Case Studies
Every case below began with a plan that would have failed. The constraint was upstream — in the problem formulation, not the model.
Wrong model class selected — the prior failure was not a data or talent failure. It was a model-selection problem.
The Situation
C-suite plan to roll out sales forecasting across all subsidiaries. An earlier forecasting effort had already failed in production. The organization attributed the failure to data quality and brought in a new team.
What GIDS Found
A model-selection problem. The prior failure had nothing to do with data quality or team capability. The wrong model class had been chosen for the forecasting problem itself — a fundamental mismatch between the problem structure and the modeling approach.
What Changed
The initiative was re-grounded on deep learning time-series modeling chosen to fit the actual problem structure, then rolled out to subsidiaries with the correct architecture in place.
Outcome
Sales forecasting accuracy improved by 33% against the benchmark set by the prior failed model. The corrected architecture was rolled out across 4 subsidiaries. The organization did not replace its data team. It replaced its problem definition.
Structured data only — an oversimplification that left a direct P&L insight source completely untouched.
The Situation
The organization treated AI as applicable to structured data only and resisted going further. Leadership believed unstructured sources were too complex and too speculative to justify.
What GIDS Found
An oversimplification. The data science of data transformations could open systematic insight from trader user files — tied directly to the department's profit and loss. The data existed. The methodology existed. The connection had never been made.
What Changed
User-file data was transformed and evaluated against the department P&L, turning an untouched data source into a live decision input.
Outcome
The methodology was adopted. Trader file data, previously untouched, became a structured input to departmental P&L analysis. A data source that had existed for 3 years without being evaluated was generating systematic signals across 4 trading desks. The data was always there. The question had never been asked.
"We told a client to spend less on AI." A right-sizing problem — a leaner solution was the better strategy.
The Situation
The company was about to invest in an expensive LLM solution for campaign recommendations. The vendor had been selected. The contract was being reviewed.
What GIDS Found
A right-sizing problem. The LLM solution was technically capable but fundamentally over-engineered for the actual recommendation task. A simpler, cheaper alternative would produce better results for this specific use case — and would be maintainable by the existing team.
What Changed
The organization pivoted to a leaner agentic recommendation engine for marketing campaigns. The vendor contract was not signed. This is the GIDS case that matters most: we told a client to spend less. That only happens when the diagnosis comes before the decision.
Outcome
The leaner recommendation engine went live across 8 campaigns within 2 months of the vendor contract being declined. Campaign recommendation performance exceeded the projected benchmarks the LLM vendor had proposed — at a fraction of the cost. The existing team maintained it without outside support.
A robust data science team focused entirely on statistical forecasting — and missing the unsupervised layer that would have made it work.
The Situation
A global retail company had a strong data science team running statistical forecasting models. The models were technically sound. The results were underwhelming. The team was being asked to do more with the same approach.
What GIDS Found
A skill gap in descriptive analytics. The unsupervised-to-supervised workflow was missing entirely. The team was skipping the step that would have told them who their customers actually were before trying to predict what they would do.
What Changed
Unsupervised learning was used to categorize customers by shopping habits, feeding segmentation into the forecasting practice — strengthening it, not competing with it. The two disciplines became complementary rather than parallel.
Outcome
The unsupervised layer produced 7 distinct customer segments that had not previously existed in the organization's modeling architecture. When segmentation fed into the forecasting pipeline, forecast accuracy improved by 20%. The data science team did not change. The sequence of what they did with it did.
Additional cases from John's ongoing advisory practice are published weekly. Subscribe on Substack to receive them as they publish.
The upstream problem takes a different form in every organization. The diagnostic is the same.
The first conversation is a diagnosis, not a pitch.