# The policy video that kept more customers after a price rise > Source: > Published: 2026-11-07 17:00:00+00:00 ## Experiment summary [Allianz Benelux](https://www.allianz.nl/service/over-allianz/allianz-benelux.html) randomised 45,685 renewal customers between the standard email and a policy-data-driven personalised video. Lont reports churn over 100 days falling from 12.8% to 11.4%, while a second account of the programme shows that the wider renewal journey changed too. - Allianz Benelux tested personalised renewal video after a price increase from May to August 2025. - Customers were randomised 90/10 between a video journey built from policy data and the standard email. Churn was measured for 100 days. - Lont reports churn of 11.4% with video against 12.8% in control, a 1.4 percentage point absolute gap and a 10.9% relative reduction. - Lont says the system combines deterministic rules with selective AI for generation, translation and variants. Public sources do not name the model or separate out the AI component. - A partner account from Pebble says timing, design and a prize draw also changed, so the cleanest reading is that the complete renewal treatment beat the incumbent communication. *Hacking Growth* lists concurrent promotions among the confounders a test summary should record, which makes the prize draw the first item to document. 13 min read A price rise turns a routine insurance renewal into a quiet referendum. The policy holder sees a higher number, remembers how rarely they spoke to the insurer and decides whether the cover still feels worth keeping. For Allianz Benelux, that decision window lasted 100 days. It was also long enough to test whether a different explanation could keep more customers. The company took 45,685 customers renewing after a price increase and split them at random. Nine in ten received a video assembled from their own policy data. One in ten kept the standard email. Then Allianz waited to see who left. This is the useful shape of an AI retention experiment. The model did not choose the business outcome, and an open rate did not stand in for loyalty. A governed content system produced the treatment, a holdout protected the comparison and churn decided whether the work mattered. ## A renewal message had to defend a higher price [Allianz Benelux](https://www.allianz.nl/service/over-allianz/allianz-benelux.html) operates across the Netherlands, Belgium and Luxembourg. It says it serves more than two million customers with 2,200 employees. The experiment covered car insurance customers at renewal, a moment when a premium increase makes the value of the policy unusually visible. According to the [trial account published by Lont](https://www.lont.ai/proof), Allianz’s existing communication was a standard renewal email. The hypothesis was that a video built from the policy itself would make the cover easier to notice and understand, remind the customer what they were buying and reduce avoidable churn after the increase. ## One policy became one video According to Lont, the treatment used the customer’s name, car brand and model, co-branded labels, policy details and no-claims recognition. The email carried an animated preview, and a click opened the customer’s video in a browser. Pebble’s account adds that recipients could enter a draw for a year of free car washing. The technical mechanism matters more than the moving pictures. [Lont describes its platform](https://www.lont.ai/) as a programmable video layer connected to CRM, billing, claims or policy systems. Customer data and business rules select approved scenes and fill the personal fields. The same inputs and rules are meant to produce the same output, which gives an insurer a traceable treatment instead of an unconstrained prompt. AI sits inside that controlled pipeline. Lont says it can draft scenes, translate, adjust tone and create variants, with locked text excluded and human review available. It can combine animation, actors, AI avatars or licensed AI voices. The public Allianz material calls the videos AI powered, yet it does not disclose the model, prompts or which individual scenes used generation. The runtime personalisation itself appears to be driven mainly by deterministic data and rules. The trial compared renewal journeys, not two AI models. Personalisation was governed by customer data and approved rules, and the public account does not isolate the effect of generated content. ## The holdout kept the standard email Between May and August 2025, Allianz put 90% of the 45,685 eligible renewal customers into the personalised video group and 10% into control. Lont calls it a randomised controlled trial and says Allianz ran and measured the experiment itself. Churn during the following 100 days was the primary business outcome. The allocation is lopsided, but the total audience is large enough that the control would contain roughly 4,500 customers if the split landed close to plan. The public report gives no exact group counts, balance checks, confidence interval, p-value or pre-registered analysis, so an outside reader cannot reproduce the statistical judgement from the published figures. Allianz also measured the delivery mechanism. Lont reports a 75% email open rate and says 41% of openers clicked through to the video. Those numbers show that the treatment reached a substantial share of the audience. They remain intermediate metrics. A customer can watch an elegant explanation and cancel the next morning, which is why the 100 day churn window carries more weight. ## Churn moved by 1.4 percentage points Lont reports churn of 12.8% in the standard communication group and 11.4% among customers assigned to personalised video. The absolute difference is 1.4 percentage points. Relative to control, that is the reported 10.9% renewal churn reduction. The absolute number is the better planning input. If a similar result held over 10,000 treated renewals, the gap would correspond to about 140 fewer cancellations. That is simple arithmetic applied to the reported rates, not a customer count published by Allianz. *Hacking Growth* treats churn as the mirror image of retention. Its Costco example reads a 91 percent membership retention rate as a 9 percent annual churn rate. Read the same way, the 100 day figures are retention of 87.2% in control and 88.6% with video. Sean Ellis and Morgan Brown argue that retention is generally the deciding factor in profitability, which is why a churn outcome carries more weight than an engagement metric. The result is stronger than a before and after comparison because customers entered treatment and control during the same period. Price conditions, seasonality and market noise should affect both groups. Random assignment also makes the control a better estimate of what would have happened to treated customers without the new journey, assuming assignment and outcome tracking worked as described. [Ruud van der Linden of Lont](https://www.linkedin.com/pulse/how-allianz-reduced-renewal-churn-109-personalized-van-der-linden-10see) published the detailed result and links [Martijn Huisman](https://nl.linkedin.com/in/martijn-c-huisman), Head of CX at Allianz Benelux, to the programme. Huisman says AI-powered personalised video improved open rates, click-through and loyalty. The quantitative public record still comes mainly from the supplier and its implementation partners, and Allianz has not published a methods report. ## The treatment was wider than video alone Lont’s proof page says everything except the medium was held constant. A [separate case published by Pebble](https://pebble.digital/case/allianz-nps/), the digital partner involved in the renewal flow, describes a broader change. It lists personalised content, a new design, delivery four weeks before renewal instead of on the renewal date, clearer next steps, consent management and the car wash competition. The two accounts do not line up perfectly. The Pebble page covers more than 60,000 customers and discusses NPS as well as churn, so it may combine a wider programme with the 45,685-customer renewal trial. It does not identify which elements were present in each randomised arm. Its engagement figures include 7.8% of recipients entering the competition, without defining the base, and it gives 41% as the share who watched the video where Lont gives 41% as the share of openers who clicked. Until Allianz publishes the protocol, the safest attribution is to the personalised renewal treatment as a bundle, not to the AI, the video or the prize draw by itself. *Hacking Growth* gives teams a template for closing that gap. Its test summary asks for the variants and target customers, key metrics, start and end dates, sample sizes, statistical confidence and power, and any confounding issues, such as other promotions that may have skewed behaviour. An earlier send and a prize draw fall squarely under that last heading, and writing them up arm by arm would turn the attribution question into a table. The commercial relationships are worth stating. Lont supplies the system and published the exact churn rates. Pebble helped build the journey and published its own case. Allianz measured the outcome, according to Lont, but has not released the raw data, uncertainty or allocation table. The result is a clear signal for the renewal journey as a whole, and the open question is which part of it did the work. ## A three-arm test could locate the gain The next test should split the bundle. One group would keep the original email and timing. A second would receive a redesigned, personalised email with the same policy facts and no video. A third would receive the full Lont personalised video journey. Holding the prize draw constant, or removing it from all three, would stop the incentive from doing invisible work. Churn at 100 days should remain the primary metric. Revenue retained, service calls, complaints, consent withdrawals and policy changes would make useful secondary outcomes and guardrails. Assignment should persist at customer level, and the analysis should report exact group sizes, the absolute risk difference and confidence intervals. *Hacking Growth* also suggests stating an expected gain in the hypothesis, so that a 1.4 point result can be read against what the team predicted. Allianz could then stratify the result by size of price increase, product, distribution partner and prior tenure. That is the cohort analysis the playbook recommends, breaking retention down by subgroup so an average does not hide who stayed and why. A video that works after a modest rise may fail when the premium jumps sharply, and a treatment that helps first-year customers may add little for someone who already understands the policy. One cheap addition would be a single question to customers who renewed: what nearly made you cancel? Ellis and Brown recommend asking people who completed a step what almost stopped them, because they met the same obstacles as those who left. The answers would show whether the explanation, the earlier timing or the price itself carried the decision. ## The copyable lever is the operating boundary A growth team does not need an ML group to reproduce the experimental structure. Pick one high-value journey, limit the model to approved content operations, map a small set of customer fields into deterministic rules and keep a holdout on the incumbent communication. Then connect exposure to the outcome the business owns. The boundary makes the AI usable here. Locked text protects mandatory language. Approved scenes keep the brand and policy explanation stable. The data pipeline personalises only the fields the journey needs. Segment-level telemetry can explain what people watched, while churn sits outside Lont’s player and judges the whole intervention. The next useful result is a comparison of the same policy explanation, delivered without video or an extra incentive, against the video at a comparable cost. That would tell Allianz which part of the renewal journey deserves further investment. | Experiment detail | What the reporting shows | |---|---| | Constraint | Reduce avoidable churn among car insurance customers renewing after a price increase. | | Hypothesis | A policy-data-driven video would make value and coverage easier to notice and understand than the standard renewal email. | | Intervention | A personalised email preview and browser video assembled from customer, vehicle and policy fields through Lont’s governed workflow. | | Comparison | Random 90/10 allocation between personalised video and standard email across 45,685 customers from May to August 2025. | | Primary metric | Customer churn during the 100 days after a price increase. | | Reported result | Churn fell from 12.8% in control to 11.4% with video, a 1.4 percentage point absolute gap and a 10.9% relative reduction. | | Decision | Allianz’s CX lead says the programme demonstrated value at key journey moments. Public sources do not state a formal rollout rule. | | Open questions | Exact group counts, uncertainty and model detail are not published, and the two public accounts differ on whether timing, design and the prize draw were part of each arm. | ## Sources and disclosure The core experiment mechanics and numeric result come from [Lont’s Allianz proof page](https://www.lont.ai/proof). Lont supplied the personalised video platform and therefore has a commercial interest in the result. The wider implementation account comes from [Pebble](https://pebble.digital/case/allianz-nps/), another delivery partner. Company scale comes from [Allianz Benelux](https://www.allianz.nl/service/over-allianz/allianz-benelux.html). The practitioner attribution comes from [Martijn Huisman’s LinkedIn profile](https://nl.linkedin.com/in/martijn-c-huisman) and [Lont founder Ruud van der Linden’s LinkedIn article](https://www.linkedin.com/pulse/how-allianz-reduced-renewal-churn-109-personalized-van-der-linden-10see). The retention and testing frameworks come from Sean Ellis and Morgan Brown, *Hacking Growth*. No source publishes raw customer-level data, conventional statistical uncertainty or a protocol that reconciles the treatment descriptions. ## FAQ ## What did Allianz Benelux test? It tested whether a personalised video built from policy data could reduce churn among car insurance customers renewing after a price increase, compared with the standard renewal email. ## How big was the test and what was the result? Lont reports 45,685 customers split 90/10 between May and August 2025. Churn over 100 days was 11.4% with video and 12.8% in control, a gap of 1.4 percentage points. ## Was it the AI or the video that cut churn? The public accounts do not separate them. Lont says personalisation runs mainly on data and approved rules, with AI used selectively, and Pebble says timing, design and a prize draw also changed. ## Who published the figures? Lont published the churn rates and says Allianz ran and measured the trial. Pebble published a related case. Allianz has not released a methods report or the raw counts. ## What can another team copy? Test a personalised explanation of the price change against the current message, keep a holdout on the incumbent journey and judge the result on churn over the full renewal window. Compare more measured experiments in the [Signal Index](https://industrycontents.com/ic-lab/growth-signal-index/), browse the [Growth Signal archive](https://industrycontents.com/category/growth-signal/), or use the benchmark tools in the [IC Lab](https://industrycontents.com/ic-lab/).