Client Health Score Model

The Challenge

During COVID-19, PepsiCo's commercial team was managing 69 key account negotiations simultaneously while operating with reduced headcount and compressed timelines. Account managers had no systematic way to prioritise which clients needed urgent attention — they were relying on gut feel and reactive check-ins rather than forward-looking signals.

The risk was significant: a single lapsed key account contract could represent millions in annual revenue, and the commercial team had no early warning infrastructure in place to prevent it.

Approach

I designed and built a composite health score that drew on three categories of signals available in our internal data:

• Sales trajectory: Rolling 12-week sales trend relative to prior-year baseline, flagging accounts with sustained volume decline.
• Engagement frequency: Recency and cadence of account manager touchpoints logged in the CRM, used as a proxy for relationship health.
• Contract tenure signals: Time since last renewal, upcoming expiry windows, and historical renewal lead times by account tier.

Each signal was normalised and weighted using a logistic regression model trained on historical account outcomes. The output was a 0–100 score per account, with RAG-status banding (Red / Amber / Green) surfaced in a Power BI dashboard updated weekly.

I worked closely with the commercial team to validate that the score aligned with their expert judgement before deployment — this buy-in was essential to ensuring account managers actually acted on the outputs rather than ignoring them.

Key Results

The model changed the team's workflow from reactive to proactive. Account managers knew each Monday which accounts had deteriorated in score, allowing them to schedule calls and tailor retention offers before problems became visible at the contract level.

Lessons Learned

• Stakeholder trust in a model's outputs matters as much as the model's accuracy. Spending time validating the score against expert intuition before launch was the most important investment.
• Simple, interpretable models outperform complex ones when the output needs to drive human action. A logistic regression with clear feature weights was more effective than a black-box ensemble.
• Operationalising the model into a weekly Power BI refresh — rather than a one-off analysis — was what turned it from a project into a lasting business tool.