YundaDeskSwitching support software starts with picking a system, but the part that actually trips teams up is the move itself. Owners want the new system to catch customers faster. Agents worry about losing old conversations, macros and customer context. Ops teams watch whether entry points like WhatsApp, email, the website widget, Instagram and TikTok stay live through cutover day.
A good migration lets the new system inherit what the team already knows: the knowledge base feeds the AI agent, historical conversations become test cases, and every channel lands in one workspace. Use this checklist in order.
Define the migration scope: what moves and what gets archived
Do not start with “move everything.” Full migration sounds safe, but it often drags expired policies, discontinued products, unused channels and duplicate customer records into the new system.
Split assets into three groups:
| Type | Recommended handling | Examples |
|---|---|---|
| Must move | Import and verify before launch | Current policies, best-selling SKU FAQs, active customer records, conversations from the last 6 to 12 months |
| Archive | Keep searchable, but exclude from default AI knowledge | Old campaigns, discontinued SKUs, historical complaint files |
| Drop | Do not migrate; keep only in old-system backup | Expired auto-replies, duplicate tags, unused macros |
The goal of this step is a clean starting point for the new system. For AI support, answer quality depends on what the system can read — old wrong scripts moved into the new knowledge base just make the new system repeat old mistakes faster.
Switching Support Software: one bad experience sends customers elsewhere
Move the knowledge base first: give AI evidence from day one
In a support software migration, the knowledge base should come before interface polishing. For an AI agent, the knowledge base is not a document cabinet. It is the evidence used to answer customers.
Organize it in four layers:
- Policy layer: shipping times, delivery countries, return conditions, warranty and customs notes.
- Product layer: sizing, materials, compatibility, usage steps, installation and common misunderstandings.
- Process layer: order lookup, address changes, shipping nudges, cancellations and what information to collect for after-sales requests.
- Boundary layer: for refunds, compensation, price changes and escalated complaints, AI may calm the customer and collect context, but must hand off for human approval.
The import can be simple: upload documents, crawl website pages, and add manual Q&A. Formatting does not need to be polished, but every item should be traceable, testable and revertible when something is wrong. In YundaDesk, the knowledge base is the foundation that feeds the AI agent. Later, when agents correct AI or answer what AI missed, those items become learning suggestions that an owner confirms before they take effect.
That is the difference between “gets smarter over time” and ordinary auto-replies: AI does not quietly rewrite business rules by itself — the team turns experience into skills first, then a human approves them. For the full mechanism, see teaching AI that gets smarter.
Use historical conversations as a test set, not just a backup
Many teams migrate historical conversations only so agents can search them later. The bigger value is testing whether the new system can handle real customers.
Export three groups of conversations:
- High-frequency questions: tracking, shipping, sizing, discount codes and return conditions.
- High-risk questions: refunds, compensation, complaints, review threats and platform disputes.
- Multilingual questions: sample beyond English based on target markets, such as Spanish, Portuguese, Arabic or Thai.
Before launch, feed these questions to the AI agent one by one. Check whether it answers from the knowledge base; whether it hands off when it does not know; and whether refund, compensation or price-change requests always go through approval. What matters most is whether it made anything up, overstepped, or handled something that should have gone to a human — sounding human is the easy part.
Reconnect channels with a schedule: avoid customer-facing gaps
For cross-border sellers, support is never just one email inbox: website widgets, custom API, email, WhatsApp, Telegram, Messenger, Instagram, TikTok, LINE, WeChat, VKontakte, Zalo and YouTube can all be customer entry points. The risk is simple: an entry point disconnects before the new one is ready.
Manage channel cutover with a table:
| Channel | Owner | Cutover window | Acceptance check |
|---|---|---|---|
| Website widget | Ops or engineering | Low-traffic hours | Start a test chat and confirm it enters the new workspace |
| IT or support lead | After forwarding or MX changes | Test inbound and outbound mail from an external address | |
| WhatsApp / Messenger | Social lead | Same day as reauthorization | Test DMs, notes and customer-profile merge |
| Content team | Ad low point | Test whether comments and DMs enter the same queue |
Every channel needs a verifiable cutover window: connect the new system, verify it, then disconnect the old one — never the other way around. After each channel is connected, confirm that messages land in one workspace, identities can merge, and agents can see context. For the broader logic, read omnichannel inbox explained.
Set the AI/human boundary during migration
Teams often ask “can AI answer this?” and forget “what must never be automated?” Set that boundary before launch.
Use three tiers:
| Risk | What AI can do | Where humans must step in |
|---|---|---|
| Low | Answer tracking, sizing, policy and usage questions | When the customer explicitly asks for a person |
| Medium | Answer first and collect information, such as address changes or shipping nudges | Conflicting information, unhappy customer, request outside policy |
| High | De-escalate, summarize and collect order context | Refunds, compensation, price changes, complaint escalation |
The AI agent faces customers and can answer 24/7 from the knowledge base. Yuna is the merchant-facing AI assistant: it helps teams ask business questions, configure settings conversationally and teach experience to the AI agent. It does not talk to customers. The shared workspace lets AI and humans switch in one click, so customers do not have to repeat themselves.
Refunds, compensation and price changes always require human approval and audit. AI should not execute them automatically.
Roll out in stages: observe, confirm, then automate
Do not flip every channel to full automation at once. A steadier path is to roll out by channel, market or question type.
Use three stages:
- Observe only: connect the new system, but do not auto-reply yet. Watch classification, knowledge hits and handoffs.
- Confirm every reply: AI drafts responses, agents approve before sending, and corrections become reviewable learning suggestions.
- Auto-send: allow automation only for low-risk, high-confidence questions. High-risk cases still hand off.
YundaDesk can start conversations at appropriate moments, but the guardrails stay on: cooldowns, frequency caps, quiet hours, no interruption while the customer is already chatting, do-not-disturb lists and human approval for sensitive actions. Observe first, confirm next, automate last — so the team sees what the system is doing before a customer ever feels interrupted.
Review after launch: turn migration into the next upgrade
One to two weeks after launch is the real review window: which knowledge items missed, which channels still feel fragmented, which human answers should become reusable knowledge, and which customer records did not merge.
Review these five points:
- Did questions the AI missed become learning suggestions?
- Did owners review and accept useful agent corrections?
- Is every new skill or knowledge item traceable, testable and revertible?
- Do all channel messages enter one workspace and one customer profile?
- Does billing match expectations, and do included AI credits cover current usage?
Put these five checks on a weekly review cadence. What surfaces in the first two weeks is more honest than anything predicted in a pre-launch meeting.
Clean the knowledge base first, reconnect channels carefully, test AI before launch, and keep humans in charge of high-risk actions. Get those four steps right, and switching support software actually holds up. Selection decides which tool you buy; migration decides whether that tool actually catches customers — surviving launch day is only a passing grade, and whether the system gets easier to run afterward is the real score. YundaDesk’s knowledge base, channel setup and AI/human boundary are all built around this sequence. See what the workspace looks like on the product page.
