AI Integration with Tokopedia and Shopee: A Practical Guide

Most Indonesian online sellers run on Tokopedia and Shopee. Both platforms have APIs. Both also have meaningful restrictions on what you can do automatically. AI integration on these channels is genuinely useful — but only if you understand what’s actually allowed before you start building.

Here’s the practical guide we wish someone had given us before our first integration.

What the platforms expose (and don’t)

Tokopedia (Open API)

What you can pull:

• Order data (statuses, line items, customer info per order)
• Product catalog (your listings, stock levels, prices)
• Logistics info (shipping providers, tracking)
• Sales reports

What you can push:

• Order status updates
• Stock adjustments
• Price changes
• Product creation and updates

What’s restricted:

• Direct customer messaging is gated — you can read but interaction goes through Tokopedia Chat (separate API, lower throughput limits)
• Marketing campaign automation is mostly off-limits

Shopee (Open API)

Roughly similar surface area to Tokopedia, with platform-specific quirks:

• Order and product APIs are mature
• Chat API exists but rate-limited
• Voucher and promotion APIs are more permissive than Tokopedia’s
• Webhook reliability has improved a lot since 2024

Both platforms throttle aggressively. Plan for 100–500 requests per minute as a typical ceiling, and design assuming you’ll occasionally get rate-limited.

High-leverage AI use cases

1. Order intake and tagging

Pull every order across both platforms into one system. AI tags each order with metadata your platform doesn’t provide: customer segment, urgency level, special handling flags, fraud likelihood scores.

Useful especially when you’re selling on multiple channels and want a single operations queue rather than two admin panels open simultaneously.

2. Listing optimization

AI generates and optimises product titles, descriptions, and bullet points based on what’s converting on each platform. Tokopedia and Shopee favour different keyword densities and formatting; AI handles the per-platform variation automatically.

Worth doing because manually tuning 100+ listings is impossible, but the gap between an OK listing and a great one routinely doubles conversion.

3. Customer message drafting

Read incoming chat messages, draft contextually appropriate responses, route to a human for approval before send. Both platforms allow this pattern (read + suggest + human-sends), even when fully automated chat is restricted.

The implementation is more nuanced on these platforms than on standalone WhatsApp because of the rate limits, but the value is comparable.

4. Pricing and stock decisions

AI reads sales velocity, competitor prices (where the platforms expose them), and inventory levels. Suggests price changes to a human approver. Almost never fully automated — small price errors can be expensive.

5. Returns and complaint triage

Pull returns and complaints, classify by likely cause (product defect, shipping issue, customer error, fraud attempt), suggest response. Big operational value because returns on these platforms carry brand penalties — getting the response right and fast matters more than people think.

What you shouldn’t try

Three patterns that consistently fail:

• Bulk auto-response in customer chat. Both platforms penalise sellers who send template responses without actually reading the message. AI-drafted, human-sent works; full automation doesn’t.
• Aggressive scraping of competitor data. Tools that scrape pricing data from competitors’ listings violate platform terms in subtle ways. Even when they work technically, they’re a brand and account risk.
• Auto-generating fake review responses. Don’t.

Implementation realities

Things that bite first-time builders:

• API access requires merchant verification. Both platforms gate API access behind seller-status checks. Allow 1–2 weeks for approval.
• Rate limits are not negotiable. Plan for them. Build queues and retry logic.
• Webhook delivery isn’t guaranteed. Combine webhooks with periodic polling for critical data (orders, status changes).
• Schema changes happen. Both platforms occasionally change API responses without strong deprecation warnings. Build with abstraction layers so changes touch one file.

Architecture pattern that works

The shape we recommend:

1. Sync layer: pulls all order/product/chat data from both platforms into your own database (D1, Postgres, whatever). Runs every 5 minutes.
2. AI layer: reads from your database, applies whatever AI logic, writes recommendations or drafts back to your database.
3. Human approval layer: a simple admin UI where humans review AI suggestions and approve.
4. Push layer: when humans approve, pushes the action back to Tokopedia/Shopee via API.

This keeps the AI from talking directly to the platforms (which avoids rate-limit and reliability issues) and gives you a single operational view across channels.

Cost ballpark

For an SME selling Rp 500 juta–5 miliar/year on Tokopedia and Shopee combined:

• Initial integration build: Rp 50–150 juta
• Monthly running cost: Rp 1.5–4 juta (hosting + API usage + minor maintenance)
• Payback: typically 3–6 months when measured against operations time saved

If you’re trying to figure out which AI integrations make sense for your specific Tokopedia/Shopee operation, an hour of conversation usually clarifies it. We do those at no cost.