In a regulated distributor a forecast is not a suggestion; it moves product with an expiry date, a cold chain and a paper trail. When an inspector asks how a reorder number was built, “the model decided” is not an answer.
In most industries a demand forecast is an internal planning aid; if it is wrong, someone reorders. In pharmaceutical distribution the same number sits inside a chain of records that a regulator can ask to see. It drives purchase orders for product that carries a lot number, a storage condition and an expiry date. It sets the stock that determines whether a patient gets a medicine or faces a shortage. And it feeds the write-offs, returns and destructions that show up in the books. Every one of those downstream events is documented, and documentation in this sector is governed.
That governance has a name. ALCOA, introduced by the FDA in the early 1990s and now expanded to ALCOA+, holds that records must be Attributable, Legible, Contemporaneous, Original and Accurate, plus Complete, Consistent, Enduring and Available. It is the minimum expectation under 21 CFR Part 211 and the equivalent Good Distribution Practice rules in Europe. The principle is simple to state and hard to fake: you must be able to show who did what, when, on what data, and get the same answer when someone checks later.
Hold a typical black-box or agentic planning tool up to that standard and the cracks appear fast. The FDA defines an audit trail as a secure, computer-generated, time-stamped record that lets a reviewer reconstruct the creation, modification or deletion of a record. A model that produces a reorder quantity nobody can reconstruct does not have an audit trail; it has an output. If the same inputs can yield a different figure next week, the record is not Consistent. If no one can explain, line by line, how the number was built, it is not Attributable in any meaningful sense.
This is not a hypothetical worry. Data-integrity findings have become one of the most common reasons the FDA issues warning letters and Form 483 observations; by one analysis 61% of warning letters in 2021 contained a data-integrity finding, and audit-trail review is a stated focus for 2026 inspections. A planning layer that cannot show its work is a data-integrity finding waiting to be written up.
The stakes on the shelf are just as real. US drug shortages reached 323 active shortages in early 2024, the highest count since tracking began in 2001. Shortages are a planning failure as much as a supply failure; when the forecast and the reorder logic cannot be trusted or explained, buffers get set by fear rather than by math, and the result is stockouts of some products next to expiry write-offs of others.
Expiry is where opaque planning turns into a number on the P&L. Dated product that was ordered on a forecast nobody can defend does not get a second chance; it is quarantined, returned or destroyed. The amounts are not trivial even at small scale: one US pharmaceutical filer reported expired-inventory write-downs rising from roughly 22,000 dollars in one year to roughly 445,000 dollars the next, a twentyfold jump, driven largely by a single quarter. Multiply that pattern across a distributor's catalogue of thousands of dated SKUs and the working capital tied up in stock that will never be sold becomes the quiet tax on the whole operation.
The cruel part is that bad inventory costs twice. First you pay to buy and hold product that expires; then you pay again in the stockouts and expedites on the items you under-bought while cash sat in the wrong place. A forecast you cannot audit tends to be wrong in both directions at once.
The fix is not to abandon advanced methods; it is to draw a clean line between the parts of the problem that benefit from learning and the parts that must be deterministic. Let statistical and machine-learning methods sharpen the demand signal, where more pattern-finding genuinely helps. But the decisions that create regulated records, how much safety stock to hold against a given service level, when to reorder, how much to write down, should run on transparent math that returns the same answer every time and can be read by a human.
In practice that means three properties an inspector can verify. Reproducibility: rerun last month's result on the same data and get the identical number. Traceability: show, for any single figure, the inputs and the steps that produced it. Durability: the number survives a review six months later, when the planner who ran it has moved on. A system built this way does not fight the audit; it hands the auditor the trail. That is the difference between a tool that adds regulatory risk and one that reduces it.
When the forecast drives regulated records, purchase orders, lot-level stock, expiry write-offs and shortage decisions, the data and logic behind it are expected to meet the same ALCOA+ expectations as other GxP records: attributable, contemporaneous, accurate, consistent and available for review. A number that cannot be reconstructed is a data-integrity exposure.
Because regulators require an audit trail: a secure, time-stamped record that lets a reviewer reconstruct how a figure was created. A model that cannot reproduce the same output on the same inputs, or explain how a reorder quantity was built, cannot provide that trail. The output exists; the defensible record does not.
Yes, if you separate learning from deciding. Machine learning can sharpen the demand signal, while the decisions that create regulated records (safety stock, reorder points, write-downs) run on deterministic, transparent math that returns the same answer every time and can be read by a person. The model proposes; an auditable method disposes.
Can you rerun last month's result on the same data and get the identical number? Can you show, line by line, how any single figure was built? Will that number survive an audit six months later? If the answer to any is no, the tool adds regulatory risk rather than removing it.
In pharma, a forecast you cannot reproduce is not a planning problem; it is a data-integrity finding waiting to be written up.
Picture your next S&OP meeting: one number on screen that nobody argues with, and a trail behind it that an auditor could follow. See it free, on your own data, in 45 minutes; and if it does not hold up, we part as friends.