Lemonade

Lemonade, a digitally native insurer operating in the competitive renters insurance segment, faced a core conversion challenge: email engagement metrics were decoupling from actual purchase intent. Open rates and click-to-open rates (CTOR) responded well to certain messaging variants — particularly price-led ('From $5/mo') and convenience-focused ('No hassle, instant') copy — but these same variants failed to drive quote requests at the decision step. In P&C insurance, where acquisition costs are high and qualified lead volume directly influences loss ratio economics, spending campaign traffic on underperforming variants carries real cost. The team was effectively subsidizing engagement theater while conversion lagged.

Lemonade deployed an AI decisioning platform built on reinforcement learning — specifically a multi-armed bandit architecture — to dynamically allocate email traffic across content variants in real time. Rather than optimizing for open rate or CTOR, the system was configured to treat conversion and quote-intent sessions as the reward signal. The platform ran across both new-lead acquisition and cross-sell campaigns, testing multiple messaging frames simultaneously at scale across millions of users. Critically, the bandit was configured to archive non-performing variants aggressively rather than keeping 'interesting' tests alive on depleted traffic. This shifted the operating model from manual campaign overrides to always-on learning with faster predict-approve-ship cycles, enabling the CRM team to reduce regret from underperforming variants without constant human intervention.

Campaigns delivered double-digit CTOR lifts across the portfolio, with conversion gains concentrated in variants using a 'friendly + speed' messaging frame. Key outcomes included:

• CTOR lift: Double-digit percentage gains across multi-campaign rollouts
• Allocation consolidation: Bandit traffic collapsed behind the top 1–2 performing variants, eliminating wasted spend on laggards
• Unsubscribe rates: Held stable as allocation tightened, indicating the winning frames were not fatiguing subscribers
• Operational shift: Team moved from reactive manual overrides to an always-on decisioning loop with faster iteration cycles

The clearest qualitative outcome was a validated messaging thesis: human, trust-building tone paired with concrete time-to-value language outperformed price-first and generic convenience claims at the conversion step.

• Optimize to the outcome you're paid for — in insurance, that's quote requests and conversions, not opens. Use engagement metrics only as early exploration signals.
• Aggressively shrink regret: auto-archive underperformers quickly rather than keeping 'interesting' variants alive on scarce traffic budgets.
• 'Friendly reminder…' combined with concrete time-to-value ('in seconds') consistently outperformed both price-first and convenience-cliché messaging at the decision step.
• Statistical ambiguity is a feature, not a bug — let the reinforcement learning system mature through gray-zone windows rather than forcing early manual calls.
• AI decisioning works best as an operating system for marketing decisions, not a one-off test; the structural shift to always-on learning compounds over time.