AI in the hospital supply chain forecasts demand for supplies and medicines, sets inventory levels, flags shortages early and finds waste. Forecasting is the mature part. The larger savings sit in clinical supply decisions, especially surgeons’ preference cards, which only change when clinicians agree to change them. The algorithm finds the problem; people fix it.
At the end of a routine neurosurgical case, a nurse gathers what was opened but never used: sutures, sponges, drapes, a specialised device still in its sterile wrapper. It all goes in the bin, because once a sterile package is opened in the operating room it cannot go back on the shelf.
Researchers at the University of California, San Francisco counted this waste directly. In a study published in the Journal of Neurosurgery in 2017, they observed 58 neurosurgical procedures and found that opened but unused supplies cost an average of $653 per case, about 13% of total surgical supply cost. Adjusting for the mix of cases, they estimated about $968 of waste per case and roughly $2.9 million a year for one department.
Most of that waste was not caused by a bad forecast. It was caused by a list: the preference card that tells operating room staff which supplies to pull and open for each surgeon and procedure. That list is where AI in the hospital supply chain meets its hardest problem. Predicting how many boxes of gloves a hospital will need next month is a solved problem, more or less. Persuading a surgeon to take three items off a card they have used for fifteen years is not.
What does AI do in the hospital supply chain?
Hospital supply chains manage tens of thousands of items: medicines, implants, surgical instruments, disposables, linens, food and equipment. AI is being applied at several points.
Demand forecasting predicts usage by item and location, using historical consumption, scheduled procedures, census and seasonal patterns. Inventory optimisation turns those forecasts into par levels and reorder points, reducing both stockouts and cash tied up on shelves. Shortage and risk monitoring watches supplier data, manufacturing disruptions and recalls to flag items at risk. Data cleaning matches messy item descriptions across purchasing, inventory and clinical systems, which sounds trivial and is often where the first real savings appear. Utilisation analytics compares what was planned, opened and used, by procedure and by clinician.
The first four are largely operational problems that supply chain teams own. The last one crosses into clinical practice, and that is where the money and the resistance both concentrate.
Why is forecasting the easy part?
Because most hospital demand is predictable in aggregate. Elective surgery is scheduled weeks ahead. Seasonal patterns in respiratory illness repeat. Many consumables track census closely. Good forecasting reduces emergency orders and expired stock, and it frees nurses from counting and hunting for supplies. Those are real benefits, and hospitals that still order from spreadsheets and manual counts can capture them quickly.
But forecasts have a ceiling. They predict how much the hospital will use given current practice. They do not ask whether current practice is sensible. A perfect forecast of a wasteful habit is simply a well-stocked habit.
Forecasts also struggle with the events that matter most. When Hurricane Helene flooded a Baxter International plant in Marion, North Carolina, in late September 2024, a site reported to produce about 60% of the sterile IV and dialysis fluids used by US health facilities was temporarily shut, and hospitals across the country rationed IV fluids. No demand model predicted that. What helped were things outside the model: knowing exactly which products depended on that single site, having conservation protocols ready and having clinicians willing to change practice quickly, for example by switching some patients from IV to oral hydration. We argued in The Next Hospital Crisis Will Arrive Through Your Supply Chain that resilience depends on that kind of mapping and readiness more than on prediction.
Why are preference cards the real project?
A preference card is a surgeon-specific supply list for a given procedure. In many hospitals there are thousands of them. They accumulate items over years, rarely lose any, and are often out of date: surgeons change techniques, products change, and cards copied from one surgeon to another carry along items nobody uses.
The consequences are waste, when items are opened unnecessarily, and wasted time, when staff pick and return items that are never needed. Inaccurate cards also cause delays when an item a surgeon actually needs is missing.
This is a data problem that AI handles well. A team at Stanford, publishing in the Journal of the American Medical Informatics Association in 2021, built an algorithm that compared what was listed on preference cards with what was actually used, and estimated the likelihood that each listed quantity was wrong. It assessed the accuracy of about 469,000 surgeon-procedure-specific items. Nurses then used those estimates to revise 309 preference cards across eight surgical services. Direct supply cost per case fell by 8.4% for procedures with revised cards, while it rose by 13.2% in a comparison group, a difference that was statistically significant.
Notice who did the work in that study. The algorithm identified likely errors. Nurses revised the cards, with surgeons’ involvement. That is the pattern everywhere: AI can rank which cards and items to change, but the change itself is a clinical conversation.
Why do clinicians resist changing supply choices?
Surgeons are not being stubborn for the sake of it. A preference card encodes how a surgeon likes to work, what they want immediately at hand if something goes wrong and, sometimes, products they trained on and trust. Removing an item feels like accepting risk for someone else’s savings. If a card is trimmed and the missing item is needed urgently mid-case, it is the surgeon and patient who face the consequence, not the supply chain analyst.
Product choices add another layer. Standardising on fewer suppliers for implants, sutures or staplers can save a great deal, but surgeons may have strong views on device performance and, in some cases, relationships with manufacturers. Physician preference items have long been among the hardest spending categories for hospitals to manage. Procurement decisions shape clinical practice more than most hospitals admit, a point we made in The Most Powerful Quality Regulator in Your Health System Is the Procurement Office.
A worked example: 1,200 cards, one project
Take a hypothetical 500-bed hospital with 1,200 active preference cards and 18,000 operations a year. An AI utilisation tool compares each card with a year of usage data and flags 40% of listed items as rarely or never used.
The supply chain team’s first instinct is to send surgeons a spreadsheet of recommended removals. Response rate: low. A few surgeons agree; most ignore it; two complain to the chief medical officer that the supply team is interfering with patient care.
The second attempt is different. The hospital picks the 50 highest-volume cards, which cover a large share of cases. For each, a perioperative nurse sits with the surgeon for 15 minutes with the data: here is what is opened, here is what is used, here is the cost of each unused item per year. Items needed occasionally but critically are moved to “hold”, available in the room but not opened. Items never used are removed. Surgeons see their own data first and can override any recommendation, and the hospital promises to restore any item within a day if a card change causes a problem.
Over six months, most of the 50 cards are revised, with a meaningful fall in supply cost per case for those procedures and fewer complaints about missing items than before. The AI did the hard analysis in days. The project took six months because the hard part was the conversations. The scenario is illustrative, but it mirrors the approach in the published Stanford work: the algorithm prioritises, and clinicians decide.
Is supply chain AI overhyped?
The strongest counterargument is that the biggest supply chain problems are not in the hospital at all. Shortages of medicines and fluids are driven by fragile manufacturing, low margins on generic products and concentrated suppliers. No amount of AI inside a hospital fixes a sole-source plant under floodwater. From this view, hospital AI projects are rearranging shelves while the real risk sits upstream.
That is correct about shortages. Forecasting cannot create supply that does not exist, and the largest resilience gains come from purchasing policy, supplier diversification and national stockpiles. But it does not make hospital-level work pointless. Waste from opened, unused supplies is entirely within a hospital’s control, and so is knowing which products depend on a single source before a disaster reveals it. Supply chain AI is best understood as two separate things: an operational efficiency tool for the problems a hospital owns, and an early-warning system for the ones it does not.
What this means for hospitals, surgeons and suppliers
For hospital leaders, fund the clinical side of supply chain projects, not just the software. That means perioperative nurse time, surgeon engagement and a clinical owner, usually a surgical leader, who treats card accuracy as a quality issue. Report savings back to departments and, where possible, let them reinvest part of it, which gives surgeons a reason to engage.
Sequence the work sensibly. Clean item data first, because forecasts and utilisation analysis are only as good as the item master. Then tackle high-volume cards and high-cost physician preference items, where the returns are largest.
For surgeons, the useful stance is to treat the data as feedback rather than as an audit. Most surgeons do not know what is opened on their behalf and never used. Seeing it is often enough to prompt change, especially when it is framed around environmental waste and staff time as well as cost.
For supply chain teams, map dependencies before you need them: which critical products come from one plant, one country or one supplier, and what the substitute plan is. Share the map with clinical leaders so conservation decisions can be made quickly.
For suppliers, expect hospitals to use utilisation data in negotiations. Suppliers that can show their products reduce total procedure cost, rather than simply unit price, will do better as that data becomes common.
Frequently asked questions
How is AI used in hospital supply chains?
AI forecasts demand for supplies and medicines, sets inventory levels, monitors suppliers and news for disruption risk, cleans and matches item data across systems and analyses what is opened and used in each procedure. The most mature uses are forecasting and inventory optimisation. The largest clinical savings usually come from analysing utilisation to revise preference cards and standardise products.
What is a surgical preference card?
It is a list of supplies, instruments and equipment that a specific surgeon wants prepared for a specific procedure. Operating room staff use it to pick and open items before surgery. Cards often become outdated, listing items that are rarely used, which leads to waste when sterile items are opened and discarded, and to lost staff time.
How much do hospitals waste on unused surgical supplies?
It varies widely. A 2017 UCSF study in the Journal of Neurosurgery found opened but unused supplies averaged $653 per neurosurgical case, about 13% of surgical supply cost, and estimated about $2.9 million a year in waste for the department. Other specialties and hospitals will differ, which is why local measurement matters.
Can AI prevent drug and supply shortages?
Not by itself. AI can flag risks earlier by monitoring supplier data and disruptions, and can help hospitals manage scarce stock once a shortage starts. It cannot prevent shortages caused by manufacturing failures, disasters or market problems upstream. Those require purchasing, policy and supplier diversification. Hospitals benefit most by mapping single-source dependencies in advance.
How do hospitals get surgeons to update preference cards?
The most effective approaches pair data with conversation: showing surgeons their own usage and costs, starting with high-volume cards, moving rarely used but important items to “hold” rather than removing them, allowing quick restoration if a change causes problems and giving clinical leaders ownership. Spreadsheets sent by email rarely work on their own.
A hospital can forecast its gloves to the box and still throw away millions in sterile supplies that nobody touched. The algorithm can point to the waste. Only a surgeon can take it off the card.
Has your hospital tried to clean up its preference cards? Tell us what finally persuaded the surgeons.
