Pharmacy AI Helper: How We Built a Recommendation Engine for 100+ Pharmacy Locations

The problem

Walk into a pharmacy and ask the pharmacist for a recommendation. Behind that simple question is an impossible cognitive task.

A typical pharmacy carries thousands of products. Each has an active ingredient, a dosage, a list of contraindications, potential drug interactions, and a set of conditions it treats. A pharmacist is expected to know all of this, plus which products complement each other, which alternatives are cheaper, and which treatment protocols apply to the customer’s symptoms.

No human can hold all of this in their head. So what happens in practice is:

• Pharmacists default to the products they know best, missing cross-sell opportunities
• When a requested product is out of stock, the customer leaves instead of buying an alternative
• Treatment advice becomes generic (“drink water and rest”) instead of specific and actionable

Across 100+ pharmacy locations in the anri-pharm chain, this meant thousands of missed revenue opportunities every day, not because the pharmacists were bad at their jobs, but because the information they needed was not available at the speed of a customer interaction.

What we built

The Pharmacy AI Helper: an intelligent recommendation system deployed at pharma.cone.red.

The system works in four steps:

1. Product scanning

The pharmacist scans a product barcode or types a name. The system instantly identifies the product, its active ingredient, and its position in the medical knowledge graph.

2. AI analysis

The system traverses a knowledge graph of 2M+ medical relations: connections between drugs, symptoms, conditions, active ingredients, and treatment protocols. It identifies what this product treats, what interacts with it, and what complements it.

3. Smart recommendations

The pharmacist sees three types of recommendations:

• Alternatives: equivalent or cheaper medications the customer could buy instead (critical when the requested product is out of stock)
• Complementary products: items that enhance the treatment (e.g., vitamin C alongside a cold medication, throat lozenges alongside an antiseptic spray)
• Medical summaries: active ingredients, contraindications, and interaction warnings

Every recommendation is medically verified. The pharmacist can recommend with confidence, knowing the system pulled from official protocols, not a marketing database.

4. Treatment protocols

Beyond product-level recommendations, the system includes treatment protocols for common conditions: common cold (8 recommended substances), tension headache (7 substances), GERD (8 substances), allergic rhinitis (8 substances), and more. A pharmacist searching by symptom gets an evidence-based treatment path, not a guess.

The technology

The core of the system is a medical knowledge graph, a structured database of 2M+ relations:

• Drug-to-drug relations: which medications are equivalents, alternatives, or interact negatively
• Drug-to-symptom relations: which medications treat which symptoms
• Drug-to-condition relations: which medications are used for which conditions
• Symptom-to-condition relations: which symptom patterns indicate which conditions
• Condition-to-protocol relations: which treatment sequences are recommended for each condition

The graph covers 99% of standard pharmacy inventory. When a new product enters the market, the system’s ingestion pipeline maps it into the graph automatically.

This is not a vector embedding system. It is a structured knowledge graph: deterministic, explainable, and medically verified. For healthcare, you do not want “approximately relevant” suggestions. You want correct ones.

Results

Metric	Before	After
Inventory coverage	What pharmacists could remember	99% (all products known)
Medical relations	Manual lookup in reference books	2M+ instant relations
Cross-sell suggestions	Ad-hoc, dependent on pharmacist experience	AI-driven, medically verified, real-time
Out-of-stock recovery	Customer leaves	Alternative suggested automatically
Availability	Pharmacy hours	24/7 (system never sleeps)

The ROI driver is intelligent cross-selling. Every customer interaction now comes with medically verified complementary product suggestions. The pharmacist is not pushing products; they are making better recommendations because the system gives them the knowledge they could not hold in their head.

Across 100+ locations, the cumulative effect is significant: higher average basket size, fewer lost sales from out-of-stock situations, and consistently high-quality advice regardless of pharmacist experience level.

What we learned

The best AI does not replace the expert. It gives them instant knowledge they could not hold in their head. The system does not diagnose. It does not prescribe. It gives the pharmacist the information they need, at the speed of a customer interaction, so they can make better decisions.

Cross-selling is not manipulation when it is medically sound. Suggesting vitamin C alongside a cold medication is not a sales tactic. It is good pharmacy practice. The system makes it systematic instead of dependent on individual pharmacist knowledge.

Knowledge graphs beat embeddings for healthcare. In domains where accuracy is non-negotiable (medicine, banking, legal), you need structured, verified relations, not semantic approximations. The 2M+ relations are curated, not learned from internet text. This is the same principle behind VERA: retrieval is the product, and it must be correct.

Product-level AI creates compounding ROI. Unlike a one-time consulting engagement, a product deployed across 100+ locations generates returns every day, in every transaction. The system gets better as the knowledge graph grows. The locations get better as pharmacists trust the recommendations more.

---

This system was built by the team at Cone Red and is live at pharma.cone.red. If you are in healthcare, pharmacy, or medical retail and want to explore AI that enhances your team rather than replacing them, let’s talk.