Adding AI to a Decade-Old Shopify App Without Breaking It

How We Added AI Safely

• We scoped this build deliberately. Early conversations imagined something much bigger — an assistant with broad access to funnel creation, statistics, and every corner of OneClickUpsell. We pulled that scope back on purpose, and it’s a big part of why the first release actually worked.
• We started narrow and expanded from a position of strength. The first version did three things well: surface recent performance stats, flag unused opportunities, and promote Zipify’s newer post-purchase feature. That restraint wasn’t a limitation — it gave merchants something reliable from day one, and gave us a stable foundation to build on.

• We split the work across four specialized agents. An orchestrator reads each merchant request and routes it. A data agent handles the numbers. A strategist agent shapes recommendations. A fourth agent manages database queries. Each one owns a single job instead of asking one model to understand the entire product.
• We designed the Assistant as a standalone module that integrates with the OneClickUpsell app rather than living inside it. This kept it isolated from production by default, minimizing any risk of impact on the thousands of merchants who rely on OneClickUpsell daily. The Assistant runs on its own database and its own infrastructure, fully separate from Zipify’s core application. A daily sync feeds it exactly what it needs — store statistics, funnel performance, health signals — into an environment built specifically for this purpose. That isolation paid off twice over: it protected the core product, and it let us iterate without waiting on anyone else’s release cycle.

• We caught the hallucinations before they became a trust problem. A ten-year-old product with dozens of overlapping features is exactly the kind of environment where an AI model can be confidently wrong — and early testing showed that it could be confidently wrong. We addressed this through extensive prompt tuning, giving each agent a precise, tested understanding of the features it owns, and backed it with end-to-end monitoring that traces how requests move between agents.
• We let real conversations shape the roadmap. Every week, we reviewed live chats to see what merchants actually wanted. One pattern surfaced fast: merchants were treating the Assistant like a support line, and “build it for me” kept coming up around funnel creation. We took that straight to the roadmap — funnel creation through the Assistant is now in active development.

The same instinct guides what comes next. Agents are gradually gaining the ability to query more complex, production-mirrored data, but only in isolated, closely monitored stages. Live production access is something we’ll earn, not assume.