The Subtle art of Reconciling Medications
Why Medication Reconciliation Resists Automation—And how that's changing
Medication reconciliation should be easy to automate. Every prescription is tracked. Every refill is logged. Every dose is documented. And yet, across healthcare today, “med rec” remains one of the most error-prone, time-consuming tasks clinicians perform.
The disconnect reveals something important about healthcare, but really systems as a whole: the hardest problems aren’t computational—they’re human.
A med rec aims to do exactly what its name suggests: reconcile the provider’s chart with the medications a patient is actually taking. Straightforward in theory. In practice, this is where the neat digital trails break down—and where AI can genuinely help, if applied correctly.
How Prescription Medication Is Supposed to Work
The standard process for prescription medication looks clean on paper:
Patient encounter: A patient meets with a healthcare provider to discuss symptoms and health concerns.
Prescription: The provider prescribes medication as part of a treatment plan.
Pharmacy processing: A pharmacy receives the prescription, doses the medications, and packages them.
Dispensing: The patient picks up the medication from their preferred pharmacy.
At every step, the medication, quantity, and dose are tracked and reconfirmed. Every refill is logged. Prescriptions expire, so they can’t be used indefinitely. With this much documentation, reconciling these records should be trivial, right?
Spoiler alert: It’s not.
Where things break down
The problem is that patients don’t fit into neat boxes—and they aren’t healthcare professionals.
Medications get picked up but are not taken. Adherence is notoriously poor. Patients stop antibiotics when they feel better, not when they’ve completed the course. They skip doses to stretch expensive medications. They abandon treatments with side effects nobody warned them about.
Medications aren’t taken as prescribed. Some patients halve doses to save money. Others double up when they feel worse. PRN (”as-needed”) medications add another layer of ambiguity—how often, exactly, is “as needed”?
Patients don’t know what they’re taking. This is especially common among elderly patients managing multiple conditions, but it occurs across all demographics. The confusion between medication names, active ingredients, and brand names is pervasive.
The Lisinopril Problem
Consider a patient taking lisinopril for hypertension. They might tell one provider they take “Prinivil” (a brand name), another that they’re on “that blood pressure pill,” or “an ACE inhibitor”, or some collection of our personal favorites, “the green one”, “the round ones”, or “the ones that look like pills”. If they switched pharmacies and received a different generic manufacturer, the pill might look completely different—new shape, new color—leading them to believe it’s a new medication entirely.
Now multiply this confusion across a patient taking eight medications. Some of them stopped months ago. Some take it differently than prescribed. Some of them received as samples and were never filled through a pharmacy at all.
This is why a clinician can spend twenty minutes, if not more, on a med rec for a single patient—and still get it wrong.
The Data We Have (And Its Limits)
There’s more data available today than ever, but each source has significant gaps.
Health Information Exchanges (HIEs) aggregate prescription data and past reconciliation records across health systems. But HIEs are only as good as their network coverage, and they capture what has been prescribed, but not what was taken. We focus on what a patient is actually taking.
Pharmacy networks like Surescripts provide real-time access to dispensing data—fill dates, quantities. But again, they capture what was dispensed, not what was taken, and miss medications from non-participating pharmacies, mail-order services, or samples.
Patient interviews remain the gold standard for understanding actual use. But patients forget, misunderstand, and misreport—and interviews are time-intensive for already-stretched clinical staff.
None of these sources alone provides ground truth. Effective med rec requires triangulating across all of them.
How LLMs Can Actually Help
While clinicians are still necessary to determine reality in the context of a med rec, we have found several strategies that leverage both LLMs and other tools to reduce the time they need to spend getting there dramatically.
Pre-population and record scanning. AI can aggregate data from HIEs, pharmacy networks, and historical records to pre-populate a medication list before the patient encounter begins. Natural language processors can extract medication mentions from unstructured clinical notes, and other free text. The goal isn’t perfection—it’s giving clinicians a 70-80% complete starting point instead of a blank slate.
Risk stratification. Not all medication discrepancies carry equal risk, therefore should not be treated equally. Using AI to detect medications doesn’t need to be 100% accurate today, although one can dream, because flagging for a multi-staged triage approach will still dramatically improve results in many cases, such as:
Patients on anticoagulants where dosing errors cause bleeding or stroke
Patients with recent hospitalizations where medications frequently change,
Detecting discrepancies between prescribed and filled medications suggesting non-adherence.
By bringing a lot of this to the front without the clinician needing to spend time digging. Clinicians can focus their time on the patients and medications that matter most.
Pre-visit verification. Emerging AI voice agents can conduct structured phone calls to verify medication lists with patients before appointments—reading back current medications, probing for over-the-counter supplements and discontinued drugs, and documenting responses in structured formats for clinical review. This shifts routine verification work to AI while preserving clinician time for the complex cases.
The Payoff
When implemented thoughtfully, AI-assisted med rec delivers concrete improvements—less time on rote data gathering, freeing clinicians to focus on actual clinical judgment. High-risk items are caught first rather than discovered by accident—fewer medication errors, which remain a leading cause of preventable harm.
Conclusion
Med recs resists automation not because the data doesn’t exist, but because patient behavior is inherently messy. Until we’re able to credential an autonomous system completely, the next best step is to augment the current workflow with intelligent data aggregation, risk stratification, and preliminary verification, removing low-level tasks from it.
Let AI handle the tedious triangulation so clinicians can focus on the judgment calls that actually require their expertise. That’s not a revolution—it’s just good engineering applied to a stubborn problem.