Designing Remy AI
We reduced tickets by 48% while enhancing discovery, education, and self-service for 250K+ providers.
Mobile
AI-Design
eCommerce
Overview
Role
Lead Product
& UX Designer
Company
Remedial Health
Team
1 Product Manager
2 UX Designer
3 Engineers (1 ML)
Timeline:
16 weeks
250K
Healthcare providers on platform
62%
Reduction in support tickets
16
Week build & launch
01
Conversational layer, five jobs
Before — support queue
"Where is my order?"
"Do you have another brand?"
"What should I restock?"
"Can these medications be returned?"
After — one conversation
What should I reorder today?
3 products are low. Here are your recommended reorder quantities →
Inventory
Orders
Knowledge
Support
The Problem
Every answer already existed. Customers just couldn't reach it without a human.
Remedial Health is a digital operating platform for pharmacies, hospitals, and clinics across Nigeria — used by 250,000+ healthcare providers to buy medicine, manage inventory, and run their business.
As adoption grew, so did support volume. But when we pulled the tickets, most weren't hard problems. They were the same handful of questions, over and over — and every one of them had an answer already sitting inside the platform.
That gap cost the business response time, and it cost pharmacy owners — people whose real job is keeping shelves stocked for patients — hours they didn't have. The question we set out to answer: what if every provider had an assistant that could resolve the routine stuff instantly, and only pull in a person when it actually mattered?
Discovery
What we learned before we designed anything.
We spent the first few weeks in support interviews, ticket data, and chat logs — not Figma. Three findings shaped everything that followed.
1.
People wanted outcomes, not conversation
Nobody typed to chat — they typed to finish a task. That ruled out an open-ended chatbot from day one.
2.
The data already existed — access was the problem
Providers had to manually cross-reference several dashboards to answer something as simple as "what should I reorder today?" The insight was there; the interpretation wasn't.
3.
"Out of stock" was where journeys died
When a searched medicine wasn't available, the conversation — and the sale — usually ended there, even when a clinically equivalent alternative was one tap away.
The Bet
We could have shipped a support chatbot. Instead, we built a business assistant.
One interface for products, orders, inventory, returns, and pharmaceutical knowledge — instead of five. That also meant being disciplined about what Remy would never touch.
Design for outcomes, not dialogue
Every reply should move the task forward — structured cards and one-tap actions beat paragraphs.
Recommend, don't just report
"You have three products running low" is information. "Here are your recommended reorder quantities" is a decision made easier.
Know what not to automate
Payments, disputes, and anything Remy wasn't confident about handed off to a human, with full context — so nothing had to be repeated.
The Trust & Ethical Framework
Where our AI assistant helps, and where a human takes over was a priority for us
What Remy AI can do
Search medicines, recommend alternatives and initiate order placement
Track deliveries and answer questions about orders
Analyse user inventory (Product stock, expenses, Low stocks, expiring products
Explain regulations, new laws pharmaceutical and health laws and news.
🧑💼
Hands off to a human
Manual payments and escalated payment disputes
Overriding secure workflows
Low-confidence or financial decisions
Process & Trade-offs
Getting the conversation model right.
The biggest design risk wasn't the AI itself — it was assuming we already knew how people would talk to it. We tested three shapes.
Fully open chat
Every reply should move the task forward — structured cards and one-tap actions beat paragraphs.
Menu-driven
Predictable, but rigid — no room for a real question.
Shipped
Hybrid
Suggested actions to start, natural language once in motion, structured UI throughout.
Four flows usability testing reshaped
Order tracking
Asked too much upfront
Cut to the one detail needed to retrieve a result
Product search
Ended at "out of stock"
Rebuilt to recommend alternatives instead
Inventory answers
Presented as raw data
Turned into recommendations — "reorder these three"
Escalation
Felt like a dead end
Carries full context into the human handoff
The Solution
Remy shipped supporting five jobs — through one conversational surface
Support, procurement, inventory, healthcare knowledge, and business operations. Three moments carry the most weight.
01 · AI-first support
Starts with a tap, not a blank box
Instead of a blank chat box, Remy opens with suggested actions based on the most common requests, so customers don't have to think about how to phrase a question just to get started
Track an order
Find a medicine
Check inventory
02 · Product discovery
Search that doesn't dead-end
Search returns structured product cards with price, availability, and supplier info. When something's out of stock, Remy proactively recommends a clinically equivalent alternative — turning what used to be an abandoned search into a completed sale.
03 · Inventory intelligence
The hero of the project — data becomes decisions
Ask "what should I reorder today?" and Remy cross-references current stock, sales velocity, and supplier availability to return a prioritized, actionable list — instead of a dashboard the provider has to interpret themselves.
Before — dashboard
"Where is my order?"
"Do you have another brand?"
"What should I restock?"
"Can these medications be returned?"
After — Remy conversation
Reorder Metformin — 40 units
Selling fast, 3 days of stock left
Reorder Amoxicillin — 25 units
Below reorder threshold
04 · Human escalation
Context travels with the conversation, not just the customer
A provider reports a loan payment that isn't reflecting in their wallet. Remy asks only for what it needs — an order date, then a receipt — before recognizing this is a financial dispute outside its authority. It hands off to a support agent with the receipt already attached, so the provider never explains the issue twice.
06 · Proactive intelligence
Remy doesn't wait to be asked
Inventory intelligence isn't only a response to a question. Remy also sits inside the inventory screen itself, watching sales and stock levels in the background, and surfaces an insight the moment it's worth acting on — before the provider has clicked anything.
Outcomes & Impact
The number that mattered less than expected was automation rate.
The one that mattered more was how often people trusted what Remy told them enough to act on it without double-checking.
62%
Reduction in customer support tickets — the primary KPI
Faster
Resolution on remaining cases — agents no longer buried in repeats
Higher
Product discovery, via substitute recommendations
Better
Inventory decisions, acting on recommendations vs. reports
What I Learned
What I'd do differently.
Given more runway, I'd add these — all scoped during the project, and cut for the 16-week deadline. A trade-off, not an oversight.
Conversation memory
Predictive inventory alerts
Multilingual support
"Customers don't care about AI, they care about getting their job done."
Every debate about adding a capability, or another turn of conversation, came back to one test — does this get a healthcare provider to a decision faster, with more confidence? If yes, we built it. If not, we cut it.
Remy wasn't designed to replace the support team. It was designed to remove the friction that shouldn't have needed a person in the first place, so the humans on both sides of the conversation could spend their time on what actually required judgment.
