The Price Is (Almost) Right: Why Retailers Need More Than Deep Learning
By Jyothsna Santosh
Pricing in retail isn’t just a science; it’s a high-stakes strategy conversation. It sits at the intersection of business judgment, customer behavior, and machine intelligence. With portfolios that include hundreds of thousands of SKUs, even a 1% price shift can translate into billions in impact.
It’s no surprise that retailers are turning to AI – especially deep learning – to find the “right price.” But there’s a catch: deep learning is powerful, not all-knowing. On its own, it cannot carry the full weight of pricing strategy.
In my work building data-driven pricing systems, I’ve seen that the right price is not just about model accuracy. It’s about organizational adoption, interpretability, portfolio thinking, and disciplined experimentation. Without those, even the smartest models fall short.
Where Deep Learning Helps – and Where It Doesn’t
Deep learning models can detect subtle patterns in historical sales, competitor prices, inventory trends, and customer behavior. They are particularly good at:
- Modeling non-linear demand shifts
- Handling complex seasonality
- Learning cross-SKU substitution and complementarity effects
But pricing is not a single-output problem. A model might recommend lowering the price of a mid-range item – one that sits between entry-level and premium options – to drive higher volume. On paper, that looks like a win.
The real questions are more nuanced:
- What happens to the lower-priced alternative sitting just below it?
- How does this affect the higher-margin item just above it?
- What is the net impact across the shelf, not just on a single SKU?
- Who owns the downstream consequences on category margin and inventory?
These intra-portfolio shifts can quietly erode profitability if they are not modeled and monitored holistically. The question isn’t just “Will this price sell more?” but “What happens to the entire portfolio, across channels and time?”
What Modern Pricing Systems Really Need
To truly support decision-making, pricing systems need more than predictions; they need perspective. The most effective implementations combine analytical rigor with organizational context and explainability. Here are some of the components that matter most.
1. Causal Inference, Not Just Correlation
Price recommendations should account for why volume changes, not just that it did. Causal machine learning methods help isolate the impact of a price change from confounding factors like promotions, competitor moves, or macro shocks.
That distinction matters. A model that merely “sees” that volume went up after a price change may over-attribute causality, leading to decisions that don’t hold when the context shifts.
2. Contextual Bandits for In-Market Learning
Instead of deploying one price and waiting weeks or months for results, contextual bandits allow retailers to:
- Test multiple price points live in the market
- Learn in real time which prices perform best for different contexts
- Converge quickly on the highest-value options with less risk
This bridges the gap between “offline modeling” and “online learning,” making experimentation a built-in part of the pricing engine rather than a separate, episodic effort.
3. Agentic AI for Scenario Simulation
Agent-based and agentic AI systems can simulate pricing actions before they go live. Strategy teams can ask questions like:
- “What happens to category margin if we move SKU X by 5% across 200 stores?”
- “How do tariffs or cost shocks ripple through price ladders and bundles?”
An agentic system can run those scenarios across SKUs, stores, and channels – providing directional insight before decisions hit the shelf. It gives pricing and merchandising teams power before execution.
4. Explainability That Builds Trust
A recommendation is only as good as the stakeholder’s belief in it. That’s why explainability is not a “nice to have”; it is a prerequisite for adoption.
We have built tools that translate model outputs into human-readable narratives, such as:
- “This price is expected to increase sales by 12% without cannibalizing adjacent SKUs.”
- “Elasticity here is driven primarily by local competition and promotion sensitivity.”
When pricing, merchandising, and finance teams understand why a model recommends a price, they are far more likely to use it – and to flag when the recommendation misaligns with on-the-ground realities.
5. Portfolio and Channel Awareness
No SKU is an island. Pricing models must account for:
- Product families and price ladders
- Substitute and complementary relationships
- Channel alignment (in-store, online, marketplace, and delivery)
A change made on the physical shelf needs to be reflected consistently in the digital cart. Disconnected price logic across channels doesn’t just erode trust; it creates friction right at the point of purchase.
When Simple Models Are Not Enough
Many traditional pricing models estimate elasticity using log-transformed relationships between a product’s price, selected features, and some category-level grouping. These OLS-based models are:
- Easy to interpret
- Qu