How we automated 68% of our own support — the ONNOA & LOVETALES story

Before Sondercase was a product, it was our own problem. We run two hardware D2C brands — ONNOA (BLE camera accessories) and LOVETALES. Physical products, EU customers, DHL and DPD in the mix, warranty cases with real diagnostics. This post is the honest version of how our support ran before, what we changed, and what the numbers look like now.

What our support actually looked like

On paper we “did support ourselves.” In practice, that meant 40+ hours a month of founder time spent on tickets that all felt identical:

• WISMO — “Where is my order?” We’d copy a tracking number out of the shop, check the DHL or DPD portal, and type out the same status reply. Again.
• Returns — check the order, check the return window, generate a label, adjust the order, reply.
• Warranty — the worst one for a hardware brand: figure out whether it’s a user error or a defect, check the purchase date, decide repair / replace / refund, then execute that decision across shop and shipping.

Every single case required opening five tools: the inbox, the shop backend, the carrier portal, the shipping tool, and whatever spreadsheet held the warranty state. None of the work was hard. All of it ate the hours we needed for products and marketing.

What we tried first — and why it didn’t stick

We did what everyone does. We wrote macros in the helpdesk. We looked hard at the established stack: a helpdesk with AI drafts on top, plus a returns portal next to it.

The problem is structural, and you only see it clearly after living with it:

• AI drafts don’t act. A drafted reply to “where is my parcel?” still needs a human to look up the tracking status, or it’s a template that says nothing. The typing was never the expensive part — the lookup and the decision were.
• Portals don’t resolve. A returns portal takes the request in nicely, but the judgment calls — is this in the window, is it warranty, do we refund or replace — still land on a person, now split across one more tool.

The realization that became Sondercase: the ticket isn’t the work. The actions behind the ticket are the work. Anything that only handles the text layer automates the cheap 20% and leaves the expensive 80% untouched.

What we changed

So we built the thing we couldn’t buy: a system that reads the case, pulls the real order and tracking data itself, makes a decision under rules we set, and then executes the backend actions — refund in the shop, return label, replacement order — with the customer reply as the last step, not the only step.

Rolling it out on ONNOA took 14 days from zero to production, and the sequence matters more than the speed:

1. Everything started in draft mode. Every reply and every store action was proposed with a confidence score and the reasoning shown. We approved or corrected. Nothing went out that we hadn’t allowed.
2. We switched categories to automatic one at a time. WISMO first — it’s high-volume and low-risk, and the tracking data makes right answers checkable. Then standard returns inside the return window. Warranty diagnostics stayed in approval mode the longest.
3. Escalation was a feature, not a failure. Anything the system wasn’t confident about landed with us, with the full case context attached — not a guess sent to a customer.

The numbers today

• 68% of support cases at onnoa.de resolve end-to-end without a human touching them. We count strictly: a case is “automated” only if nobody on our side read it, decided it, or clicked anything — reply, backend actions and all. Cases we merely approve don’t count toward the 68%.
• €4,247 per month saved across tool costs and the founder hours we no longer spend — that’s the conservative version of the math, valuing our time at what we’d pay an ops hire, not at founder opportunity cost.
• The remaining ~32% is exactly what should reach a human: unusual warranty judgment calls, goodwill decisions, angry customers, edge cases. That’s the 32% where a founder actually adds something.

What we’d tell other founders

Even if you never use Sondercase, three things transfer:

1. Measure cost per resolved case, not tickets per month. Tickets feel free when founders answer them. They aren’t. (We did the full math in what a support ticket really costs.)
2. Automate categories, not the inbox. “AI for support” fails as one big switch. WISMO, standard returns, address changes — each category has its own risk profile and its own point where you can trust automation. Start where wrong answers are cheap and verifiable.
3. Never accept automation that stops at the text layer. If a tool can’t execute the refund, the label and the replacement order, you’ve automated typing — the part that was never the problem. (WISMO is the clearest example: here’s that playbook in full.)

We built Sondercase for ourselves first, and ONNOA and LOVETALES still run on it every day. If you want to see what it would do with your store’s cases, the live demo resolves a real one in front of you — no signup, just your store domain.