AI Escape assisted with sales-data analysis across more than 1,000 promotional product SKUs.
Back to success stories
Promotional product analytics
TD Bank
AI Escape assisted TD Bank in analyzing 1,000+ promotional product SKUs from 1-2 years of disorganized, unstructured sales records across many different file formats, using AI and statistical analysis to normalize categories, identify slow-selling SKUs, forecast likely sales trajectories, estimate inventory drawdown, compare counterfactual strategies, and backtest error estimates.
Inventory intelligence
A sales-history mess turned into forecasting and decision support.
The TD Bank work started with uneven promotional-product sales data, then used AI-assisted normalization and statistical modeling to make category, forecast, drawdown, and error-estimate questions easier to reason through.
The analysis covered 1-2 years of disorganized, unstructured sales records across many different file formats.
AI and statistical analysis supported category normalization, slow-selling SKU identification, likely sales trajectories, and inventory drawdown estimates.
Older data was used to backtest the techniques and provide error estimates around forecast behavior.
From disorganized SKU history to modeled inventory choices.
What changed
Promotional product history is hard to forecast before it is organized
TD Bank had 1-2 years of disorganized, unstructured sales records across many different file formats for 1,000+ promotional product SKUs. Before any useful forecasting could happen, the product categories needed to be normalized and organized so each SKU could be compared inside a cleaner category structure.
AI Escape assisted with that analysis using AI and statistical techniques: normalizing and organizing product categories, identifying slow-selling SKUs, forecasting likely sales trajectories, estimating inventory drawdown, and backtesting techniques on older data to provide error estimates.
Category normalization map
Input files
- Sales sheets
- SKU exports
- Unstructured records
Normalized categories
- Category A
- Category B
- Category C
SKU states
- Steady priorbaseline movement
- Slow sellerseparate attention
- Drawdown watchinventory timing
What AI Escape helped analyze
The work was not a generic forecasting dashboard. It started with the unstructured sales-history problem: many file formats, more than 1,000 promotional product SKUs, and enough category disorder that sales trajectories and inventory drawdown needed a cleaner base before the forecasts could be trusted.
Max’s prior research on counterfactuals helped frame hypothetical outcomes under adjustments to sales strategies and environmental shifts. Markov-sampling techniques with AI-powered priors supported the sales forecasts, while backtesting on older data produced error estimates instead of false certainty.
Forecast discipline
Model inputs
- Older data backtesttest the technique before relying on the estimate
- AI-powered priorguide Markov sampling without claiming certainty
- Counterfactual framecompare adjusted strategies and environmental shifts
Forecast bandLikely trajectoryError estimate
Inventory drawdownWatch timing
drawdown timing signal
The useful output was a decision-support frame: category-normalized SKUs, slow-seller identification, likely sales trajectories, inventory drawdown estimates, counterfactual comparisons, Markov-sampling with AI-powered priors, and backtested error estimates that made the forecast limits visible.
Analytical architecture
Category cleanup, forecast priors, counterfactuals, and backtests connected into one analysis path.
The work combined practical data organization with modeling discipline: normalize product categories first, identify slow sellers, model likely trajectories, estimate inventory drawdown, compare hypothetical strategy changes, and backtest against older sales history.
Unstructured product history became organized categories.
The first step was normalizing and organizing promotional-product categories so scattered SKU sales records could support analysis instead of staying as disconnected file fragments.
Slow-selling SKUs could be separated from broader category noise.
AI and statistical analysis helped identify products whose sales behavior needed separate attention before forecast and drawdown estimates were considered.
Likely sales trajectories and inventory drawdown estimates were modeled together.
The work paired trajectory forecasts with drawdown estimates so the analysis could reason about how inventory might move over time.
Hypothetical strategies and Markov-sampling priors were tested against older data.
Max's prior counterfactual research informed hypothetical sales-strategy and environmental-shift comparisons, while Markov-sampling techniques with AI-powered priors were backtested on older data for error estimates.
How the analysis moved
A sales file could move from many formats to forecasts and error estimates.
The operating path stayed grounded in the supplied data: clean the product categories, separate slow-moving SKUs, model sales trajectories, estimate drawdown, compare counterfactual changes, and test the technique on older data.
Bring many sales-file formats into one analysis path.
The starting point was 1-2 years of disorganized, unstructured sales records across many different file formats.
Organize product categories before forecasting.
AI-assisted category normalization made the 1,000+ promotional product SKUs easier to compare and reason about.
Identify slow-selling SKUs that needed separate attention.
The analysis separated slow-moving products from broader category behavior before estimating sales trajectories and drawdown.
Estimate likely trajectories and inventory drawdown.
Statistical modeling and AI-powered priors supported forecasts of likely sales trajectories and inventory drawdown estimates.
Use counterfactuals and backtests to bound the estimates.
Counterfactual comparisons imagined hypothetical outcomes under sales-strategy or environmental shifts, then older data backtests produced error estimates.
The modeling tools inside the work.
Analysis methods
- AI and statistical analysis Used for category normalization, slow-selling SKU identification, likely sales trajectories, and drawdown estimates.
- Counterfactual research Used to imagine hypothetical outcomes under adjusted sales strategies and environmental shifts.
- Markov-sampling priors AI-powered priors informed the sampling approach for sales forecasts.
- Backtesting Older data was used to test the techniques and provide error estimates.
After the case study
Bring one operating workflow to the table.
If the TD Bank story feels close to a workflow you want to move, book a working conversation with AI Escape. We will use the case study you just read as the starting point and map the first concrete system worth building.