The Quote on the Earnings Call
In early 2024 Progressive's CEO Tricia Griffith was asked, on an earnings call, why the company kept widening its lead in personal auto profitability. Its largest competitors were struggling to write business at the same combined ratio. Her answer was the kind that sounds dull until you sit with it. Progressive, she said, had spent more than a decade collecting driving-behavior data through its Snapshot telematics program. The company now had visibility into how individual drivers actually behave at the wheel that the rest of the industry did not have. The pricing followed the data.
A Carrier Management Telematics Master Class the year before had made the same point in plain language. Progressive was not running a smarter pricing algorithm than State Farm or GEICO. It was running a pricing algorithm on a different kind of input. The 2024 Annual Report made the scale concrete: tens of billions of vehicle-miles of behavior data, each mile attached to claim outcomes Progressive observes directly through its own underwriting and claims operations.
Run against the two-question test in The Berkshire Test for AI, Griffith was describing both axes at once. The location axis (Goldratt's Theory of Constraints) lined up because telematics was deployed at the single step in personal auto that determines profitability: risk selection at the point of quote. The compounding axis lined up because every condition the AI factory framework from Iansiti and Lakhani names was present at scale. Two decades of Snapshot policies had built a record of driver behavior labeled by claim outcome that no rival can buy, replicate, or shortcut. The flywheel was running.
Where Progressive Makes Its Money
The bottleneck of personal auto insurance is the moment the insurer commits to a price. Get the price right (charge the safer drivers less and the riskier drivers more, accurately, within the same rate class) and the company wins on loss ratio for decades. Get it wrong and the safer drivers shop elsewhere for a better price while the riskier drivers stay, the book skews steadily toward the riskier half of every rate class, and the loss ratio bleeds. Economists call that failure mode adverse selection. The insurer who knows less about driver behavior than the customer does ends up paying for the gap.
Customer acquisition is not the bottleneck (cost-per-quote is comparable across carriers). Claims handling is not the bottleneck (the operational efficiencies are broadly distributed). Capital deployment is not the bottleneck (every large carrier accesses the same reinsurance markets). The bottleneck is the quote screen, the single moment where the insurer commits a price to a risk it has only partially observed.
In Goldratt's vocabulary from The Goal (1984), the underwriting decision is the constraint because every dollar of profit or loss flows through it. Everything else in the company either feeds the underwriting decision or is downstream of it. Investments that improve underwriting precision are the ones that move the P&L; investments that improve adjacent steps without strengthening underwriting do not.
The historical underwriting input set was a small handful of static signals: age, ZIP code, vehicle make and model, prior claim history, credit score where state regulation permitted. Combined, the carriers in the 1990s could segment risk into a few dozen rating cells. The problem with the static-signal approach is that it captures the driver class a customer belongs to, not the driver they actually are. Two thirty-five-year-olds in the same ZIP code driving the same model at the same coverage tier can have wildly different actual risk. The static input set cannot tell them apart, so the carrier prices the average and gets adversely selected the way the previous section described.
Telematics broke that pattern by adding a behavioral input set the static signals could not generate. The question at the quote screen is no longer which class does this driver belong to? but given how this specific driver actually behaves at the wheel, what is the loss-cost expectation on their next policy term? The carrier that asks the second question consistently, with a large enough data set to support it, wins on loss ratio.
How Snapshot Works
Progressive launched Snapshot in 2008. The program was opt-in for new customers willing to plug a small device into their car's diagnostic port (and later, through a mobile app). The premise was simple in shape and operationally hard to execute. Collect driving-behavior data from the policyholder's own car (acceleration profiles, braking patterns, hours of day driven, sudden-stop frequency, miles driven). Attach it to the policy's eventual claim outcomes. Feed the labeled pairs back into the pricing model. Over time the model learns which behavior patterns predict which kinds of claims. Progressive then prices new business with a behavioral score on top of the static signals every other carrier already had.
The 2024 Annual Report is the closest thing the public has to a scale check: tens of billions of vehicle-miles, accumulated across multiple policy generations, attached to claim outcomes Progressive observes directly. State Farm and Allstate run telematics programs of their own, but their public disclosure on accumulated mileage and on integration depth with the live pricing model has been thinner. The head start matters in ways the next section will spell out.
The machine-learning model itself is a fairly standard ensemble layered on top of the existing actuarial pricing model. The interesting engineering is the data pipeline. Three pieces matter. The first streams data off millions of vehicles in something close to real time. The second joins that behavioral data to claim records in the policy administration system. The third produces the rate filings for state insurance departments that justify the use of behavior-based pricing factors. The model is the visible layer. The pipeline is the moat.
| Condition | Score (0–4) | Evidence sentence |
|---|---|---|
| Proprietary data origin | 4 | Snapshot beacon data is generated by Progressive's own policies at the wheel of customer-owned vehicles, not bought, scraped, or shared with a third-party data marketplace. |
| Self-labeling workflow | 4 | Every policy term either does or does not produce a claim, and Progressive's claims operation supplies the loss-cost label that re-trains the pricing model on the next cycle. |
| Decreasing marginal cost | 4 | After 17 years of operating the program, the marginal cost of an additional Snapshot policy is dominated by the app and processing pipeline that are already paid for; the hundredth million miles cost less to process than the first million. |
| Defensible asymmetry | 3 | Competitor telematics programs (State Farm Drive Safe and Save, Allstate Drivewise) exist and accumulate their own data, but the multi-policy-generation head start on integrating behavior data into the live pricing model produces a competitive lag a rival cannot close by buying a vendor product. |
The Four Conditions
Condition 1, proprietary data origin. Snapshot data is generated where the data is, to use the phrasing James Currier of NFX uses in his data network effects piece. It does not exist in a public dataset, a vendor marketplace, or any aggregated bundle a competitor can subscribe to. Each beacon reading is attached, by Progressive's policy administration system, to a policy number, premium tier, coverage selection, and (over time) to claim outcomes Progressive observes directly. A rival carrier wanting the same input set today would face two decades of opt-in customer acquisition, regulatory approvals for behavior-based pricing in every state, and the customer trust required to sustain telematics adoption at scale. The provenance is the moat.
Condition 2, self-labeling workflow. The work itself produces the labels. Every Snapshot-enrolled policy generates a stream of behavior data during the term. At the end of the term, the policy either has produced a claim or has not. The presence, severity, and category of any claim is the ground-truth label for the pricing decision that opened the term. The labels arrive on a natural cadence (the policy term). They are first-party. They are attached to the behavioral input set without any external labeling budget. This is exactly the AI factory pattern Iansiti and Lakhani describe, applied to underwriting: data, predictions, and outcomes produced continuously by the same operation, each cycle's outputs sharpening the next cycle's inputs.
Condition 3, decreasing marginal cost per cycle. Brynjolfsson, Rock, and Syverson's productivity J-curve names the dynamic. Big new technology investments suppress measured productivity early and amplify it later, because the supporting infrastructure investment has to be paid off before the gains land. Progressive is on the rising side of that J-curve. The ingestion pipeline, the rating-engine integration, the regulatory-filing infrastructure, the customer-facing app, and the policy administration tie-in have all been paid for. The marginal cost of one more Snapshot policy in 2026 is dominated by the app and the cloud processing fees, both of which decline per-unit at scale. The first model retraining cycle, back in 2010, required a heroic engineering effort. The 2026 retraining cycle is a scheduled operations event.
Condition 4, defensible asymmetry. This is the hardest condition to argue for any case, because Casado and Lauten's Empty Promise of Data Moats critique is real: most claimed data moats plateau quickly, and competitors with less data can usually reach the plateau by buying a vendor product. The Progressive case survives on two grounds. The first is that the relevant asymmetry is not raw mileage count, where State Farm and Allstate are catching up. It is the integration depth between behavior data and the live rate-filing pipeline. Borrowing from Carlota Perez, this is a deployment-phase capability rather than an installation-phase one, and it is exactly the kind of asymmetry an incumbent can build that a vendor product cannot replicate. The second ground is that the moat is regulatory as well as technical: Progressive has rate filings approved in every state for behavior-based pricing factors, and a rival catching up on data still has to catch up on the multi-year state-by-state approval cadence. The asymmetry is real but capped, which is why I score the condition 3 rather than 4 in the table above.
The composite verdict: Snapshot satisfies all four conditions at material strength, and has been satisfying them long enough for the J-curve to have inflected. Progressive passes the test.
The Easier Wrong Choice
Imagine a Progressive that, in the early 2010s, had spent its AI budget on marketing personalization instead of Snapshot. That is the wrong-place alternative. The reason to walk through it is to surface the version of the same choice that a thoughtful executive at any non-technology firm in 2026 is making right now without realizing it.
The wrong place. AI-driven marketing personalization, in the form of behavioral targeting on display advertising, dynamic landing-page optimization, lift-modeled outbound calling campaigns, and recommendation-driven cross-sell of homeowners' insurance to auto customers. This is a defensible-on-paper investment for any large personal-lines carrier. The vendor ecosystem (Adobe, Salesforce Marketing Cloud, the major demand-side platforms) is mature. The metrics tooling is built. The consultancies have reference architectures. The early-cycle wins on cost-per-acquisition are large and easy to attribute.
Why it would have looked attractive. Cost-per-acquisition is a metric every CFO understands and every CMO is measured on. An alternate-history Progressive that deployed personalization aggressively in 2012 would have reported a 15-20% drop in CPA within the first eighteen months, against a baseline where direct-response auto-insurance marketing had been stable for years. The internal narrative would have been compelling. AI driving real efficiency in the largest line item under the CMO's control. Wins showing up in board-level metrics quarterly. The same toolset portable to homeowners' and small commercial lines.
The failure mechanics. Casado's critique applies with full force here, because the underlying data and machine-learning capability are vendor-owned. State Farm, GEICO, and Allstate would have observed the CPA improvement, hired the same agency and consultancy, and bought the same vendor packages within 18-to-24 months. By 2014 the CPA gap would have closed. By 2016 it would have inverted in places, as later competitors bought improved third-generation tooling Progressive could no longer differentiate against. Meanwhile the loss ratio, the metric that determines whether a policy was good to write in the first place, would not have moved at all, because the marketing investment did not touch the underwriting decision.
The time to failure. Approximately 24 months from the first CPA improvement to the first quarterly investor question about why the combined ratio was not following. By month 30 the analyst notes would have shifted from celebrating marketing efficiency to asking whether the company had a structural underwriting disadvantage relative to GEICO. By month 36 the technology leader who had championed personalization would have been at a different company, and the successor would have inherited a sunk cost on marketing tooling and a deteriorating loss-ratio trend, with no AI program at the bottleneck.
The early-warning signal. A careful observer in 2012 could have flagged this in advance with one question. Does this AI investment generate first-party labels that feed back into the metric that determines whether the business makes money? For marketing personalization the answer is no. The labels are click-through and conversion rates, which feed back into marketing-channel optimization, which is downstream of the bottleneck. For Snapshot the answer is yes. The labels are claim outcomes, which feed directly into the underwriting decision that is the bottleneck. The 2012 executive who applied the two-axis test would have steered the budget to the boring program at the boring bottleneck, and would have been right.
What Progressive Teaches
The general lesson from Progressive is that the bottleneck of a regulated, capital-intensive, claims-driven business almost always lives at the moment the company commits to a price or a decision, in whatever vocabulary the industry uses. In auto insurance it is the quote. In healthcare it is the diagnosis. In commercial lending it is the credit decision. In manufacturing it is the bill-of-materials commitment. In each case the bottleneck is the step where the firm commits resources to an outcome it has only partially observed. The AI investment that compounds is the one that improves first-party visibility at that step.
In my experience working with leaders outside insurance on their own AI capital allocation, three questions in sequence usually find the right program. Where is your business committing resources to outcomes it cannot yet see? What first-party data, generated by your own operations, would let you see those outcomes earlier or more accurately? Does the candidate AI investment feed that labeling loop? The answers usually point at a single program. The discipline is to fund that one and refuse to fund the others, even when the others are showing wins on local metrics.
The other lesson is patience. Snapshot took the better part of a decade to inflect into a P&L signal large enough that competitors started copying it. The gains compounded because the data set compounded, and the data set compounded because the program kept running through years when quarterly returns did not justify it on their own. The four conditions predict the J-curve shape, but predicting it is different from sitting with it through three executive transitions and four budget cycles.
What You Can Do
If you run a non-insurance business, the Progressive pattern translates into a Monday-morning exercise that takes ninety minutes and produces an actual capital-allocation answer. Pull up your top three active or proposed AI investments. Write the bottleneck of your operation in one sentence. Place each investment on the two-axis matrix from the framework essay.
For each investment, run the three diagnostic questions from the previous section. Where is your firm committing resources to outcomes it cannot yet see? What first-party labels, generated by your own operations, would close the visibility gap at that step? Does this AI investment feed that labeling loop? If yes, you have a candidate compounder. If no, you have a candidate Roman Candle, regardless of how compelling the early-cycle metrics look.
The Snapshot lesson on patience is the harder follow-through. A compounder at the right bottleneck will look unremarkable for two-to-four budget cycles before the J-curve inflects. The discipline is to keep funding it through the unremarkable years, and to defund the visible-but-wrong-place alternatives. Progressive's leadership held that discipline through three CEO transitions. Most firms cannot, and that is where the asymmetry lives.
Back to the framework: The Berkshire Test for AI.
Continue the series: Deere's Physical-World Data Loop — physical-world labels at the row-crop spraying step. Mastercard's Network of Labeled Outcomes — chargebacks as the label substrate, with a candid asterisk on the Visa peer. Mayo Clinic's Outcome-Labeled Corpus — longitudinal clinical outcomes labeling decade-old ECG readings.