What we've shipped elsewhere. What it would mean at Halfords.

MOT reminders were sent on a fixed 30/14/1-day cadence to every vehicle in the file. Reminder-driven bookings had flattened, garages in high-demand postcodes were saturated while others sat empty, and the retailer had no way to route customers to available slots.

• MOT booking survival model
  Deep survival network on 4 years of registration, service and booking data — predicts the day booking probability peaks per vehicle.
• Garage capacity solver
  Constraint solver matching predicted demand to per-postcode Autocentre slots on a rolling 21-day horizon.
• Uplift-matched reminder
  Causal model chooses between reminder-only, discount, or free-check offer — only where each is incremental.

10% randomized holdout kept on the legacy 30/14/1 calendar. Treatment cohort received model-timed reminders with slot-aware routing. Readout on booked MOTs and attach revenue.

The retailer retired its calendar-rule reminder engine. Load moved off saturated garages onto quieter ones, colleagues reported fewer 'we're full' conversations, and the P&L saw the attach lift as pure margin.

The membership base was mature; churn was creeping up post-renewal price change. The save-desk was discounting every at-risk member with a blanket 20% offer — retaining some, but destroying margin on members who would have stayed anyway.

• Survival churn model
  Time-to-churn network on visit frequency, benefit utilisation, service history and app engagement — with SHAP reason codes surfaced to agents.
• Uplift save-matching
  Causal forest matching each at-risk member to one of: no-offer, benefit-nudge, service credit, tier downgrade, discount. Ranked by net expected value.
• Contact suppression layer
  Members with negative net EV (retention offer would cost more than the churn) suppressed from outreach entirely.

15% randomized holdout worked with the legacy blanket-discount playbook. Treatment cohort worked the ranked save queue with matched offers.

The retention team shrunk the discount envelope by half while retaining more members. Finance now models subscription revenue with a probabilistic churn curve rather than a linear roll-off.

The retailer sold a bike, then lost the customer to the internet or the local independent for servicing. Only 18% of bike buyers came back for the 6- or 12-month service — a huge margin hole given how profitable Bike Care hours are.

• Service booking propensity
  Gradient-boosted model on bike SKU, ride season, weather, distance to store and prior visits.
• E-bike battery health signal
  Rules-augmented model on ride telemetry and battery age for the £2k+ e-bike segment — early flag before failure.
• Local-store routing
  Match predicted demand to store-level Bike Care capacity on a rolling 4-week window.

20% randomized holdout received the legacy 'come back in 6 months' generic email. Treatment cohort got model-timed outbound to the local store with pre-slotted appointment.

Bike Care moved from a break-even service to one of the highest-margin lines in the retailer's P&L. The 12-month service reactivation rate more than doubled.

42 concurrent offer variants across email, app and outbound. Product managers argued for their own audiences; the CRM had no way to arbitrate. Response rates flat at 1.6%, unsubscribes climbing.

• Product-level propensity (12 products)
  LightGBM ensembles with product-affinity embeddings and monotonic constraints on tenure and balance.
• Two-model uplift
  Causal treatment-effect estimation using an X-learner over 18 months of randomized holdout exposures.
• Decisioning layer
  Real-time ranker with contact cap of 2/week, channel affinity, and margin-weighted expected value.

10% end-to-end randomized holdout across all channels. Weekly incrementality readout on revenue per contactable customer.

The bank retired 42 campaigns and consolidated onto one ranked feed into the CRM. Marketing headcount reallocated from campaign ops to experimentation.