What the CFO sees at month-end
The close comes in late, and when it does it carries a freight line that has grown with no obvious cause. There were expedited shipments to cover stockouts, trucks that came back half empty and a warehouse holding months of cover on items that stopped moving. All three lines share one root cause: a demand forecast that missed weeks earlier, in a spreadsheet someone updates on Mondays.
That is why, in a company that buys, stores and distributes, the place where AI moves the P&L has little to do with drafting emails. It sits in the predictive models that have been solving this problem for years: forecasting from sales history, reading seasonality, picking up market signals such as promotions, prices or weather, tracking turnover by SKU and optimizing stock and routes. It is AI that rarely makes headlines, and it is the kind a CFO can audit line by line.
So how much is a better forecast worth in hard currency? Some studies put numbers on it, and they are worth reading with an auditor's pencil before they reach the board.
What the widely quoted study actually measured
The figure that shows up in plenty of vendor decks comes from a 2025 paper in the Journal of Business and Entrepreneurship by three researchers at the Quevedo State Technical University in Ecuador. They compared three AI models (recurrent networks, a Transformer and gradient boosting) with two classical statistical methods, ARIMA and exponential smoothing, on 24 months of weekly demand from a Latin American retail chain, enriched with inflation, exchange rates, weather, seasonality and promotions.
Forecast error. The Transformer reached a mean absolute percentage error (MAPE) of 3.9% and gradient boosting 4.5%, against 7.2% for ARIMA and 8.1% for exponential smoothing. That is where "up to 50% lower error" comes from: the gap between the best model and the weakest classical method.
Inventory cost. They simulated six months of running a distribution center on each method's forecasts. Holding plus stockout cost fell from $78,100 to roughly $53,500, a 31.5% reduction, driven mostly by less overstock, and their attribution analysis ranked promotions and seasonality as the variables that mattered most.
Empty miles. An LSTM model run on a fleet's GPS data, with demand and weather as inputs, cut empty kilometers from 1,200 to 1,020, a 15% drop, while emissions went from 4.8 to 4.1 tonnes of CO₂.
These are plausible numbers and the study is useful. It also has limits an executive committee should understand before using it as an argument.
Four reasons not to put 31.5% in the budget
It is one company, and it is a simulation. The inventory saving comes from replaying six months with each model's forecasts; nobody changed replenishment in a real warehouse for half a year to check. A well-built simulation gives you the order of magnitude, and only live operations confirm the realized saving.
The validation may flatter complex models. The authors used five-fold cross-validation. With time series, if the folds ignore chronological order, the model trains on weeks that come after the ones it later predicts, and the measured error ends up lower than what you will see in production. The paper does not say how the folds were built, and that doubt alone is enough to temper expectations.
The text has inconsistencies. The conclusions cite a forecast error reduction "of up to 50% (45% vs. 38.1%)", a comparison that matches none of the paper's own tables. The tabulated results still stand, though the sentence suggests a light editorial review.
The best model is only tested against the classics. The statistical tests confirm the Transformer beats ARIMA and exponential smoothing, but they report nothing on whether it beats gradient boosting, which landed six tenths of a point behind and is far cheaper to train, explain and maintain.
One more figure from the study deserves attention: model accuracy explained 68% of the variance in cost. The rest depends on things no algorithm changes by itself, such as the replenishment policy, lot sizes, real supplier lead times and the freight contract.
"A model that forecasts better saves nothing while the buying policy stays the same. The saving shows up the day someone signs off on a new reorder point."
Where the money is: the AI that doesn't write
For three years, AI conversations in executive committees have been dominated by generative models, which is understandable because they are the ones you can see. The AI that goes after logistics cost is older and less glamorous: predictive models trained on your own history, which learn how each product family behaves in each channel, which weeks break the pattern and which outside signals give advance warning.
That AI works on four levers a CFO recognizes in the P&L.
Seasonality and calendar. Campaigns, local holidays, customer shutdowns and the end-of-quarter bump in B2B orders. Classical methods capture regular seasonality, and machine learning models also pick up the interactions, such as a promotion that lands on a long weekend.
Market signals. Your prices and your competitors', promotions, weather and sector indicators. In the Ecuadorian study, promotions and seasonality carried the most weight in the inventory saving.
Turnover by SKU. Separating the items you can forecast well from the ones you can't matters because the right safety stock for an erratic SKU is very different from that of a stable one. Applying one rule across the whole catalog is how warehouses fill up with what doesn't sell.
Stock and route optimization. With a reliable forecast and its error band, reorder points and safety stock are set SKU by SKU according to the service level the committee decides to pay for, and load and route planning runs on expected demand instead of last week's.
Generative AI has a place in this setup too, as the interface for querying the data and understanding the numbers. The next section covers it.
How to design a project you can measure
The pattern that works for a mid-sized company that buys, stores and delivers has five parts, and the sequence matters more than the technology you pick.
A baseline before touching anything. Current forecast error by product family, holding and obsolescence cost, working capital tied up in stock, stockouts with their estimated lost sales, expedited shipments and empty miles. If you don't measure first, there will be no honest way to say afterwards how much you saved.
A simple method as the bar to clear. The first model competes against the same week last year or against exponential smoothing, and it only goes live if it beats them consistently on weeks it has never seen. A well-fed gradient boosting model is often enough, and it is easier to explain to the purchasing team than a neural network.
A forecast wired to a decision. The model produces a number per SKU and week with its confidence interval, and that number feeds the reorder point, the safety stock and the load plan. If the output lives in a dashboard someone checks when they have a spare minute, the saving depends on that person's calendar.
A conversational assistant over structured data. This is where the generative model comes in, as a query layer: the operations director asks why cost per order in the northern region rose in August, and the assistant turns the question into a query against the data warehouse, returns the figure and shows where it came from. The design rule is strict: numbers always come from a query, never from the model's prose, and definitions (what counts as an order, what goes into cost per order) live in a semantic layer the assistant cannot reinterpret.
Measurement with a control group. Some warehouses, families or routes run on the new system while others stay on the current method for an agreed period. Finance validates the comparison using the same allocation rules it applies at month-end, and the cost side includes what it takes to run the system: platform, generative model usage, retraining and the hours of the team that maintains it.
What rarely works the first time
The stumbling blocks in these projects repeat so reliably that they belong in the budget from day one. The study's own authors name data quality and technical training as the main challenges, and practice fills in the detail.
History records sales and hides demand. When an item was out of stock for three weeks, sales for those weeks were zero even with customers waiting, and a model trained on that learns to under-order exactly where stockouts happened. Correcting that censored demand is mandatory groundwork.
The item master can't take the scrutiny. SKUs that changed codes, mixed units of measure and substitute products that cannibalize each other make the model blend histories that have nothing in common until someone cleans them up.
Promotions were never recorded. If the commercial calendar lives in emails and slide decks, the variable that mattered most in the study doesn't exist for the model, and rebuilding it retroactively takes weeks.
The forecast improves and stock doesn't fall. This is the most expensive failure and the most common one: the model gets more accurate, but planners override it out of caution and the replenishment policy stays as it was. You need an agreed rule for when overrides are allowed, and a log of every override so you can tell whether it helped or hurt.
Empty miles depend on contracts. A better forecast lets you consolidate loads, but empty backhauls also depend on carrier agreements and on whether there is return freight at all. Without renegotiating, the routing saving stays at a fraction of the theoretical one.
None of these problems gets solved by a better algorithm, because all of them are about data and governance. That is where most of the real effort goes, and it is why AI money so often goes to the wrong layer.
Six questions a CFO should ask before signing
1. What is the baseline, and who signed it off? Current error, inventory cost, stockouts and expedited shipments, measured before the project and accepted by finance. Without it, any later saving is a matter of opinion.
2. What is the model being compared against? If the answer is "against what we do today" and what you do today is a moving average in a spreadsheet, ask for the comparison with a simple statistical method as well. Part of the improvement doesn't need AI, and you want to know how much.
3. How was it validated? The test has to mimic real use, training on the past and predicting weeks the model hasn't seen, in chronological order. Any other validation produces prettier numbers than the ones you will get in production.
4. Which decision changes, and who owns it? Reorder point, safety stock or load plan. If the project changes no policy with a named owner, the saving stays in the slide deck.
5. What is the full cost? Build, platform, model usage, retraining and internal hours, set against the net saving with a payback period. A 20% inventory saving that requires a dedicated team may not pay its way in a small company.
6. What happens if it doesn't work? An improvement threshold agreed in advance, a trial period with a control group and a commitment to switch the system off if it falls short. Agreeing on that up front is what separates a project from an open-ended expense.
Once these questions have answers, the numbers in the studies stop being someone else's promise and become a hypothesis your own operation can confirm or refute within a quarter. The business case for the committee and the board we published this month shows how to present that calculation.
What a committee can decide this quarter fits in one sentence: pick one product family or one warehouse where inventory and stockout costs are visible in the P&L, measure its baseline for a month and commission a model that has to beat it under measurement rules agreed with finance. A project like that fits in a quarter and can be switched off if it doesn't prove the saving, which is the best protection against miracles.
This is the work we do in enterprise AI implementation: a business case closed before anything gets built, predictive models where the money is and a conversational assistant only where it helps people decide faster. If every month-end brings a logistics line that is hard to explain, that is the conversation worth having.
