How one agentic workflow became shared infrastructure across our portfolio

Ten months ago, if you'd asked someone on the Releasit support team what a normal day looked like, the honest answer was: behind. Conversations were coming in faster than anyone could read them properly, let alone answer them well. The obvious fix was to hire more people. We tried something else first.

Cody is what we built: an AI agent that sits inside Releasit's support inbox, reads each conversation, including any screenshots a customer sends along to show what’s going wrong and decides what it can safely answer versus what genuinely needs a person to look at it. Nothing exotic about the idea. What surprised us was how well the same approach held up once we pointed it at completely different problems.

Illustrative example. Store names and data shown are fictional.

The logic is simple enough to say in one sentence: read the conversation, decide whether you actually know the answer, and if you don't, don't guess. Today, Cody resolves 65.8% of Releasit's 63,000 conversations directly. Separately, first contact now lands in about ten seconds on average, down from roughly three minutes before, that's how fast someone hears back. When it can't resolve something, it doesn't just drop the conversation back in the queue empty-handed. It gathers what a person would need first, setup steps, diagnostic details, and attaches that before handing over. That one detail, doing the groundwork before a human even opens the ticket, turned out to matter more than we expected.

Underneath, it's simpler than it sounds and more particular than “AI reads your support inbox” suggests. Cody runs on retrieval, not improvisation: each company's support team builds a knowledge base of the questions their merchants actually ask, and Cody's job is to find the closest match and answer from that, not invent something plausible-sounding. When a question is niche, ambiguous, or clearly needs a person's judgement, it hands over rather than guesses. A human can also step into any conversation directly at any point, which automatically triggers the same handover.

Since Cody went live, we estimate it has saved around 2,746 hours (~343 working days) of support time in total, and about 382 hours last month alone. Methodology: resolved sessions × 4 minutes, our estimate of how long a human takes to actually resolve the same kind of conversation (resolution time, not the first-contact time above). Since Cody resolves these without any human involvement, that full 4 minutes counts as saved per session. 

We didn't stop there. The same core architecture was deployed across other Releasit applications, and adapted for Wholesale Helper, an ecosystem with significantly different support flows and operations.

Support for B2B e-commerce spans two ends of the spectrum: first-time store owners setting up basic wholesale forms, and enterprise brands running multi-currency tiers, custom NET-30 checkout rules, and complex account-level discount matrices. Merchants rarely describe these intricate configurations in text; they send screenshots of broken cart drawers, failing dynamic pricing tables, or layout conflicts.

To bridge that gap, we expanded Cody with computer vision capabilities designed to analyse visual context directly. Cody now parses raw UI screenshots and diagnoses misconfigured pricing tiers or layout bugs without requiring the merchant to write a single line of explanation.

That multimodal vision layer didn't stay isolated to Wholesale Helper. It retrofitted every Cody deployment group-wide, giving the original Releasit agent the ability to also diagnose issues visually.

Illustrative example. Store names and data shown are fictional.

This loop captures the core logic of our business model: build foundational architecture once, adapt it deeply to the unique operational setups of each business, and immediately cross-pollinate those breakthroughs group-wide. By solving a high-stakes B2B challenge for Wholesale Helper, we unlocked a compounding advantage for our entire portfolio.