Issuing a policy is not selling it

In short

I walked La Caja's digital channel through its public flows to understand how a channel like this gets measured. I ran into a problem before the one I was looking for: online quoting lives on six separate platforms that measure different things, and today you cannot even compare conversion between products. Then I looked at how ten other leading insurers quote, to see whether the same problem repeats. Five give no signal from outside; of the five that do, one shares it and four solved it.

The heart of the case is the metric. Measuring a team by "new policies" rewards what generates no value: a policy that lapses before clearing its first payment has already spent the acquisition and the commission, and left no premium. The scale of the discard is published by Argentina's insurance regulator (SSN): in motor insurance, 12.76% of the stock cancels every quarter, for any reason and across the whole book. I chose La Caja as the anchor because motor is 82.7% of its production, so almost all of its business depends on measuring a single line well.

To argue with numbers I built a metrics model calibrated against the same SSN tables. Of the 53 numbers that feed that simulation, I chose 43 myself, each one declared; only 6 come straight from published tables.

Issuing a policy is not selling it. The unit of measure for an insurance digital channel is the premium still alive at six months, divided by what it cost to bring it in.

What is data and what is my own choice

Before any number, the line that separates the verifiable from what is mine.

Six quoters that measure different things

Before deciding what to measure I went to see what exists, and walked the public quoting flows. Everything in this section is verified against the production site as of August 10, 2026.

The first thing was counting. Car, home, motorcycle, bicycle and health each quote on their own subdomain, and there is a sixth, legacy stack that quotes notebook, handbag and elder-care coverage on two routes of a host whose root, on top of that, redirects to the institutional site. That makes six separate quoters; counting the personal, business and broker portals, the ecosystem lives on at least nine hosts. Only two of the six declare which version they serve.

In practice this means that today you cannot compare car conversion against motorcycle or home. They are six implementations measuring different things, each on its own domain and with its own session. The question "does the digital channel work?" has no table to rest on.

I was left wondering whether this was one company's quirk or a debt of the whole industry, so I looked at how ten other leading insurers quote (August 11, from the code their sites serve). Five give no conclusive signal from outside, so I make no claims about those. Of the five with signal, only Mapfre shares the problem: it splits car, home and personal accident into separate subdomains. The other four quote unified: Provincia Seguros uses the same quoter changing only the route; Rivadavia picks the product by parameter; San Cristóbal and Sancor concentrate quoting on one domain. When everything lives on a single platform where the product is a route, the funnels share session and measurement, and adding a product does not open a new silo. The problem repeats in the market, but it is not inevitable: most of the insurers with signal already quote unified.

The walkthrough turned up another finding. Of the six, the only one that asks for personal data before showing the price is the motorcycle quoter: date of birth, phone and email in the first step. The car quoter goes the other way, asks for year, make, model and version, shows the price, and only then asks who you are. When someone compares the conversion of the six products side by side, motorcycle will look worse, and someone will read that gap as lack of demand when the origin is in the form design.

The walkthrough also surfaced two product rules that change the math.

• The car policy is monthly and renews automatically (it is in the quoter's own FAQ). That means twelve continuity decisions and twelve collection events a year: retention reads month by month, and every failed collection is a chance to lose the policy for a reason unrelated to the product.
• For used cars there is an inspection where the customer uploads photos from the browser. It is the only step where the work comes after the decision to buy, with the intent already won, and it mixes three situations worth separating: brand-new cars skip inspection, the in-person one is a different process, and within the photo flow the user who abandons and the system that rejects are different problems.

From all of this I took away a sequencing conclusion: the area's first deliverable is an agreement on what counts as a quote and what counts as a new policy across the six platforms, before any screen.

On the side, two small defects still in production. The «CONOCÉ NUESTRA APP» button in the header (the "check out our app" call to action) points to seguros-dev.lacaja.local, an internal development server that does not exist outside the company's network. In the code the server delivers the link is fine: it breaks once the page finishes loading in the browser, which is why it can only be found by using the site. The other one is in the broker portal, the channel that concentrated 69.9% of the market's distribution in 2025: its three self-service guides (second-factor authentication, password reset and electronic payment) have their «Descargar» button with no address assigned, they download nothing. Neither moves premium on its own, but every failed attempt probably ends in a call to the contact center, and both could be fixed quickly.

Why new policies are not enough

In motor insurance, the SSN's Q1-2026 property policy table says 12.76% of the standing stock cancels each quarter, and that cancellations equal 16.36% of what gets issued. Across all property lines they are 18.91% of issued policies, almost nineteen in every hundred. Part of that is endorsements and replacements from a change of vehicle, so I read the figure as a ceiling on churn rather than a pure measure. Even as a ceiling, the scale is enough for the point: measuring by new policies means counting policies that in good part will never be collected.

And there is a simpler reason. If you measure the team by new policies, it produces new policies: cutting price, loosening underwriting, buying cheap traffic. Choosing the metric is choosing a behavior.

The north star metric and its counterweight

The one I chose is Persistent Annualized Premium: the annualized premium of digital-channel policies still in force and current six months after issue. The usual alternatives each fail on their own side:

• New policies: they treat a coverage that holds and one that lapses in the second month the same.
• Issued premium: it measures the promise; collection comes later.
• Collected premium: correct but slow. By the time the figure closes, the decision that caused it is three months old.

What mattered to me about persistent premium, more than precision, is one property: no single area can move it alone. Marketing can show a falling cost per sign-up and underwriting a rising quality, each on its own, while the business gets worse. Persistent premium forces them to share the result, and for that same reason it will be awkward to adopt.

It is not enough on its own either, because any volume metric inflates by overspending. It comes paired with persistent premium over investment, which is the one that decides the budget: where the next peso yields most.

Correcting cost per sign-up does not change the ranking

The correction that looks obvious is to divide cost per sign-up (CPA) by the persistence rate and keep the cost of the policy that survives: CPAP = CPA / persistence. I tried it in my simulation and it did not move a single rank. The five channels stayed in the same order they already had by cost.

And it had to happen, because there is an asymmetry underneath. Persistence is a proportion: as long as the worst channel keeps more than half its book, the gap between the best and the worst cannot exceed 2 to 1. Cost per sign-up has no such ceiling. The correction moves the level and leaves the order where it was.

What does reorder is persistent premium per peso invested, because it adds the missing variable, ticket size. In my simulation, the embedded channel brings the second-cheapest sign-ups and ends up last in premium left behind; the conversational one, the most expensive per sign-up, climbs to third.

Where every number comes from

The number that weighs most in the simulation is the quarterly motor cancellation rate, which the SSN publishes. Before using it I checked my extraction of the tables against three figures I had already published in an earlier piece on collections (18%, 25.6% and 18.9%): my extraction gives 18.04%, 25.57% and 18.91%, so all three reproduce. The load runs ten checks of that kind, and if any fails nothing gets in.

A failed collection is a monthly leak

With a monthly policy there are twelve collections a year, and each one can lose the policy for a reason that is not about the product. The distinction I care about is between the cancellation the customer asks for and the one that happens because the debit bounced through its retries: the first is a value problem, the second a process one, and different teams with different budgets fix them.

Collection capacity is finite: the team works a capped list. The product decision is which list to give it and which yardstick to judge it by, and the yardstick I care about is how many real failures land in the list the team actually gets to work. To size it I simulated a prioritization of the month's collections by risk of failure.

All of this runs on a simulation, but the exercise is the same with the company's real data. That is where the output stops illustrating and starts deciding: what list size is worth working and what recovery to expect from it. In a case like this my recommendation would be to start with the short list, where each action yields most, and widen it as long as the recovery keeps paying for the cost of working it; and between two models that perform alike, pick the one the collections team can explain. Building and maintaining the model belongs to the data team; on the product side sit the success metric and the list cap.

The lever already exists in the product and can be verified from outside: the quoter applies the promotional price to automatic debit. What is missing is the number only the company has, how much mix that discount buys and whether the extra persistence pays for the premium given up.

The hypothesis left open

Three observations from three different pieces point the same way. Among vehicle lines, motorcycle cancels quite a bit more than car: 25.57% of what it issues against 16.36%, per the SSN's Q1-2026 table. The motorcycle quoter is the only one of the six that asks for personal data before showing a price. And the conversational channel does the same: when I tested it as a customer for another project, it asked for my ID, date of birth and license plate before any price, and in the end deferred the quote.

My hypothesis: flows that capture the lead before delivering value produce lower-quality sales, and that quality comes back later as cancellation. I cannot prove it from outside, because the correlation is market-wide and each flow design belongs to one company. But inside it is quick to test, measuring early cancellation of motorcycle against car, segmented by lead origin and by whether the price was shown before or after asking for the data. If it holds, it connects a form-design decision to a line on the balance sheet. If it does not, the alternative explanation interests me too: the line is simply more fragile and the flow has nothing to do with it.

What this work cannot support

It is an external analysis, on public data, and it carries the limits of that starting point.

• 43 of the 53 numbers that feed the simulation I chose myself. Everything that comes out of it (channel rankings, funnel rates, collection-model results) shows how the simulation behaves, nothing more.
• No funnel rate is public. The step structure is verified by walking the quoters; the levels are chosen.
• There is no public figure for the digital channels' weight in production, and it is the denominator of the whole sizing.
• The cancellation rate the regulator publishes includes non-financial reasons, so it works as a ceiling on churn. The simulation uses it anyway, overstating the loss on purpose.
• In a simulation the truth about each policy is known, so I can measure how much the prioritization recovers. With real data that yardstick does not exist.
• The two list errors do not cost the same: working a collection that was going to clear costs one extra action; not working one that fails costs a policy. The list's cutoff should come from that asymmetry, and Operations has that number.
• The regulator's channel split does not break out by line: the most granular cut is property versus personal.

The four questions to start with

None of the above runs without answers that only exist inside. The first three decide whether the model is worth building; the fourth sizes the collection lever.

1. What was the last important decision you made with data, and what was missing to make it with more confidence? I want to know which concrete decision would improve with better data.
2. What share of production comes from digital channels? Without it there is no telling whether two points of conversion are worth a meeting or a quarter.
3. Is there campaign traceability down to the issued policy? The whole efficiency axis depends on a join that may not exist.
4. How much does the automatic-debit discount return? It is the lever from the collections section, and its return is a number that only exists inside.

With the first three answered, the first move is still the definitions agreement across the six platforms, because without it any comparison between products produces a false conclusion. Measurement arrives once the numbers mean the same thing in every column.

One front was left out on purpose: the business portal is measured with a different yardstick, because it saves operations instead of bringing premium, and I am working on it as its own case.

This is independent work, done with public sources. I have no commercial relationship with La Caja, and never have.