ca 30 min walkthrough
Take-home · for CRM & Lifecycle candidates

Customers are slipping away at month twelve. What do you do?

We care about your thinking — not the production value. Read the scenario, work through the four prompts below, then walk us through your approach.

The scenario

Tibber has ~550,000 customers across SE, NO, DE, NL. We see strong engagement in the first 90 days — app installs, adoption, smart charging setup — but a measurable churn spike between months 12-18, especially in DE and NL where the product is younger.

~550k customers across SE · NO · DE · NL
90 days strong onboarding engagement
12-18 mo measurable churn spike
DE · NL newer markets, higher loss

What you have to work with

Braze

Multi-channel orchestration — email, push, in-app, SMS, Canvas flows, Liquid, Connected Content.

Databricks CDP

Customer data layer with the warehouse, event streams and the modelling environment.

Churn propensity score

Refreshed weekly. Available as a user attribute for segmentation and Liquid.

AI toolkit

LLM-based content generation and agentic workflows that operate without human intervention.

Your task · four parts

Prepare a 30-minute walkthrough.

Click any card to expand
01

Diagnosis

Before you build anything

What would you want to know before building anything? Name the 3-5 questions you'd answer first, and how.

We're listening for

  • You don't accept the brief at face value — "churn at 12-18 months" is a symptom, not a diagnosis.
  • You can tell us how you'd answer each question, not just what you'd ask.
  • You think about the propensity score's calibration on this specific cohort, not just its existence.
  • You consider market-level context (regulatory, competitive, meter rollout) — not just behavioural data.
02

A Braze Canvas sketch

One journey, whiteboard-level

Design one journey that addresses this churn risk. Whiteboard-level is fine: entry trigger, key branches, channels, timing, exit criteria, and what data points drive the splits. Be specific about what's a segment filter vs. a Liquid personalization vs. an API-triggered event.

We're listening for

  • You know when to compute in Databricks vs. in Braze — and why the propensity score isn't a segment recompute job.
  • You can name the difference between a segment filter, Liquid personalization, and a Currents/API event — and pick the right one for the job.
  • Your branches map to different churn drivers, not just different message frequencies.
  • You have real exit criteria, not just timeouts. What's the win condition?
  • You think about channel cost and consent — SMS in DE isn't free or unregulated.
03

The agentic layer

One concrete example

Where in this journey would you replace human or rule-based decisions with an AI agent? Pick one concrete example and describe what the agent does, what it has access to, and how you'd keep it safe.

We're listening for

  • You pick one place — depth over breadth. Where in the journey, why there, and what changes.
  • You can name what the agent reads, writes, and is forbidden to touch.
  • Safety isn't a list of vibes — it's deterministic guardrails, evals, human-in-the-loop defaults, scoped credentials, a kill switch.
  • You think about fallback behaviour when the agent fails or is uncertain.
  • Bonus: you can describe how you'd measure whether the agent is actually doing the right thing.
04

Measurement

The one number you'd commit to

What's the one KPI you'd hold yourself accountable to, and what's the leading indicator you'd watch weekly?

We're listening for

  • Your headline KPI is hard to game — controlled, attributable, not retargetable by shrinking the segment.
  • The leading indicator is observable in days, not quarters, and closely tied to the eventual outcome.
  • You've thought about holdouts or quasi-experimental design, not just before/after.
  • You can explain what you'd do if the leading indicator drops — i.e. it's actually decision-relevant.

How we'll spend the time

Roughly 30 minutes for you to walk us through it, then a conversation. Present it however you like.

Expect us to push on the parts where the answer is non-obvious — especially the diagnosis and the agent's failure modes. You don't need to answer everything perfectly; tell us where you're guessing.