Supply chain analytics, explained.

Supply chain analytics explained: plain questions about stock, orders and shipments, answered from the systems of record

One Veltrimab question, from the morning load to the shortage call

1. The morning loadStock, orders, shipments and the approved forecast land in the supply warehouse at about 6 a.m.
2. A question comes up"Weeks of cover for Veltrimab by market," before the 11 a.m. shortage call.
3. The answer is builtFree stock by market divided by the weekly forecast, from the right tables, cut to the asker's markets, with its sources. This is the step this guide is about.
4. People decideAllocate, transfer between distribution centers, expedite a shipment, or tell a hospital when to expect stock.
5. The question comes backNext Monday the same question needs the same answer on new data, and a regulator may ask how a shortage was foreseen.

Step 3 is the subject of this guide. Veltrimab and every figure in the example are fictional.

Why a regional total misleads planners, and three more surprises in supply data

Weeks of cover, free stock, backorder, forecast bias

Fourteen terms are defined in the one-page summary.

What pharma supply chain analytics covers: six families of questions and the systems each one needs

In shortMost supply questions fall into six families: cover, backorders, inbound shipments, forecasts, shelf life and transfers. Each needs data from two or more systems and one decision about definitions, such as which stock counts as free. The arithmetic is easy. The hard part is joining the right tables with the same meaning every time.

Analysts sometimes sort analytics into descriptive (what happened), diagnostic (why), predictive (what will happen) and prescriptive (what to do). Supply planners rarely talk that way. They ask questions in their own words, and the questions repeat with small changes: a different product or SKU (stock-keeping unit: one product in one pack), a different market, a different week. A useful way to see the work is by the families those questions fall into, and by the systems each family has to read.

Two things follow from the grid. A dashboard built on one system answers only part of most questions, which is why so many planners end up in spreadsheets. And master data is in every row: if a product code means one syringe strength in the ERP and a family of products in the planning system, every join on it is wrong in a way no chart will show.

The same applies to time. Stock is a snapshot from the morning load. Orders change hourly. The forecast is a monthly plan, approved at a sales and operations planning meeting, and replaced the next month. An answer that mixes this morning's stock with last month's plan without saying so is still a number, and a planner will act on it.

Who holds which supply number: your plants, contract manufacturers, distribution centers and wholesalers

In shortStock passes through hands the company does not own, but responsibility does not pass with it. FDA's guidance on quality agreements keeps the product owner's quality unit responsible for final release of what a contract facility makes, and EU distribution guidance keeps the contract giver responsible for outsourced activities. So the owner needs the partners' data in its own systems, on agreed terms, with the status and lot of every unit.

When an owner uses a contract facility, FDA's 2016 guidance says, the owner's quality unit "is legally responsible for approving or rejecting drug products manufactured by the contract facility, including for final release". The quality agreement should define which records each party keeps and explain how they "will be made readily available for inspection".8 EU guidance on good distribution practice says the same of outsourced storage and transport: "The contract giver is responsible for the activities contracted out."5

For analytics, that has a practical meaning. A planner asking about cover needs to know whether a lot made at a contract manufacturer has been released, whether it is in transit to a distribution center run by a logistics provider, and whether a wholesaler already holds it. Each of those facts lives in someone else's system first and reaches the company's ERP by a feed, a portal or an email, on that partner's timetable.

Distribution rules that become fields in your data

Contract partners also explain why the same product can carry different codes. A contract manufacturer may book a product under its own item number, a wholesaler under another, and the owner's ERP under a third. The mapping between them sits in master data. When a new pack size or a second manufacturing site arrives, that mapping changes, and every report that joins on the old code quietly loses rows.

How to calculate weeks of cover, days of supply and safety stock shortfalls, with the Veltrimab numbers

In shortWeeks of cover is free stock divided by average weekly forecast demand. Days of supply is the same idea per day. Both are only as good as the two definitions under them: which stock is free, and which forecast. Worked by market, the example's region drops from a comfortable 5.2 weeks to two markets below the minimum.

The weeks-of-cover formula, step by step

The regional figure is right as arithmetic: 8,699 ÷ 1,660 = 5.24. It is wrong as a guide to action, because stock does not move between markets for free. A transfer takes time, may need a relabel for the receiving country, and leaves the sending market short if its own demand rises. So planners read cover market by market and decide transfers one by one.

Days of supply and safety stock at one plant

Days of supply divides free stock by average daily demand. Safety stock is the buffer the plan keeps against forecast error and supply delays; falling below it is an early warning that comes before a shortage. In the example's plant report, seven product codes at Plant 3 sit below safety stock, up from four the day before. One infusion set has 220 free units against a safety stock of 900, which is 3.9 days of supply. A wound dressing has no free stock at all against 1,800.

Two questions should follow every safety stock answer. Is the stock below safety level because demand rose or because a receipt slipped? And is anything on hold that quality could release soon? The first needs the forecast and receipts tables; the second needs the quality status. A report that shows only stock against safety stock answers neither.

Forecast accuracy and bias for supply planners: MAPE, the sign of the error, and how a biased forecast turns into a shortfall

In shortAccuracy says how far off the forecast was on average; bias says in which direction. A product family can meet its accuracy target and still run short if the error always points the same way. In the example, infusion sets miss the accuracy target and run 14% above forecast, and one infusion set's shortfall grows each week until its next receipt.

The common accuracy measure is MAPE, the mean absolute percentage error: for each week, the gap between forecast and actual as a share of actual, ignoring the sign, averaged. Bias keeps the sign. In the example's convention, positive bias means demand ran above forecast, so the product was under-forecast and stock runs down faster than planned. Teams use different conventions, so the convention belongs on every chart.

A worked week: Dravenzo IV cannulas in Germany were forecast at 1,300 units for the week of September 7 and sold 1,420. The error is 120 units: 9.2% of the forecast, or 8.5% of actual demand, the share MAPE uses. Over eight weeks the family's MAPE is 4.7% and demand ran 2.0% above forecast: small, and mixed in sign. That family is in good shape.

Infusion sets are different: MAPE of 24.5% and a bias of +14%. The family is noisy and consistently under-forecast. That is the pattern that produces shortages, because safety stock is sized for random error and cannot absorb an error that always points one way.

The projection is simple arithmetic: free stock today, plus this week's arrivals, minus backorders, minus the weekly forecast, rolled forward week by week until the next confirmed receipt. The infusion set starts 280 units short, loses 400 a week to forecast demand, and recovers only when 2,000 units arrive on November 2. If the 14% bias continues, the real gap is larger than the projection, because the projection uses the forecast. A good answer says that, and shows the bias beside the line.

Forecast versions add one more trap. Plans are approved monthly at sales and operations planning (S&OP) meetings, and old versions stay in the planning system. A question asked on October 6 should read the September plan, approved on September 29, unless the planner asks otherwise. When the October plan is approved, the same question will give a different answer, and both are right for their date.

The shortage notice rules a supply team answers to in 2026: FDA's six months, EU platforms for drugs and devices, Canada's five days

In shortIn the US, makers of certain drugs must tell the FDA six months before a discontinuance or interruption likely to disrupt supply, or as soon as practicable; makers of listed devices owe the same notice during or ahead of a declared public health emergency. EU device makers have owed a similar notice since January 10, 2025, and holders of centrally authorized medicines report shortages through EMA's platform. Canada asks for six months or five days. The EU's pharma package and Critical Medicines Act would add more, but neither is law on November 3, 2026.

These rules put a date on how far ahead a supply team must see. To give six months' notice, a company has to see six months ahead: cover by market, confirmed receipts, supplier and contract manufacturer status, and demand. The duty to notify falls on the manufacturer or marketing authorization holder; the numbers behind the notice come from the same tables planners query every day.

The US duties sit in sections 506C and 506J of the Federal Food, Drug, and Cosmetic Act (21 U.S.C. 356c and 356j).1,2 The EU device duty is Article 10a of the Medical Devices Regulation (MDR) and the In Vitro Diagnostic Medical Devices Regulation (IVDR), which covers in vitro diagnostic devices (IVDs).9

The Commission's Article 10a Q&A walks through a case worth reading: a manufacturer outside the EU confirms on May 1, 2025 that a device will be interrupted from December 1, 2025, and must inform the authority and its importers by June 1, "at least six months before" the interruption. The importers then inform their distributors, and the distributors the health institutions.9

Each link in that chain is a company whose stock and customers the manufacturer's data team may never have seen. The manufacturer can only start the chain on time if it sees the problem six months out.

Trade rules add a second reason to see ahead. A US proclamation of April 2, 2026 set tariffs on imported pharmaceuticals and their ingredients, with lower rates for some countries and for companies with approved onshoring plans.18 The Commerce Department opened a separate investigation into imports of protective equipment, medical consumables and medical equipment, including devices, on September 2, 2025; we found no report or proclamation on devices by November 3, 2026.19

The BIOSECURE provisions in Public Law 119-60 will limit federal contracting with certain biotechnology providers once their limits apply.20 Each of these can change which plant or contract manufacturer should supply a market, and each change shows up first as a supply question.

Drug shortages in 2025 and 2026: what FDA, US pharmacists, EU auditors and Health Canada counted

In shortFDA counted only four new drug shortages in 2025 and says it helped prevent 330. US pharmacists counted 227 active shortages in mid-2026, rising for three quarters. The two counts measure different things. Both say the same thing to a planner: most of the work is early warning, and most causes are never published.

FDA's count of new shortages and the pharmacists' count of active shortages are often quoted side by side, and they disagree by design. FDA counts shortages its own staff confirm, nationwide, and counts the ones it prevented separately. The pharmacists' service counts any product that pharmacists cannot get in the usual way, including those FDA would not call a national shortage. A supply team should know which count a customer, a journalist or a regulator is using.

The causes that are disclosed point back at supply data. A supply and demand mismatch is a forecast that was wrong in one direction for too long. A manufacturing issue is a receipt that did not arrive, at a company's own plant or at a contract manufacturer. A business decision is a discontinuance that, under US law, should have been notified six months earlier.1 Each one shows up in the tables a planner queries before it shows up in a count.

The same report found that sole-source products made up 48% of new shortages in 2026, and controlled substances 16% of active ones.6 One supplier for a product means one plant's receipts decide its cover, which is the case where seeing confirmed receipts and supplier status beside stock matters most.

Where supply planners' time goes: data preparation, manual tracking and inventory that keeps growing

In shortPublished figures on planners' time come from consultancies and software vendors, so read them as indications, with their dates. They agree on one thing: preparing and tracking data takes a large share of the week. One survey puts it at 41% of planning time. Meanwhile pharmaceutical companies hold more inventory than a year ago, which makes knowing where it sits worth more.

None of these surveys is about pharmaceutical or device manufacturers alone, and none publishes its method in full. We found no regulator or academic study that measures how long a supply planner spends finding data. Part C shows how to measure it on your own team in a week.

More stock does not mean better cover where it is needed. Days of inventory are counted across a company's whole balance sheet: raw materials, work in progress and finished goods in every market. The example's region shows the point in miniature: plenty of stock in total, too little in two markets. The more inventory a company carries, the more it gains from knowing exactly which part of it can serve which customer this week.

Eight ways a supply answer goes wrong, from quality-hold stock counted as free to a market leak through a shared account

In shortMost wrong supply answers are right arithmetic on the wrong rows: stock on hold counted as free, the wrong forecast version, stock counted where it sits instead of where it is promised, a shipment counted before it is received. The worst error is of a different kind: a reader sees another market's stock, or supplier prices, because the query ran with more access than the reader has.

The access leak deserves a closer look, because it is the one a screen cannot fix. In PostgreSQL, a widely used database, "superusers and roles with the BYPASSRLS attribute always bypass the row security system", and "table owners normally bypass row security as well".7

Other databases have their own versions of the same rule. A self-service tool that checks access in its interface and then queries through one service account returns every row to its own code; whether the user sees them depends on the interface getting every case right. A filter written into the query itself, with the asker's own markets, keeps the rows in the source.

The other errors share a cure: write the definitions down once, apply them in every query, and put them on the answer. "Free stock = unrestricted stock minus allocated units; quality hold excluded; September plan; syringes" is one line under a chart, and it is the line that lets a quality reviewer or an auditor check the number.

The data integrity principles that regulators use, known as ALCOA+, ask for records that are attributable, legible, contemporaneous, original and accurate, and complete, consistent, enduring and available.27 A supply answer with its source tables, refresh time and query meets that standard far better than a forwarded spreadsheet.

Asking supply data in plain words: what AI text-to-SQL gets right, where it fails on enterprise data, and what regulators say

In shortLarge language models can turn a planner's question into a database query, and on classroom benchmarks they do it well. On real enterprise databases they do it badly: the best systems on two published enterprise benchmarks solve about 11 to 21% of tasks. The failures are the same ones human analysts make, wrong columns and unstated assumptions, so the cure is narrow scope, written definitions and answers checked against agreed results. FDA's AI draft guidance leaves this kind of operational use out of its scope.

Text-to-SQL means turning a question in plain words into a query in SQL, the language databases understand. Research benchmarks measure it by running the generated query and checking the result. The best-known benchmarks, Spider and BIRD, use small public databases. Two newer ones use the kind of data companies actually have.

BEAVER, built by researchers at MIT, AWS AI Labs and TUM from private enterprise data warehouses, is harsher still. In its January 2025 version, GPT-4o answered 69.5% of Spider questions and 30.9% of BIRD questions correctly end to end, and none of BEAVER's.29 Its May 2026 version reports that the best published frameworks on a newer model reach 10.8%, and 30.1% when given human hints for each sub-step.29 The error analysis explains why, and it reads like a list of supply reporting mistakes.

"Missing implicit assumptions" is the free stock problem: the question says "stock" and the right answer needs "unrestricted stock minus allocations". "Incorrect column mapping" is the location-for-market problem. A model that sees hundreds of tables guesses; a model that sees five tables, with a written rule for what free stock means, has far less to guess. That is the design argument for narrow specialists over one general model, and for a set of real questions with agreed answers that is asked again every night to catch drift.

The security risks of letting people ask a database in plain words

What FDA, EU GMP drafts and the AI Act say about supply analytics

FDA. The January 2025 draft guidance on AI to support regulatory decision-making for drugs "does not address the use of AI models … when used for operational efficiencies (e.g., internal workflows, resource allocation, drafting/writing a regulatory submission) that do not impact patient safety, drug quality, or the reliability of results".31 Answering supply questions sits there, as long as no answer decides a release or a quality outcome. The draft was still a draft on November 3, 2026.

Device software. FDA's final Computer Software Assurance guidance of February 2026 covers software used in medical device production and quality systems.32 A device maker that uses a supply tool inside its quality system should read it for that use; a planning question is outside its scope.

EU GMP. EudraLex Volume 4 still lists Annex 11 on computerized systems at its January 2011 revision and no Annex 22.33 The draft Annex 22, consulted on from July to October 2025, says large language models "should not be used in critical GMP applications".34 EMA's report of a June 2026 workshop shows consultation comments asking to allow them under a documented, risk-based control strategy; no revised text is published.35 A tool that reports cold-chain logger readings and leaves the release decision to quality stays on the non-critical side of that line.

Records. Where supply records are GxP records, FDA's Part 11 asks for "secure, computer-generated, time-stamped audit trails" of entries and actions.36 An answer that keeps the question, the query, the tables and the time is the record an inspector or quality reviewer would ask for.

EU AI Act. After the AI Omnibus, the 2026 amendment that reset the AI Act's application dates, in force July 27, 2026, the rules for high-risk systems listed in Annex III apply from December 2, 2027, and for AI in regulated products from August 2, 2028.37 Whether a supply analytics tool is high-risk depends on its use; a question-answering tool for planners is not in the Annex III list as such. Check with your own counsel.

Row-level and column-level access for supply data: why each reader should see only their markets, and how to prove it

In shortSupply data is commercially sensitive inside the company: one market's allocations, another's standard cost, every supplier's price. The safe design applies each person's markets inside every query, masks columns they do not read, counts what was left out without naming it, and keeps a record of who saw what. A leader can then preview what a colleague will see before sharing an answer.

In the example, the same Veltrimab question has three correct answers. The regional director sees nine markets and the value at standard cost. A planner for Central Europe sees four markets, with cost masked, because planners do not read cost. The commercial lead for the UK and Ireland sees one market. Each answer is complete for its reader, and each says how many markets it left out, so nobody mistakes a smaller total for the whole.

A data owner, usually the person responsible for the supply warehouse, decides two lists: which tables each kind of question may read, and which markets each person sees. Both change rarely and should change by request and approval, with a record. Adding a table, or a market for a person, is a decision with a name and a date on it.

How supply teams answer a new question today: who asks, who builds the answer, and where the days go

In shortA planner or leader asks; a dashboard answers if someone foresaw the question; otherwise an analyst pulls extracts from three systems, joins them, agrees definitions, cuts the result for each reader and sends a chart. The answer often arrives after the meeting it was for. Most of the elapsed time is waiting in a queue.

Planner, director, commercial lead, quality and data owner

A new cross-system question passes through four people

In-house reports, outsourced analytics or software

The Veltrimab cover question: asked at 9:42, answered in 2.6 seconds, previewed as two colleagues before it is shared

In shortLena Ortiz, the supply chain director for Europe, the Middle East and Africa, asks for Veltrimab's weeks of cover by market. A screener reads it as a read-only request, her access is applied, two specialists query stock and forecast, and the chart draws in 2.6 seconds with its sources and queries beneath. She then previews it as a planner and as a commercial lead before she shares it.

1

A question box, with answers already run for the team

Lena's home is the question box and suggested questions, on data loaded at 6:10 a.m. Three answers already ran for her team this morning: 6 of 14 European backorder SKUs have stock arriving this week, 7 SKUs at Plant 3 are below safety stock, and 5 of 6 product families are inside the 20% forecast accuracy target.

On the right, the access panel says who she is answering as and which markets she may see: nine, with four others shown only as locked.

2

Each step shows as it runs

"Weeks of cover for Veltrimab by market." The question screener reads it as read-only, brand Veltrimab, grouped by market, in 0.2 seconds. Lena's access is applied: nine markets, cost visible. The orchestrator routes it to two specialists: Inventory for free stock in the ERP, net of allocations, and Forecast for weekly demand in the September plan.

Each specialist runs one read-only query on its own tables. The chart draws with four markets outside her access left out.

3

Nine markets, two below the minimum

The bars draw against the 3-week minimum and the 4-week target. Belgium at 1.6 and the United Kingdom at 2.2 are below the minimum; Italy sits at 3.1; the Nordics lead at 9.1. Three numbers sit under the chart: 5.2 weeks overall, 2 of 9 below the minimum, 8,699 free syringes.

The lock line reads: "Showing 9 of 13 markets · 4 outside your access are left out of the chart, the total and the query." The 1,500 syringes arriving on October 8 count once received.

4

The sources and the exact query, open to read

Under the chart: stock from ERP inventory, refreshed at 6:05 a.m., 118 rows read; demand from the supply planning system, September plan of September 29, 36 rows; the product master. Both queries are shown in full. Each carries the line market_code IN (:my_markets) with the number of markets it allowed.

The reasoning tab lists the six definitions the answer used, such as "Stock in transit or on quality hold is left out until it is received and released."

5

Previewed as Sam Patel, then as Grace Whitfield

Before sharing, Lena previews the answer as Sam Patel, the Central Europe planner. The chart redraws with his four markets, Germany, Switzerland, the Netherlands and Belgium: 6.8 weeks overall, 1 of 4 below the minimum, 4,239 free syringes, cost masked. As Grace Whitfield, the UK and Ireland commercial lead, it shows the United Kingdom only, at 2.2 weeks.

Preview only narrows. Every preview goes into the audit trail.

The example follows the screens exactly: Veltrimab's cover in each market, 8,699 free syringes and 1,660 a week are the product's sample data. The people, products, plants and figures are fictional.

Twelve more supply screens: backorders against arrivals, a forecast shortfall, cold-chain loggers, the access matrix and the audit trail

In shortOne question is the story. These screens show the follow-ups and the controls: backorders matched to this week's arrivals, the forecast impact, stock below safety level, accuracy by family, cold-chain logger status, transfers between distribution centers, an answer outside the asker's access, a stopped request to change data, the library of saved answers, who sees which markets, and the dashboard.

Six read-only agents answer supply questions; a supply chain leader validates shared answers and the data owner decides what is read

In shortEach agent has one job and limits it does not cross. None of them changes data, judges product release or widens anyone's access. A supply chain leader validates the answers the team shares, and the data owner approves every new table for a specialist and every change to a person's markets.

From question screener to forecast specialist: the six supply agents

The director validates shared answers; the data owner approves tables and markets

The connection to the warehouse accepts read-only statements, enforced on the connection as well as by the agents. Every question, routing decision, query, row count, preview and stop is kept in the audit trail for six years, with who asked and when. The 34 validated questions are asked again each night and compared with their expected answers; a drift goes to a person, as on October 1, when a new plan version moved one answer and the director confirmed the new figure.

ERP stock, the approved S&OP forecast and shipments in; charts with their sources, queries and what was left out

In shortThe solution reads the systems the supply chain already runs on, writes nothing back to them, and leaves behind answers that carry their tables, refresh times and queries, inside each reader's access. Saved answers re-run after the morning load.

Systems a supply analytics rollout connects to

Seconds to a sourced chart, no write paths, and 30+ validated questions at a medtech manufacturer

Refresh schedules, cover thresholds and who may preview are settings your team chooses. Read-only queries, answers inside the asker's access and the count of what was left out are always on.

What a supply question that waits days costs you in planner hours, late shortage notices and hospitals left waiting

In shortStart from your own numbers, because nobody publishes how long a supply question waits or how often a supply report is wrong. A few published figures frame them: how much planning time goes to preparing data, what shortages cost the hospitals you supply, and how long the law gives you to warn a regulator.

Finding your own supply analytics baseline in a week

Hospitals waiting, trust lost and notices that came late

What plain-language supply questions return when you count planner and analyst hours against the tool's yearly cost

In shortPut in your own numbers. Only the 41% share of planning time on data preparation comes from a published survey; every other default is an assumption to replace. The calculator values the planner and analyst hours released and sets them against what the tool and your testing cost a year.

Planner time, the analyst queue and answers you can trace: where the value comes from

Planners, hours on data and the analyst requests you send a month

At 10%, the yearly value is $132,020 of planner time plus $126,000 of analyst time, $258,020 in all, or $4,962 a week: 30 weeks to cover $150,000. At 25%, $456,050 a year and 17 weeks; at 50%, $786,100 and 10 weeks. A team of five planners that rarely sends a question to an analyst sees a much smaller figure; run your own numbers before buying anything.

How ready your supply data is for plain-language questions: tables, definitions, access rules, a question set and an owner

In shortPick the answer closest to your team today. Eight questions score 0 to 2; the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

What to ask any supply analytics or conversational BI vendor about access, definitions, read-only queries and proof

In shortAsk every vendor the same questions, us included. The answers show whether each reader's access is applied in the query, whether definitions are yours and written down, whether anything can change data, and how you would measure the tool on your own questions.

Pilot plain-language supply questions on a copy of your tables, with 30 agreed answers, then decide on three numbers

In shortMeasure first, then test on your own tables with questions whose answers you already agreed, then decide with three numbers: how many answers matched, how long they took, and whether any reader saw a row outside their access. None of these steps commits you to a vendor.

1. Time your baseline: five planners log a week; count last month's analyst requests and their days to answer.
2. Write the definitions: free stock, plan version, units, market assignment, one page each, agreed by planning and quality.
3. Collect 30 questions from recent shortage calls and S&OP meetings, with the answers you used, across cover, backorders, inbound, forecast and shelf life.
4. Set the access test: three people with different markets ask the same questions; add five questions that should be stopped.
5. Run and decide: match rate, median time, rows outside access, and what your planners say after four weeks.

Match rate, time to answer and rows outside access

Supply chain analytics questions planners and buyers search for

In shortThe questions people search about pharma and medtech supply chain analytics, each answered on its own.