The problem
Small businesses in Egypt sell and support customers inside DMs. A clinic, a real estate office or a training center gets the same questions all day across WhatsApp, Instagram, Messenger and comments. The answers already exist somewhere: a price list, a PDF, a website page, or the owner's head.
The pattern was the same everywhere: replies arrive hours late, leads go cold, and hiring someone just to answer "how much?" isn't affordable. Most tools available were either rule-based flows that break on the first unexpected question, or products priced far above what an SME can pay, with weak Arabic and dialect support.
My role
I co-founded Agentech Assistant with an engineering partner. He owned the backend and AI core; I owned the product: what we build, for whom, how it's structured, how it looks and feels, and how it's priced. I also built frontend parts myself with AI coding tools, which removed most of the back-and-forth between design and engineering.
The goal: a shop owner with no technical background should be able to set the assistant up alone, trust it with real customers, and pay for it monthly in local currency.
The process
01 · Discover
We started from the businesses around us and the conversations they were already having with customers. Three things stood out:
- The questions repeat. Most incoming messages ask about prices, availability, location and how to book.
- The channel is the customer's choice, not the business's. The same business gets messages on four or five platforms, and each one lives in a separate app.
- Owners don't trust a bot they can't control. Every conversation about automation came back to one fear: "What if it says something wrong to my customer?"
02 · Define
That turned into one problem statement: help small teams answer every customer, on every channel, in seconds, without losing control of what's said.
I set four product principles that every later decision was checked against:
- Answers come from the business's own content, never from the model's imagination.
- A human is always one tap away.
- Setup is self-serve. If a business needs a call with us to start, we failed.
- Plain language everywhere. No AI or API jargon in the interface.
03 · Design
I designed the product around the four moments every business goes through, in order.
01 · Connect
Channel onboarding
- Screenshot needed
public/work/agentech-assistant/connect-01.png - Screenshot needed
public/work/agentech-assistant/connect-02.png - Screenshot needed
public/work/agentech-assistant/connect-03.png
The first moment of value is seeing the assistant reply on a real channel, so connecting channels is the core of onboarding, not a buried settings page.
- One hub for every channel: website widget, WhatsApp, Instagram, Messenger, Telegram and comment replies, each with a clear connected / not connected state.
- Official WhatsApp connection, explained in plain language instead of API terms.
- Trade-off: first connection asks only for what's needed to go live. Everything else can be tuned later, so first-time owners don't stall.
02 · Teach
Knowledge sources
- Screenshot needed
public/work/agentech-assistant/teach-01.png - Screenshot needed
public/work/agentech-assistant/teach-02.png - Screenshot needed
public/work/agentech-assistant/teach-03.png
Principle 1 lives here. The assistant answers from what the business teaches it.
- Multiple source types: documents, website URLs and plain text, so owners use what they already have.
- Answers grounded in sources, with citations in the dashboard so the owner can check where a reply came from.
- Assistant settings in business language, not prompt-engineering language.
03 · Handle
Unified inbox & human handoff
- Screenshot needed
public/work/agentech-assistant/handle-01.png - Screenshot needed
public/work/agentech-assistant/handle-02.png - Screenshot needed
public/work/agentech-assistant/handle-03.png
The flow I'm proudest of, because it answers the fear we heard most. Owners only hand customers to an AI when they know they can take back control.
- One inbox for all channels, so a team doesn't jump between five apps.
- AI replies first, a teammate takes over in one tap, inside the same thread with the full history.
- Clear ownership of each conversation, so the team always knows whether the AI or a person is replying.
04 · Grow
Campaigns & packages
- Screenshot needed
public/work/agentech-assistant/grow-01.png - Screenshot needed
public/work/agentech-assistant/grow-02.png - Screenshot needed
public/work/agentech-assistant/grow-03.png
- WhatsApp broadcast campaigns to reach existing customers, with replies flowing back into the same inbox.
- Four billable products (Assistant, WhatsApp, WhatsApp outbound, Comments) combined into packages from Starter to Business, plus a WhatsApp-only plan for businesses that come for campaigns first.
- Prepaid plans in EGP and SAR, matching how SMEs in Egypt and Saudi buy software.
04 · Build & ship
Design and build ran side by side. My co-founder built the multi-tenant backend, the RAG pipeline and the channel integrations; I designed the dashboard and widget and built frontend parts directly with Claude Code.
- Multi-tenant from day one, so every business's data stays isolated at the database level.
- Arabic RTL as a core requirement, not a translation layer added at the end.
- Becoming a Meta Tech Provider let us offer the official WhatsApp connection instead of unofficial workarounds that get numbers banned.
05 · Launch & iterate
Real usage changed the product more than any plan did:
- WhatsApp campaigns became the most requested feature, and some clients wanted only that. This shaped the WhatsApp-only plan and moved campaigns closer to the center of the product.
- Big campaigns used up AI replies fast, because every customer reply to a broadcast consumed the plan's allowance. I'm reworking how packages count and limit replies, since raising the allowance would break the margin.
- Many trial users signed up and left before connecting a single channel. They never reached the moment where the assistant answers a real customer.
Results
- 80%+ of incoming messages are answered by the assistant without a human.
- First response time for client businesses went from hours to seconds.
- 10+ businesses pay monthly, across sectors like real estate, education and clinics, reached in 3–6 months from design start to first paying client.
- 6 channels run from one inbox, in Arabic and English.
What I learned & what's next
- In AI products, the first real answer is the activation event. Everything before it should be as short as possible. Next: a guided setup checklist and lifecycle emails that bring back users who signed up but never connected a channel.
- Usage beats the roadmap. Campaigns weren't the plan; customers made them the product's front door.
- Pricing is product design. How a package counts usage shapes behavior as much as any screen does.
Stack & tools
Design & product
- Figma
- FigJam
- ClickUp
- Notion
Frontend
- Next.js (dashboard & website)
- Embeddable chat widget
- Arabic RTL
- Claude Code (AI-assisted build)
Backend & AI
- Java / Spring Boot
- RAG pipeline
- Qdrant (vector database)
- PostgreSQL with row-level security (multi-tenant)
- Scribe v2 (audio transcription)
Channels
- WhatsApp Business API (Meta Tech Provider)
- Instagram DM
- Messenger
- Telegram
- Facebook & Instagram comments
- Developer API
Infrastructure
- Coolify (hosting)
- Public widget CDN