AI for inventory management: what the data actually tells you
Most ecommerce businesses think they have an inventory problem. They are usually right, but for the wrong reasons. The problem they name is stockouts, overstock or margin erosion. The problem they have not yet named is that their decisions are running on data that is weeks old, manually processed and structurally unable to reflect what is actually happening in demand.
The conventional response is to buy a forecasting tool, tune it once and assume it is working. That assumption is expensive. It produces safety stock that masks the real issue, write-downs that appear in quarterly reviews but never trace back to a decision, and a planning cycle that is permanently one step behind.
This article will not tell you that AI will fix your inventory. It will tell you what the data inside your inventory operation actually reveals about where margin is leaking, which forecasting assumptions are wrong and what a well-designed system should be doing differently. If you are a marketing director, head of ecommerce or growth leader, this matters directly to you - because inventory availability and demand-side activity are the same problem, not separate ones.
The forecast is not wrong - it is answering the wrong question
Standard inventory forecasting asks: given historical sales, what will we sell next week? That is a useful question. It is not the right question.
Historical sales data tells you what customers bought when stock was available, at the price you were charging, with the promotions you were running. It does not tell you what demand actually was. If a SKU was out of stock for six days in a quarter, the sales data for that SKU is suppressed. A model trained on that data will underforecast. You will run out again. The suppression compounds.
This is not a hypothetical problem. Consider a mid-size direct-to-consumer apparel brand running seasonal campaigns. Their paid social drives a measurable spike in traffic to a hero product. The ecommerce team sees the spike. The demand planner sees a stock level. Neither is looking at the same signal at the same time. By the time the stockout occurs, the campaign has already spent its budget driving people to a page that cannot convert.
The data you need to answer the right question - what is true demand, adjusted for availability, price elasticity and promotional effect - already exists in most businesses. It lives across your ecommerce platform, your ad accounts, your CRM, your warehouse management system and your returns data. The issue is not data availability. It is that nobody has connected those sources into a single model that updates in something close to real time.
What the data actually reveals when you connect it properly
When organisations connect these sources - even imperfectly - the picture that emerges is usually a surprise.
The most consistent finding is that between 15 and 30 per cent of lost revenue in any quarter is attributable to demand-supply timing failures rather than actual demand weakness. A product with flat sales is not necessarily a product with flat demand. It may be a product that was never available when the demand signal was strongest.
A second finding is margin-related. Businesses that rely on end-of-season markdowns often find, when they model it properly, that the markdown was not caused by over-buying - it was caused by the wrong timing of replenishment on adjacent SKUs. The full-price sell-through on Product A collapsed not because demand dried up but because Product B, which shares a customer segment, was out of stock and customers substituted, then disengaged.
A third finding is about promotional effectiveness. Marketing teams calculate ROAS on the revenue generated. They rarely model the revenue not generated because inventory was not in position when the campaign fired. Attribution looks fine. The actual commercial outcome is materially worse than it appears.
When you run these analyses, the data does not just tell you that something went wrong. It tells you which decision, made at which point in the planning cycle, caused the problem - and how much it cost.
Where AI adds genuine signal and where it does not
AI does not add value uniformly across inventory management. Being precise about where it helps is more useful than the general claim that it improves forecasting accuracy.
AI models, specifically gradient-boosted trees and certain recurrent neural network architectures, outperform statistical methods meaningfully in three situations: when demand is driven by factors external to historical sales patterns (weather, trend cycles, competitor availability); when you have high SKU count and cannot manually tune a model per product; and when promotional calendars interact with demand in non-linear ways.
They do not reliably outperform simpler methods when your data is thin - fewer than 18 to 24 months of clean transaction history per SKU, or significant data gaps from stockout periods that have not been corrected for. Training a sophisticated model on corrupted input does not produce a more accurate output. It produces a more confident wrong answer.
The practical framework for deciding where to deploy AI forecasting looks like this:
- Audit your historical data for stockout periods and correct for suppressed demand before any model training
- Identify the SKUs or categories where external signal (search trend, social velocity, wholesale signal) meaningfully predicts demand ahead of sales data
- For those categories, test a model that ingests external signals against your current baseline - measure lift over 8 to 12 weeks before committing
- For stable, low-volatility SKUs, a well-tuned statistical model is cheaper and more interpretable - save the AI budget for where variance is high
- Connect your demand forecast to your marketing calendar explicitly, not as an afterthought - if your media spend is going up on Thursday, your demand model should know that on Wednesday
The last point is the one most businesses miss entirely.
The organisational gap that technology cannot fix on its own
There is a structural problem in most ecommerce businesses that no forecasting system resolves: the people responsible for generating demand and the people responsible for fulfilling it are not looking at the same data, and they are not making decisions in coordination.
Marketing plans campaigns. Buying plans inventory. They operate on different cycles, report into different functions and use different metrics. When AI forecasting is deployed, it is almost always deployed within the commercial or supply chain function. Marketing carries on planning campaigns against availability assumptions that are either stale or simply assumed.
A consumer electronics retailer in this position might run a Black Friday campaign that drives above-forecast demand on a flagship product. The inventory team had planned for a more conservative scenario. The product sells out within hours. The campaign continues to run for two more days before anyone intervenes - spending budget to drive traffic to a sold-out page, generating returns from oversold allocations and damaging the post-purchase experience for customers who did get through.
The fix is not a better forecast. The fix is a shared data environment where marketing sees inventory position and supply sees demand signal - and both update frequently enough to act before the problem compounds. This is what a connected business intelligence layer actually does in practice. It is not a dashboard. It is a single version of truth that changes decisions before they become costs.
Turning analysis into a decision system
Most AI implementations in inventory management stop at the forecast. The forecast is an input. What matters commercially is the decision downstream of it - how much to order, when to reorder, when to redirect demand and when to accept the markdown.
Building a decision system on top of a forecast requires three things. First, defined decision rules: at what stock cover level does reorder trigger, and does that rule vary by product velocity and lead time? Second, exception logic: which SKUs need human review and which can be automated? Third, feedback loops: does the system learn from decisions that turned out to be wrong, and does that learning update the model?
Without the third element, you have automation, not intelligence. The system will confidently repeat mistakes at speed.
For businesses at the scale Rodan typically works with - technology-led consumer businesses between £100m and £1.5bn in revenue - the right entry point is not a full-system rebuild. It is a structured diagnostic that maps your current data architecture against the decision points where margin is leaking, quantifies the cost of the gap and produces a clear picture of what needs to change and in what order.
That work takes two to three weeks. The output is specific enough to act on immediately, and scoped enough to build a business case for whatever comes next.
The cost of the current state is not theoretical
If your inventory decisions are running on weekly batch data, your forecast is trained on stockout-suppressed history, and your marketing and supply functions are not connected to a shared demand signal - you are not in a stable situation. You are in a situation that looks stable because the costs are distributed: across write-downs, suppressed conversion, wasted media spend and customer churn that never traces back to a stockout.
The question is not whether AI can improve inventory management. Applied correctly, in the right data environment, with connected decision logic, it demonstrably does. The question is whether your current architecture is capable of supporting it - and whether the problem has been diagnosed precisely enough to know where to start.
If you are not confident the answer to both is yes, a diagnostic engagement is the right first step. Speak to Rodan to scope one.
