Industries / Health / Healthcare / Pharmacy
Filling prescriptions
Filling a prescription means turning a prescriber's order into the right medicine, in the right amount, with a label the patient can follow, whether it is handed over a shop counter, sent up to a hospital ward or mailed.
7
Milestones achieved, from 1951 to 2024.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1951
Prescription-only by law
The Durham-Humphrey Amendment to the US Food, Drug, and Cosmetic Act defined the kinds of drugs that cannot be used safely without medical supervision and restricted their sale to a licensed practitioner's prescription. Since then a US pharmacist may hand over such a drug only against a prescription.
Credit
the US Congress, in an amendment named for its sponsors, Carl Durham and Hubert Humphrey
1960s
Medicines packed dose by dose
In unit-dose dispensing the hospital pharmacy sends each dose to the ward in its own labeled package, ready to give, in place of bulk bottles kept on the ward. It was developed in the 1960s to help nurses and cut waste; in a 1994 survey, 92% of US acute care hospitals used it, and it became a standard for hospital accreditation by the Joint Commission.
Lessona review for the US Agency for Healthcare Research and Quality found the evidence that unit dose cuts errors "modest", although it had long been standard.
Credit
Kenneth Barker, who began observing nurses' medication errors in 1961 and published the observation method with W. E. McConnell in 1962, and hospital pharmacy research groups funded by US Public Health Service grants
1983
Prescriptions sent electronically
In 1983 the computers of a medical clinic and a nearby pharmacy in Jönköping, Sweden, were linked, which Bengt Åstrand and colleagues call the world's first electronically transferred prescription outside hospital. Gunnar Klein's history of the Swedish system dates the first online prescribing pilot to 1984 instead. Swedish rules made such prescriptions legal to dispense in 1984, and by December 2007, 68% of new Swedish prescriptions were sent electronically.
Credit
a clinic and a pharmacy in Jönköping (the sources do not name them)
2010
Controlled drugs prescribed electronically
A US Drug Enforcement Administration (DEA) rule published on 31 March 2010, in effect from 1 June 2010, gave practitioners the option of sending prescriptions for controlled substances, drugs such as opioids whose handling the DEA tracks, electronically, and let pharmacies receive, dispense and archive them. The DEA said this could cut forged prescriptions and errors from illegible handwriting.
Credit
US Drug Enforcement Administration
2016
Directions in five languages
A 2010 survey of 286 pharmacies in the Bronx, New York, found that 73% gave Spanish labels, mostly made by computer programs, and that half of the 76 such labels examined were partly untranslated or contained errors. From January 2016 California required pharmacies to print translated directions on request, and its Board of Pharmacy publishes 15 standard directions translated into Spanish, Vietnamese, Korean, Russian and Chinese.
Lessonlabel software translated directions badly when nobody checked its output.
Credit
I. Sharif and J. Tse (the Bronx study); California Assembly Member Ting, author of AB 1073; the California State Board of Pharmacy
2021
More pharmacies closed than opened
Of the 88,930 US retail pharmacies that operated at some point in 2010–2020, 29.4% had closed by 2021, by the count of Jenny Guadamuz and colleagues using industry data; independent pharmacies and those in mostly Black and Latinx neighborhoods closed more often. An earlier study by the same group found that when a pharmacy closed, its older patients' adherence to statins fell by about 6 percentage points within 3 months and stayed lower for a year.
Lessonlosing a pharmacy changes whether people keep taking their medicines, not only how far they travel.
Credit
J. S. Guadamuz, D. M. Qato and colleagues
2024
Paper prescriptions end in Germany
From 1 January 2024 the electronic prescription, the E-Rezept, became mandatory for prescription-only medicines for people in Germany's statutory health insurance, replacing the pink paper form. Patients redeem it at a pharmacy with their electronic health card, an app or a paper printout of its code.
Credit
German Federal Ministry of Health
Watch for
Same-day delivery to 4,500 towns
Amazon says it will extend same-day prescription delivery to nearly 4,500 US cities and towns by the end of 2026, adding about 2,000, with statewide coverage in Idaho and Massachusetts. It gives pharmacy closures, staff shortages and transport barriers as reasons. Amazon entered the business by buying the mail-order pharmacy PillPack in 2018.
Achieved whenan outside count, by a state agency or researchers, shows same-day delivery reaching those places.
Claimed by
Amazon Pharmacy, February 2026
Tracking a prescription by app
NHS England launched prescription tracking in the NHS App in May 2025, so patients can see when a prescription is ready to collect, and said 5,000 pharmacies would sign up within a year. On 12 October 2026 its chief pharmacy and medicines information officer, Rahul Singal, said only 30% of pharmacies had it switched on and that adoption had "stalled".
Achieved whenNHS England or Community Pharmacy England reports that most community pharmacies in England offer tracking.
Claimed by
NHS England, May 2025
Open
01
A public label-direction test
A public set of at least 2,000 made-up prescriber directions (the "sig", such as "1 po bid x10d"), each paired with the label wording two pharmacists agree on and with its parts (dose, route, how often, how long) filled in, plus a scoring script. AI models and rule-based parsers are scored on how many directions they turn into a wrong dose, frequency or duration. No real prescription or patient record is used.
The briefCredit · unclaimed02
Machine-translated directions, scored
A published tally of how well machine translation tools and AI models translate standard label directions into the languages pharmacy customers speak. Start from California's 15 standard directions, add 200 made-up ones, and have two native-speaking pharmacists or professional medical translators grade every output.
The briefCredit · unclaimed03
An open count of closures
A public, monthly count and map of US community pharmacy openings and closures built only from public records: the federal NPPES registry of provider numbers, with its monthly deactivation file, and state boards' license lists, checked against states that publish their own counts.
The briefCredit · unclaimed
Open 01 · the brief
A public label-direction test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 2,000 made-up prescriber directions (the "sig", such as "1 po bid x10d"), each paired with the label wording two pharmacists agree on and with its parts (dose, route, how often, how long) filled in, plus a scoring script. AI models and rule-based parsers are scored on how many directions they turn into a wrong dose, frequency or duration. No real prescription or patient record is used.
Why it is still open
The published tests run on private pharmacy records. Amazon Pharmacy's MEDIC system cut near-miss errors in directions by 33% in production, but its authors say its data cannot be released under US health privacy law; PharmMT was tested on more than 530,000 prescriptions from a mail-order pharmacy; a 2020 daily-dose study used 1,000 real orders from 427 patients at one medical center; a 2025 French study turned free-text French dosing instructions into structured form with AI models, reaching 91% accuracy with a hybrid pipeline, on French prescriptions only. Open-source sig parsers exist (ParseRx; parsigs, whose authors call its training data private) without a shared answer key, and a search of Hugging Face datasets for "sig prescription" and "prescription directions" returned nothing.
Built with what exists
RxNorm drug names from the US National Library of Medicine; openFDA labeling for real strengths and dose forms; ParseRx and parsigs as baselines; general AI models and open-weight models such as Mistral Large 4 (see the Note "A trillion open parameters"); small models that can be fine-tuned on one graphics chip, such as Liquid AI's LFM3 (see the Note "One GPU to fine-tune").
First step
Write 200 made-up directions covering hard cases (tapering doses, "as needed", weekly doses, "as directed", mixed-up units) and have two pharmacists write the label wording for each on their own.
Done looks like
2,000 directions published under an open license, with the two pharmacists' agreement reported and an error tally by type for at least five systems, listing every direction a system turned into a wrong dose.
Open 02 · the brief
Machine-translated directions, scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A published tally of how well machine translation tools and AI models translate standard label directions into the languages pharmacy customers speak. Start from California's 15 standard directions, add 200 made-up ones, and have two native-speaking pharmacists or professional medical translators grade every output.
Why it is still open
The one test of label directions found was small and old. In 2021 Cornelison and colleagues ran the 38 directions for the top 100 US drugs through Google Translate alone, into three languages, and found 71% of the Spanish, 76.3% of the Arabic and 89.5% of the Chinese directions accurate; 29.1% of all the inaccurate translations, counting counseling points, were judged highly significant or possibly life-threatening. It did not test AI models or other languages, and its translations are not published as a reusable set. Newer tests use hospital discharge instructions: Kong and colleagues (BMJ Quality & Safety, 2026) found at least one inaccuracy in 56% of ChatGPT-4's Russian instruction sets. California's board publishes checked translations in five languages only.
Built with what exists
California's published translations as a first answer key; Google Translate, DeepL and general AI models; back-translation by a second system; a harm grading sheet like the one Kong and colleagues used.
First step
Run California's 15 directions through three systems into the board's five languages and grade the results against the board's own translations.
Done looks like
At least 200 directions in at least 10 languages, each output graded by two reviewers as correct, wrong but harmless, or possibly harmful, published with the reviewers' agreement and a tally by language and system.
Open 03 · the brief
An open count of closures
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, monthly count and map of US community pharmacy openings and closures built only from public records: the federal NPPES registry of provider numbers, with its monthly deactivation file, and state boards' license lists, checked against states that publish their own counts.
Why it is still open
Public maps exist, but the counts behind them rest on paid industry data that cannot be downloaded. Guadamuz and colleagues and the Associated Press both used data from the National Council for Prescription Drug Programs (NCPDP), which the AP called proprietary. In November 2025 NCPA, the trade group for independent pharmacies, and the University of Southern California launched what they call the first publicly available tool to find pharmacy shortage areas, an interactive map; the University of Pittsburgh's 2024 closure map, by its own notes, uses NCPDP data held under a data use agreement. Ohio's Board of Pharmacy publishes yearly closure totals from 2012, and England's NHS Business Services Authority publishes monthly openings and closures under an open license, but no downloadable US national series built from public records turned up. A Europe PMC search for NPPES and pharmacy closures found no study that uses the registry this way.
opb.orgncpa.orgpharmacy.ohio.govopendata.nhsbsa.netdownload.cms.gov
Built with what exists
the NPPES monthly and deactivation files from the Centers for Medicare & Medicaid Services; state board license lookups and open-data portals; US Census ZIP code population data; a coding agent to write and maintain the download scripts.
First step
Download the NPPES file, keep Ohio records with a community or retail pharmacy taxonomy code, and compare their 2024 deactivations with the Ohio board's 2024 total.
Done looks like
A dataset for all 50 states, refreshed monthly, whose Ohio closures for 2015–2024 come within 10% of the Ohio board's yearly totals, with the method and the misses published.