ML capability framework, roadmap, and pilots delivered
Team upskilled through co-delivery
Pilot campaign demonstrated positive ROI vs control group
Propensity models for donation, sponsorship, and segmentation deployed
Delivered a prioritised AI roadmap, ML capability framework, and live multi-channel marketing pilot for a large UK healthcare charity. The pilot used propensity models, NLP-derived features, and a custom allocation algorithm to demonstrate measurable fundraising uplift against a randomised control group, providing the evidence base for a data-led supporter engagement strategy.
The Problem
A large UK healthcare charity had stated an ambition to become data-led. Senior leadership understood that supporter data could transform fundraising and engagement, but intent had not translated into action. Multiple technology and data projects were running with no shared priorities and no assessment of where AI could add the most value. Supporter communications still followed broad, static segmentation: donors grouped by giving band, with little use of behavioural signals or predictive scoring. The team had limited experience with machine learning in production, and there was no clarity on model ownership, refresh cadences, or how predictions would feed into operational systems.
The Solution
We delivered three interconnected workstreams within a single engagement: a strategic assessment, a capability uplift plan, and a live marketing pilot.
Strategic assessment. We reviewed the charity's data estate, technology stack, and governance, then mapped AI use cases across fundraising, marketing, supporter experience, and operations. Each was scored on expected value and delivery feasibility in collaboration with leadership. The output was a tiered roadmap: high-value, high-feasibility use cases first, followed by waves requiring platform or data investments.
Capability plan. We defined the skills, roles, and operating rhythms needed to sustain ML delivery beyond the engagement: model ownership, refresh cadences, approval processes, monitoring, and integration from Snowflake into Salesforce campaign lists. The plan included a skills development pathway combining structured training with pair-working on the pilot, so learning was tied to a real deliverable.
Marketing ML pilot. We built propensity models for donation and sponsorship, reflecting the different decision dynamics behind each giving type. NLP techniques extracted sentiment, topic, and engagement intensity features from supporter messages, enriching the models where behavioural data alone was sparse (newer supporters, or those with infrequent but emotionally driven giving). A custom allocation algorithm optimised total expected incremental value across channels, subject to budget and capacity constraints, replacing the previous approach of sending everything to everyone through the usual channel. The entire campaign was designed around proper test-and-control measurement from the outset, with randomised holdout groups and pre-specified success metrics.
Results and Impact
| Metric | Value |
|---|---|
| Data strategy contribution target | Supporting an ambition of approximately GBP 90M additional contribution over eight years |
| Pilot campaign ROI | Strongly positive: model-targeted campaign outperformed control on incremental revenue per supporter |
| Propensity models delivered | Donor propensity, sponsor propensity, supporter segmentation |
| NLP features extracted | Sentiment, topic, and engagement intensity from supporter messages |
| Capability framework | Model ownership, refresh, approval, monitoring, and Salesforce integration defined |
| Measurement standard | Randomised control groups embedded in campaign design from the outset |
The GBP 90M figure represents the charity's own long-range ambition for additional contribution enabled by a data-driven approach. The strategy, roadmap, and capability plan were designed to underpin that target. The pilot demonstrated that the first wave of ML use cases could deliver measurable uplift, providing the evidence base for continued investment.
Key Takeaways
-
Measurement designed upfront is what made the pilot credible. Specifying control groups, success metrics, and required sample sizes before the campaign launched meant the results stood up to scrutiny from fundraising leadership and the board. ML pilots that skip experimental design produce results nobody trusts.
-
Capability uplift has to answer the boring questions. Teaching a team to build models is the easy part. The harder part is answering: who owns this model in six months? What happens when prediction quality degrades? How do scores get from Snowflake into Salesforce? If these questions are left unanswered, the models stop working within a quarter of handover.
-
Strategy and proof of value have to be delivered together. A roadmap without a pilot is a document that gets filed. A pilot without a roadmap is a one-off experiment with no path to scale. Delivering both in a single engagement meant the charity had a plan it believed in and a pilot it could build on.