The Human Check on an AI-Extracted Prescription: A 90-Second Routine

Every AI prescription reader worth using produces a draft, not a decision. The draft is the AI's best guess at what the doctor wrote; the decision is yours, made in 90 seconds by comparing the draft to the photograph in the same screen. Skipping this step is where wrong medicines enter the record — and once they are in the record, they get taken.

The four fields to check, in this order

• Field: Drug name · What to compare against: The spelling on the photograph · What to look for: Similar-looking substitutions (Tamoxen/Tamoxifen)
• Field: Strength · What to compare against: The number on the photograph · What to look for: 5mg vs 50mg — a 10x difference is easy to misread
• Field: Frequency · What to compare against: Timing marks (1-0-1, TDS, BD) · What to look for: 1-0-1 (twice daily, morning and night) is not the same as 1-1-1 (three times)
• Field: Duration · What to compare against: How many days written by the doctor · What to look for: Antibiotics especially — an antibiotic short by two days is a treatment failure

The specific traps

Six errors an AI reader will confidently make on Indian prescriptions:

• Reading a badly-formed 5 as an S. 'Metformin S00mg' looks like a typo; the real reading is 500mg.
• Splitting a combination drug into two. 'Amoxicillin + Clavulanate 625mg' is one tablet; getting it as two makes the pharmacy misinterpret.
• Confusing 'q6h' (every 6 hours) with 'q6d' (every 6 days). A big difference for an antibiotic.
• Getting the frequency but missing the meal timing. 'Before food' vs 'after food' matters for absorption and gastric side effects.
• Missing the 'as needed' qualifier. A painkiller written 'PRN' (as required) is not a scheduled medicine.
• Adding a plausible drug where the writing was illegible. This is the most dangerous failure mode.

What to do when a field is unclear

If a field is genuinely unreadable, do not guess. Two safer paths:

• Call the clinic and ask. Most receptionists will read the prescription back if you name the doctor and the date.
• Show the photograph to the pharmacist when dispensing, before adding to the record. Pharmacists read doctors' handwriting for a living; their guess is usually better than yours or the AI's.

Filling an unclear field with 'my best guess' in the record is worse than leaving it blank. A blank is a known unknown; a wrong entry is an unknown unknown that gets acted on.

The salt-name cross-check

A useful habit once the extraction is confirmed: compare the extracted salts against the family's allergy list. Most good record apps do this automatically. A prescription for 'Ceftriaxone' in a person with a recorded 'cephalosporin allergy' should stop you at the confirmation screen — that is exactly what the record is for.

Confirming, then acting

Only after the checks above should the record be saved. Only after the record is saved should the medicine be started. This is not paranoia; it is the workflow that catches the errors an AI reader by itself cannot. The reader saves you the typing. The confirmation saves you the wrong medicine.

References

• JAMA — technology-assisted medication reconciliation
• BMJ — prescription errors and error prevention
• PMC — human-in-the-loop for AI in healthcare