The Ohio Board of Pharmacy released eight proposed rules on August 14, 2026, rewriting how central fill, remote order processing, prescription kiosks, and automated dispensing operate in the state. Comments close September 30 [Ohio Board of Pharmacy, 2026].
One number in that packet decides whether a machine may complete a dispense with no pharmacist looking at the drug. Proposed rule 4729:5-3-17.1(F) requires a pharmacist to verify everything the system puts out for a continuous forty-five-day period, and sets the passing mark at “ninety-nine and nine hundred eighty-five thousandths (99.985) per cent” [Ohio Board of Pharmacy, 2026].
That is one error in 6,667 dispenses. Ohio is one of a small number of boards willing to write a number instead of a prohibition, and the instinct behind it is correct. The arithmetic underneath it is not.
Clinical Context
Most states regulate machine dispensing by controlling who stands next to the machine. Ohio decided years ago to regulate the output instead, which is the harder and better choice, and it is the same shift TheraIntel traced when Ohio’s peers were still arguing about the pharmacist of record.
The 99.985 figure is not new. It already sits in current rule 4729:5-3-17, effective June 4, 2021, with prior effective dates running back to March 13, 1995 [Ohio Administrative Code 4729:5-3-17]. What changes is the company it keeps. The existing rule governs machines sitting inside a hospital or a nursing home, with a pharmacist physically present everywhere except long-term care. The proposal opens two new doors: paragraph (M) authorizes automated systems that dispense outpatient prescriptions directly to a patient, and the surrounding package lets pharmacists, interns, and technicians process those orders from their residences, including from outside Ohio [Ohio Board of Pharmacy, 2026].
A threshold written for a cart-fill machine in 1995 is about to become the gate for a machine that hands a bottle to a member of the public. Nobody re-examined the math on the way through.
Key Findings
1. The threshold has no stated denominator. Paragraph (F)(2) sets the accuracy metric at 99.985 per cent without naming what is being counted. The quality assurance paragraph three lines down tiers sites by “dosage units per day” while requiring “a review of all dispensed prescriptions” [Ohio Board of Pharmacy, 2026]. Those are different units. A ninety-count fill is one prescription and ninety dosage units. Read one way, the rule tolerates one wrong bottle in 6,667. Read the other way it tolerates one wrong tablet in 6,667, which at ninety tablets a fill is a compromised prescription every 74 dispenses. The Board says it will supply the reporting form. It has not supplied the unit.
2. The validation window cannot resolve the number it specifies. A site running 200 dispenses a day through the system produces 9,000 across forty-five days. At 99.985 per cent, that window permits 1.35 errors, so the pass line is one error and two errors fails. Run the distribution: a system performing exactly at the standard throws two or more errors about 39 per cent of the time. To be 95 per cent confident that a true error rate sits at or below 0.015 per cent, you need roughly 20,000 clean observations, which is 444 dispenses a day for forty-five consecutive days. High-volume central fill clears that. A hospital outpatient window or a nursing home cart-fill does not.
3. The same window passes systems that should fail. Hold the site at 9,000 dispenses. A machine running at 99.97 per cent, twice the allowed error rate, clears the gate 25 per cent of the time. At 99.955 per cent, three times the allowed rate, it still clears 9 per cent of the time. The gate is noisy in both directions.
4. Surveillance thins out exactly where the volume is. After approval, the current rule requires ongoing review of at least 5 per cent of dispensed prescriptions [Ohio Administrative Code 4729:5-3-17]. The proposal holds 5 per cent below 10,000 dosage units a day and drops it to 2 per cent above that line [Ohio Board of Pharmacy, 2026]. At 200 dispenses a day for a year, a system operating precisely at the standard generates about 11 errors. A 5 per cent sample expects to catch one of them every other year. A 2 per cent sample expects one about every four years. Paragraph (H) is the tripwire: a single incorrect drug selection forces the pharmacy back to 100 per cent pharmacist verification until it identifies and fixes the cause. A tripwire only works as well as the sampling that trips it.
How This Plays Out
For sites running a machine under this framework, the obligations stop being theoretical:
- Board review of the system is site-specific, so a chain approving one configuration in Columbus has approved nothing in Toledo.
- The forty-five-day validation is a real labor line. Every dispense gets a pharmacist check, at a site that adopted the machine to avoid exactly that.
- The proposal bars automated outpatient systems from controlled substances, most hazardous drugs, and new pediatric prescriptions requiring reconstitution, which pulls a meaningful share of a retail queue back out of the machine.
The kiosk rule shows the same habit. Ohio would bar a self-service kiosk within ten miles of an outpatient pharmacy, while NABP’s Model Act tells states they “should not impose volume restrictions, mileage restrictions, or unnecessary limitations that would limit patient access to remote dispensing sites” [NABP, 2026]. Colorado repealed its twenty-mile rule in 2022. Ohio is drafting one in 2026, the same lag TheraIntel found in Maine.
TheraIntel Perspective
Set 99.985 per cent against the humans it replaces. The national observational study of dispensing accuracy in 50 pharmacies across six cities found an overall accuracy rate of 98.3 per cent, 77 errors among 4,481 prescriptions, with 5 of those 77 judged clinically important [Flynn, Barker, and Carnahan, 2003]. Ohio is asking the machine to make roughly 113 times fewer errors than trained pharmacists were observed making.
The machines probably are that much better. Barcode-driven counting hardware has no bad shift and no third consecutive twelve-hour day. But the rule does not demonstrate that. It asserts a number two orders of magnitude past human performance, collects a sample too small to confirm it, and then watches the running system at 2 per cent.
“A five-significant-figure standard checked against a four-figure sample isn’t a safety threshold. It’s a number that sounds like one.”
The Board is asking whether the machine is accurate. Nobody in the packet asks whether forty-five days is long enough to find out.
The Bottom Line
Here is the prediction. The 99.985 figure survives to final rule unchanged, and it survives because comment letters get written about scope and money, not about statistical power. The kiosk radius will draw fire from every operator it blocks, and the remote-work provisions will draw fire from labor. Nobody files a comment about the sample size of a validation window, so the drafting error rides through to adoption and becomes precedent for the next state that copies Ohio’s language, which is how 1995 arithmetic reached 2026 in the first place.
Ohio has taken the right position: measure the machine rather than legislate a human next to it. That position is worth more than any comment letter about kiosk mileage. It is also worth getting the measurement right, and the window has until September 30.