Recurve AI

The flagship Winter Appeal generated significant surplus, but the targeting approach had poor efficiency: pockets of unresponsive supporters were scattered throughout even high-performing segments, driving up costs unnecessarily. Rising costs meant every pound of fundraising spend needed to work harder.

The organisation needed a targeting approach that would provide better accuracy, and turned to Adroit’s Recurve ai tool — a modelling pipeline that continuously relearns rather than stagnating over time.

The culprit

The charity was using complex manual selections together with a 2016 model, built on behavioural data that was over a decade old — ancient history in donor behaviour terms that hadn’t kept pace with how people’s engagement had changed after the global pandemic and economic crisis.

As a consequence, the charity was spending precious budget on people who weren’t going to respond, buried within segments that looked promising on paper.

The selection process itself consumed significant time; creating campaign selections meant building multiple bespoke segments that needed multiple adjustments to hit the required campaign volumes.

The charity needed a better way to rank supporter likelihood to respond — one that considered multiple behavioural features simultaneously, adapted to recent trends, and simplified the selection process. The old approach was still profitable, but it was leaving money on the table.

Adroit introduced the organisation to Recurve AI, its proprietary AI scoring tool. But the real value wasn’t just another predictive model — it was a modelling pipeline designed to prevent the exact model decay problem that had plagued the previous model.

Recurve ai continuously relearns from fresh campaign results, adapting to changing patterns without the cost and delay of periodic model rebuilds.

Technical foundation

Adroit loaded 8 million supporter records into the platform. The training dataset drew from multiple years of Winter Appeal mailings and response outcomes, giving the algorithms substantial behavioural history to learn from.

Feature Engineering

Feature engineering went well beyond simple RFV segments. The platform generated features across Traditional Donations, specific Appeals (Winter and Summer), Lottery Sales and Donations, and Regular Giving products. Demographic variables included Age, Gender, Title and Mosaic classifications.

Behavioural features captured Gift Aid Declarations and Retail Gift Aid Donations. Each supporter became a rich profile of giving patterns, not just a single RFV bucket.

Multiple Algorithms

The modelling stage tested multiple propensity algorithms head-to-head across different training time periods.

This wasn’t about picking a favourite algorithm — it was about empirically identifying which combination of technique and training window produced the most predictive model for the charity’s specific supporter base.

The winning model surfaced a critical insight: the most recent campaign data proved far more predictive than longer historical periods.

This validated the model decay concerns and confirmed that recent donor behaviour patterns mattered more than distant history. Supporters were changing, and the model needed to change with them.

Scoring 8 million records created a technical challenge. Recurve ai handled this through data chunking — processing supporters in manageable batches rather than attempting one massive pass that would create processing bottlenecks.

Implementation

Implementation followed a rigorous testing methodology. The charity ran a 50/50 A/B test with half the available audience selected via Recurve ai model scores, and half via business-as-usual RFV segments.

This ensured a valid, apples-to-apples comparison of the two approaches on identical campaign timing and creative.

The model scores fed directly back into the existing marketing platform (Apteco). No workflow disruption, no new systems for the fundraising team to learn — just enhanced targeting intelligence flowing into familiar tools.

The process simplification proved as valuable as the predictive lift. Previously, building a Winter Appeal audience meant constructing multiple bespoke segments, testing different combinations, adjusting to hit volume targets, and iterating through back-and-forth revisions.

With Recurve ai, selection became a single decision: set the model score threshold that balanced volume and predicted response.

One selection replaced the entire segmentation puzzle.

The streamlined approach freed fundraising team capacity while simultaneously improving targeting precision. Less time planning segments, more time optimising creative and strategy.

And because the pipeline relearns after each campaign, the next Winter Appeal starts with a model already trained on the most recent supporter behaviour — no decay, no delays, no rebuild costs.

The head-to-head test delivered clear results:

1  8.7% response uplift

The Recurve ai model generated significantly higher response rates than the charity’s business-as-usual segments — an uplift in response rate of 8.7%.

2  £6K from a hidden segment

The model identified a segment of lapsed donors who’d fallen completely outside the charity’s traditional RFV segmentation framework. These supporters had shown responsiveness to multiple products in the recent past — appeals, lottery, perhaps regular giving — but had since gone quiet. The BAU approach missed them entirely because they didn’t fit neatly into the 0–12 month or 13–24 month buckets that defined manual targeting. That hidden segment alone generated £6K in additional net income from the single Winter Appeal campaign, not counting the future lifetime value of reactivating previously lapsed donors.

3  25% further uplift available

Post-campaign analysis revealed the model could have extended even further down the propensity scores to capture an additional 25% net income uplift beyond the original projection.

Beyond the immediate financial impact, the modelling pipeline transformed the charity’s campaign planning process.

What previously required building multiple bespoke segments, endless back-and-forth adjustments to hit volume targets, and manual reconciliation now happens through a single selection based on model scores. Campaign selection time dropped dramatically.

Model value

The platform’s value extends well beyond the Winter Appeal. The charity has since deployed the modelling pipeline for its Summer Appeal variant and built a Legacy Response Model using the same infrastructure.

The system easily extends to reactivation campaigns, cross-sell initiatives, and email responsiveness predictions.

Valuable insights

The pipeline relearns automatically for each subsequent campaign, eliminating the model decay problem that plagued the charity’s 2016 approach.

No periodic rebuild costs, no gradual erosion of predictive power, no manual intervention required.

The model stays current with donor behaviour and campaign evolution automatically, creating compound efficiency gains across the entire donor engagement programme.