into — and would the distance be large enough that anyone should care?

The fleet we looked at runs roughly 180 power units out of the Midwest. Over ten years it covered 191 million miles and reported 347 auto liability claims. In every one of those years it renewed into a traditional first-dollar program. Nobody in that decade appears to have asked whether that was still the right container for the risk.

The distance turned out to be larger than the account's annual premium.

Three things in that table matter more than the headline. The first is that the conservative option never lost. A $25,000 deductible — no captive, nothing exotic, nothing that requires a board resolution — came out ahead in all ten years. This is not a captive argument. It is a structure argument, and it has an answer for the owner who will never entertain a captive under any circumstances.

The second is that the aggressive option is a genuine trade rather than a free lunch. Captive “B” gave up $98,840 in its worst year to gain $621,080 in its best. Presented honestly, that is a decision an owner can make on their own terms. Presented as a saving, it is a trap.

The third is the one we keep returning to. Nobody ever showed this carrier the table. Not because the data was missing — every number in it came from documents the fleet already owned.

Part two

Why the question waited thirty years, and why it cannot wait another one

It is worth being fair to everyone who never asked it. For most of the last three decades there was a sound commercial reason not to. Assembling the evidence cost more than the answer was worth on all but the largest account, so the analysis was rationed to the handful of clients a broker could not afford to lose. That was not negligence. It was arithmetic.

What has changed is that both sides of the arithmetic moved, in opposite directions, at the same time.

The cost of being wrong went up

Retention analysis exists to protect against the tail. The tail has been getting heavier for fifteen years — not only more often, but far more expensively when it lands.

A broker reading that as a safety story is reading it too narrowly. It is a structure story. When the tail gets heavier, the difference between a retention chosen on evidence and a retention chosen by inheritance stops being academic. It becomes the most consequential number on the account.

And the reason for not knowing disappeared

At the same time, the three weeks collapsed into an afternoon. The mechanics of that are worth a section of their own, and they get one further down — but the consequence can be stated now. The analysis that was rationed to a handful of accounts can be run on all of them.

The excuse expired at exactly the moment the penalty for needing it went up.

There is a third change, and it is the one that will decide who benefits

When an analysis is expensive, performing it is a differentiator. When it becomes cheap, not performing it becomes a liability — because the broker sitting opposite you at the next renewal will have performed it.

Verdict severity is, in the end, the fleet's problem. A competitor who now walks into every renewal with a ten-year structure comparison is the broker's problem. That window is open, and it is the nature of such windows to close.

Part three

The market does not rate your experience. It rates your exposure.

This is where most brokers lose the argument before they have finished making it. They arrive with a loss history and expect it to price the risk. Steve McElhiney has spent thirty-five years on the other side of that conversation.

The distinction is not academic and it decides pricing. Experience is what already happened to a fleet. Exposure is what the operation is likely to produce next year. Carriers and reinsurers rate the second one, and they say so plainly when asked.

Which is why the second body of evidence matters, and why it is the half almost no broker brings to the table. It sits in the fleet management system the client already pays for: following distance, harsh braking, speeding, phone use, hours of service — every driver, every mile, time-stamped and scored. A safety director signs into that terminal every morning and works from it. It has simply never travelled from that terminal into a submission.

Loss runs describe what a fleet's risk was. Driver behavior describes what the operation is.

The cost of that omission shows up plainly in this account. Across the ten years, claim frequency fell from 1.92 to 1.72 per million miles while premium rose sixty percent. The fleet became measurably safer and paid substantially more, because nothing in the renewal file described the company it had turned into.

Behavioral evidence is also what makes a retention defensible to the people who have to accept it. No captive manager writes a fleet on the strength of a favorable loss triangle alone. They write fleets that can demonstrate the losses will stay favorable — and that demonstration is an operating rhythm, not a report.

Part four

Where the cost actually hides

Layering is the step brokers skip, and it is the step that decides the answer. A loss total tells you nothing useful about retention. A loss distribution tells you everything — which band of loss a fleet experiences routinely and could comfortably fund itself, and which band it genuinely needs an insurer for.

Sixty percent of this fleet's capped losses fall below $100,000. That single fact is what makes a $100,000 retention defensible — the carrier had been paying a full insurance load, year after year, on a band of loss it experiences as a matter of routine.

And the forty percent above $100,000, including the $631,812 that breached a million dollars outright, is exactly why the retention should not go higher. The distribution argues for a specific answer, not for maximum aggression. That is the difference between an analysis and a sales pitch.

Part five

What changed is not the analysis. It is what a machine can now read.

Part two claimed the three weeks had become an afternoon. That claim deserves more than an assertion, because the word “AI” on its own explains nothing at all. Here is the renewal as it happens today, and what specifically has to be true for it to happen differently.

Five things have to be true for the same work to finish in an afternoon. None of them is magic; all of them are engineering.

Why this could not be done in 2015

The obstacle was never computing power. It was that a loss run is not a data file — it is a document, with a carrier’s own layout, its own field names, its own conventions about what belongs in which column, printed or scanned or exported at whatever fidelity the portal allowed. Software that expects a template needs a template for every carrier, and there are thousands of them. That is why three decades of automation attempts stalled: each one solved four carriers and broke on the fifth.

What changed is that models now read documents much the way a person does — finding the date of loss whether it appears as a column header, a stamped label, a footnote, or simply by where it sits on the page. No template. No mapping project. That is the unlock, and it is genuinely the whole of it.

But reading it is not the same as understanding it

This is where a great deal of what is currently sold as AI stops being useful.

A general-purpose model will extract a loss run competently. It will not know that $0 paid against an open reserve means something entirely different from a closed claim at $0. It does not know that valuation dates must be aligned before two carriers’ years can sit in the same column. It holds no view on whether allocated loss adjustment expense sits inside the incurred figure — which varies by carrier, and changes the answer.

Extraction without that knowledge produces a tidy spreadsheet of wrong numbers. Which is worse than no spreadsheet, because it looks finished.

The 1,100 validation rules in the table above are not a feature. They are decades of that judgment written down and made executable — the difference between a model that can read insurance documents and a system that knows what it has read. It is also the honest answer to a fair question: why not simply upload the documents somewhere and see what comes back? Because the reading was never the hard part.

None of this does anything a capable analyst with three uninterrupted weeks could not do. It does it in the time a renewal actually allows.

That is the whole claim, and it is worth keeping it that modest, because the consequence is not modest at all. When the analysis costs three weeks, a producer reserves it for the one account they cannot afford to lose. When it costs an afternoon, it becomes the way they walk into every renewal on the book — and a technique becomes a competitive position.

Part six

What it changes at the table

There is a version of the renewal conversation that almost never happens. Tom Lucci has been having it for thirty years, and it sounds like this.

Read what is actually being offered. The competitive advantage does not belong to the broker. It belongs to the motor carrier, in the trucking business. A fleet carrying a lower cost of risk than the carrier bidding against it on the same lane has a margin advantage that compounds every year it holds.

That is a different conversation from a price, and it is a different relationship afterwards.

It also changes who else finds you useful. Captive managers get actuarially grounded fund sizing instead of a number somebody negotiated. Underwriters get the exposure evidence they have been saying all along is the main event. The broker stops being one of three quotes and becomes the analytical hub the whole program runs through.

Part seven

Where this does not work

Three situations where the honest answer is to leave the program where it is. We would rather say so here than have you discover it on an account.

When the safety culture will not change

Retention only pays if losses stay contained. A fleet where ownership is disengaged from safety, or where dispatch ignores telematics alerts, will not hold a retention. Putting one in place transfers risk to a client who has no mechanism for managing it.

When the client genuinely cannot absorb volatility

A carrier running thin working capital may be entirely right to pay the insurance load in exchange for certainty. Captive “B” saved this fleet $3.2 million across a decade and still cost it $98,840 in one particular year. Some balance sheets should not have to meet that year.

When the loss history is too short or too thin

Fewer than five credible years, or a fleet small enough that a single severe claim dominates the distribution, and the analysis produces a number with false precision attached to it. Volume is what makes a loss pick mean anything at all.