Two ways to buy the same policy

The rule

Score green, and the purchase stops

Sonnet scores every prospect with Policy Value Metrics, built from details like credit history, and sorts them into three tiers. When a green customer reaches the end of the quote flow, they cannot bind online. They have to phone a licensed agent, who walks them through the purchase over a shared screen.

The rule earns its place. The hypothesis behind it is that green customers finalize a better policy when an agent guides them through it: coverage gets corrected, the questions people are too unsure to ask a form get answered, and the purchase actually completes. It was also meant to feel like priority treatment: the segment worth the most gets a person who helps them fit coverage to how they actually live. Nobody wanted to delete that conversation.

The problem is that it is compulsory. The business makes its most valuable customers wait for a phone slot at the exact moment they have decided to buy.

The flow in production today. The point is the dead end at the end of it, not the screens.

What the data said

The gate leaks the customers it was built for

Before touching the flow I went into product analytics to find out what the requirement actually costs. Two things showed up, and both of them are money.

Why it leaks

I split two events by device: who initiates a purchase, and who confirms one. Initiation skews mobile. Confirmation skews desktop. People explore on their phone, then commit somewhere they can read every line.

That matters because the agent call is a desktop ritual. It assumes a screen share and an hour of attention. Dropped on someone browsing on a phone, it does not assist, it interrupts.

So the call was not the problem. Its placement and its obligation were. Keep it, offer it, stop requiring it.

What I could not change

Four fixed walls

Wall three is the one that changed how I worked. I come back to it below.

Optional, not removed

The gate was right. Compulsory was the mistake.

The easy version of this project is an argument that the call should go. I never made it, because the guidance hypothesis is probably true: a green customer walked through their coverage does finish with a better policy. Deleting the call deletes that.

1. 01

   Remove it and you lose what works. The conversation catches misconfigured coverage, answers what a form cannot, and is how this segment was meant to feel prioritized rather than processed. That value is real and worth protecting.

2. 02

   Require it and you lose the customer. A mandatory call is a queue, and a queue arrives at the one moment the customer has already decided. That is the leak.

3. 03

   Optional keeps both. Everyone who wants guidance still gets it, and the offer itself is the priority signal. Everyone who does not stops being charged a wait for it. The business gives up nothing it was actually collecting.

Autonomy was the point. Not fewer calls, not a cheaper funnel: the customer decides how they want to buy.

It is also why there is no exception list. Once the self-serve path exists, nothing in conversion needs a compulsory call. The place that judgment still belongs is retention, where a cancellation can be worth a human intervening.

Three ways to offer a call

All three dismiss. The question is what happens next.

Every option opens the same way, with a dismissible modal after the quote is generated. They differ on one axis: once someone has waved the call away, how does the offer stay reachable without becoming a pest?

I shipped B and C together. B keeps help within reach for everyone, C reaches out only to the people who look stuck. A was cut: familiarity is not worth burying the offer.

figjam · call-back system options

The board the three options were argued on, with design and dev in the room.

Building it to find out

A prototype was the only honest way to review this

Reviews kept stalling on the same question, and it was not one frames could answer: how does the offer feel when it follows you? Our tooling could not run the triggers, so I built the component instead. I worked with Copilot to code the behaviour and published it on CodePen, dismiss states and resurfacing rules included, then put the link in the review invite.

Try it below: dismiss the modal, adjust a coverage, and let it sit.

The live mockup, coded with Copilot. Or open it on CodePen.

Three things only the working version could tell us

1. 01

   Dismissal is fine, timers are not. Resurfacing on a countdown read as nagging, so the triggers moved to interaction signals instead.

2. 02

   The nudge drifted. In motion it read as a foreign element and had to be rebuilt against the coverage-details pattern.

3. 03

   Proximity carries the meaning. Separated from the coverage toggle it refers to, the nudge reads as an unrelated ad.

The critique got sharper because people were reacting to behaviour instead of imagining it.

Who had to agree

One offer, four different flows

The room was a VP, the product owner, my design manager, and directors. What moved them was not the concept. It was showing that I had found the bottleneck, understood why the gate existed, and could lay out the ways out of it with their trades attached. Options with consequences beat a recommendation.

The hard part was coverage. Sonnet does not have one purchase flow, it has four, and "the offer appears after the quote" means four different placements with four different sets of adjacent content. Ontario auto was the tension: a brand new flow, released July 1, so the offer had to land in something still settling.

Approval came once it was one rule with four placements, not four features.

Hi-fi

Saying it in Sonnet's own voice

With B and C settled, the work moved to hi-fi in Sonnet's design system. The behaviour was already decided, so this phase was about tone: how a customer is told they can decline a call without making the call sound optional in the sense of unimportant.

Three surfaces went to build: the choice modal after the quote, the nudge in coverage details, and the resurfaced modal.

What changed between wireframe and hi-fi

What we are measuring

In market, and what would prove me wrong

The hybrid is running as an A/B test on the green segment. Control keeps the required call, the variant offers it. Early signals point toward the optional path, consistent with the device split that started this. Real numbers come with the readout.

What I would do differently

I built the prototype after the concepts were already drawn. Building first would have killed option A in an afternoon instead of a review cycle.

And I set the resurfacing triggers from judgment, not data. They are heuristics, and the pilot is the first chance to replace them with real thresholds.