How we saved 40 hours a month with AI: a real e-commerce case study

Theory about AI in business is everywhere. Here are concrete numbers from a real deployment at a Czech online shop doing 2,000 orders per month.

Starting point

A small shop selling fashion accessories, a team of 4. The biggest time sinks:

• Customer support — 15 hrs/week (repeating questions about shipping, returns, sizes)
• Product descriptions — 8 hrs/week (new products, translations, SEO copy)
• Reporting — 5 hrs/week (manually pulling data, pivot tables in Excel)

Phase 1: Customer support with Claude

Instead of a classic chatbot, we deployed Claude via the API with context of the terms and conditions, shipping price list, and FAQ.

const systemPrompt = `Jsi zákaznický asistent e-shopu FashionCZ.
Odpovídej POUZE na základě přiloženého kontextu.
Pokud si nejsi jistý, napiš: "Přepojuji na kolegu."

Kontext:
${obchodniPodminky}
${cenikDopravy}
${faq}
`;

Key decision: Claude over GPT because of lower hallucination rates. In customer support you can't afford fabricated information about returns.

Results after 3 months

• 72% of queries resolved automatically (no human involvement)
• Average response time: 8 seconds (vs. 4 hours before)
• CSAT score: 4.2/5 (vs. 3.8 before deployment)
• Saved: 10 hours/week

Phase 2: Product descriptions

Every new product needs: name, description (3 lengths), SEO meta description, alt text for photos. Used to be 20 minutes per product, now it's 3 minutes — human review plus tweaks on the AI output.

Workflow:

1. Product photo + basic specs → Claude (vision)
2. Claude generates all the copy at once
3. Human reviews, adjusts tone, adds unique details
4. Batch upload into the CMS via API

const prompt = `Na základě fotky a parametrů vytvoř:
1. Název produktu (max 60 znaků, SEO optimalizovaný)
2. Krátký popis (max 160 znaků pro meta description)
3. Střední popis (2-3 věty pro kartu produktu)
4. Dlouhý popis (150-200 slov pro detail produktu)
5. Alt text k fotce

Parametry: ${JSON.stringify(produktData)}
Tón: ${tonOfVoice}
`;

Results

• 20 min → 3 min per product
• More consistent tone across the catalogue
• SEO metrics: +23% organic traffic in 2 months
• Saved: 6 hours/week

Phase 3: Automated reporting

Instead of manual data exports and pivot tables, we built a simple script that:

1. Pulls data from the Shopify API (orders, revenue, products)
2. Sends it to Claude with an analysis instruction
3. Output: a structured Markdown report plus recommendations

const weeklyReport = await generateText({
  model: 'anthropic/claude-sonnet-4-5-20250514',
  prompt: `Analyzuj týdenní data e-shopu a vytvoř report:
    - Top 5 produktů podle tržeb
    - Trendy oproti minulému týdnu
    - Anomálie (neobvyklé nárůsty/poklesy)
    - 3 konkrétní doporučení pro další týden

    Data: ${JSON.stringify(salesData)}`,
});

Results

• Report done in 30 seconds (vs. 5 hours by hand)
• AI flagged a seasonal pattern the team had been missing
• Saved: 5 hours/week

Total impact

| Metric | Before AI | After AI | Change | |---------|---------|-------|-------| | Time on support | 60 hrs/mo | 20 hrs/mo | -67% | | Time on descriptions | 32 hrs/mo | 6 hrs/mo | -81% | | Time on reporting | 20 hrs/mo | 2 hrs/mo | -90% | | Total | 112 hrs/mo | 28 hrs/mo | -75% |

AI costs: ~$150/month (Claude API). The time saved moved over to marketing and building new products.

Lessons learned

1. Start with the most painful problem — not the trendiest one
2. Human-in-the-loop is a must — AI generates, humans review
3. Claude for accuracy, GPT for creativity — use both
4. Measure everything — without numbers you don't know if it's working
5. Iterate on prompts — the first version is never the final one

AI isn't magic. It's a tool that works when you know what you want from it.