Seven months, one AI agent, and an attempt to rethink selling on Instagram

01

The problem I saw from the inside

Before NovaSale, I built Balace, an Instagram-based clothing brand. I was responsible for more than just the product — I managed the entire sales operation: hiring, sales scripts, workflows, and team training. I knew exactly when to send a payment link and how to collect the correct shipping details for Nova Poshta.

I also knew where everything could break down.

A manager called in sick — sales slowed to a halt. A new employee joined — another week of onboarding. Human error meant forgotten customers, lost conversations, and shipping mistakes. The entire process lived in people's heads and disconnected tools rather than in a reliable system.

What stayed with me wasn't some abstract theory about lost sales. It was the small problems that kept repeating every week: a customer ready to buy, but the reply arrived too late; or a shipping label with an incorrect name or pickup location, causing a delivery to fail.

Each mistake seemed minor on its own. Together, they gradually eroded customer trust.

That's when I realized the problem wasn't simply the absence of a CRM. The real problem was that selling through messaging apps still required a human manager at almost every step — from the first "Hello" to the final delivery status.

02

Why a traditional CRM isn't enough

Traditional CRMs are good at storing customer information and orders. They're much less effective at actually selling when a customer is already messaging a business on Instagram or Telegram.

At that moment, a business doesn't need another customer record. It needs an intelligent conversation that can find the right product, collect order details, process payments, and arrange delivery — without losing context.

"I didn't want to build another CRM. I wanted to create a system that could independently handle the work of a sales manager."

That's why NovaSale isn't just another chatbot with predefined responses. Instead of letting the AI model make every decision, I moved critical operations into controlled tools. The agent can communicate naturally, but product searches, price calculations, order creation, and payments follow predefined business rules.

This reduces the risk of errors and makes the system's behavior more predictable. It doesn't eliminate mistakes entirely — the model can still misunderstand customer intent or select the wrong tool.

Instagram Direct and Telegram operate through a shared orchestration layer. One agent. One customer history. Multiple communication channels. A human manager steps in when a situation becomes too complex — with the full conversation context already available.

03

Seven months from idea to product

The development process involved an intense series of decisions. Seven months is the actual time it took to go from the initial idea to the product stores can use today.

The hardest part wasn't designing the interface or connecting an AI model. It was making the business logic work reliably in real conversations, completing the Meta Verified Tech Provider process, and securing the platform well enough for real businesses to trust it.

1. Stage 01 AI agent architecture, sales scenarios, and the first foundations of the CRM and product catalog
2. Stage 02 Orders, payments, shipping, and a unified processing pipeline for Instagram and Telegram
3. Stage 03 Meta Verified Tech Provider, API limitations, phantom PSIDs, and channel stability
4. Stage 04 API key encryption, data protection, backups, and preparing the platform for real stores

Approximate development priorities over seven months — not a commit-by-commit history.

Building the business logic

The core components included order slot-filling, tool orchestration, payment integrations, and Nova Poshta shipping. But business logic took significantly longer than expected. The agent needed to retain previously collected information, escalate complex cases correctly, and complete the entire sales process — rather than simply generating convincing responses halfway through a conversation.

The Meta challenge

One particularly demanding stage involved the Meta ecosystem. Verification procedures, permissions, API restrictions, business account requirements, and obtaining Verified Tech Provider status all introduced additional complexity. Without meeting these requirements, a product like NovaSale quickly runs into platform limitations when trying to support Instagram at scale.

At the same time, issues such as phantom PSIDs and other silent bugs emerged — problems that could cause messages to disappear without obvious errors.

The less visible challenge: security

Another major challenge was something that rarely appears in polished startup case studies: security. API keys, encryption, product protection, and backups all had to be handled properly. Store data and access to connected messaging channels couldn't be managed like a personal side project.

It's not the most visually exciting part of building a product. But without it, there's no point in inviting real businesses onto the platform.

When AI doesn't follow the rules

There were also the classic AI development problems. Sometimes the model invented discounts. Sometimes it lost products from the shopping cart. The first version of the orchestration system contained conflicting logic. Fixing these issues required weeks of rewriting, testing, and refinement.

Connecting Claude or Gemini is only the beginning. The real challenge is making an AI agent consistently complete sales while operating within strict business rules.

Technology stack

Frontend

Next.js

Backend

FastAPI

AI

Claude, Gemini

Database

PostgreSQL

Integrations

Instagram, Telegram, WayForPay, Monobank, Nova Poshta

04

What NovaSale actually does

The process is straightforward: Message → Product → Order → Payment → Shipping label → Delivery status. Three layers operate within one platform — conversations, storefront, and CRM — with no disconnected collection of services.

05

What's next

NovaSale is currently in its early adoption stage. The product is already being used with real stores, but I'm deliberately avoiding impressive-looking metrics without proper context. What's more important is being transparent about what we're actually testing — the percentage of conversations handled without human intervention, the accuracy of collected order information, the stability of Meta's APIs, and the situations where human involvement remains necessary.

The question I'm still trying to answer: can AI genuinely eliminate most of the manual work involved in conversational commerce — or will there always be a layer that requires human judgment? So far, the answer is mixed. The agent handles repetitive tasks well. Complex escalations, unusual customer objections, and unexpected logistics situations remain more challenging.

That's the balance I'm continuing to refine. The goal isn't to promise that AI will completely replace sales managers. It's to reduce the number of hours people spend on repetitive tasks so they can focus on situations where human judgment actually matters.

For me, NovaSale isn't just another project in my portfolio. It's a continuation of my entrepreneurial journey. First, I built a sales operation for my own clothing brand. Then I started turning that experience into a product other businesses could use. The next step is expanding real-world production use cases and finding more reliable answers to the limits of automation.

Building something similar? I help founders turn complex ideas into working digital products — from strategy and MVP development to automation and launch.

Discuss a project ↗ novasale.app