AI Customer Support: What to Automate and What to Keep Human

AI customer support works best when it handles the repetitive, low-risk questions, such as opening hours, prices, order status and booking changes, and hands everything emotional, complex or high-value to a person quickly. The goal is not to remove your support team; it is to free them from answering the same ten questions all day so they can spend time on the conversations that actually keep customers.

Key Takeaways

• Automate questions with clear, factual answers that live in your own data: hours, prices, delivery areas, order tracking, appointment changes.
• Keep complaints, refunds, sensitive personal situations, sales negotiations and anything legal or medical with humans.
• Always offer a visible, fast route to a real person; hiding it damages trust more than slow replies do.
• Ground the AI in your own approved content so it does not invent policies or prices.
• Measure resolution and handover quality, not just how many chats the bot closed.

What Should You Automate in Customer Support?

Automate the questions you could answer with a lookup or a single paragraph from your policies. In most businesses these make up a large share of incoming messages, and customers are happy to get them answered instantly at midnight rather than waiting until morning.

Good candidates for automation:

• Information questions: opening hours, location, prices, delivery areas, required documents, course schedules.
• Order and booking status: "Where is my order?" or "What time is my appointment?" pulled from your system, not guessed.
• Simple changes: rescheduling a booking, updating a delivery address before dispatch, resending an invoice.
• Triage and data collection: asking for an order number, photo of a damaged item or preferred callback time before a human picks up.
• After-hours coverage: acknowledging messages, answering what can be answered and setting expectations for the rest.
• Language coverage: replying in Dari, Pashto or English depending on how the customer writes.

For Afghan businesses, much of this traffic arrives through WhatsApp and Facebook Messenger rather than website chat. Automating those channels properly usually means the official WhatsApp Business Platform; our guide to WhatsApp Business API automation explains how that works.

What Should Stay Human?

Keep humans in charge of anything where a wrong or cold answer costs you a customer, money or reputation. AI is good at retrieving and phrasing information; it is not good at judgement, empathy under pressure or bending rules sensibly.

Keep these human, or at least human-approved:

• Complaints and angry customers: the AI can collect details and apologise, but a person should resolve.
• Refunds, discounts and exceptions: anything that costs money or sets a precedent.
• Sales conversations with high value: corporate orders, project quotes, negotiations.
• Sensitive situations: health, bereavement, legal or financial difficulty, safety issues.
• Ambiguous or new problems: if the question is not covered by your documented answers, it should go to a person, not be improvised.
• VIP and long-term clients, who expect to deal with someone they know.

A useful rule: if a mistake would require the owner to personally apologise, the AI should not be making that decision.

How Do You Stop an AI Support Bot From Making Things Up?

You stop hallucinations by grounding the bot in your own approved content and instructing it to hand over when the answer is not there. Generic chatbots answer from general knowledge, which is how they end up inventing refund policies or quoting old prices.

The practical approach is retrieval: the bot searches your FAQ, price list, policies and order data, and answers only from what it finds. This is the RAG knowledge base chatbot pattern. Add clear rules: never promise refunds, never quote prices not in the price list, and say "Let me connect you with a colleague" when unsure.

Keep the knowledge base current. When prices or policies change, the support content must change the same day, or the bot will confidently repeat old information.

How Should the Handover to a Human Work?

A good handover is fast, visible and carries the context so the customer never repeats themselves. This is where many AI support projects fail: the bot is decent, but escalation is slow or loses the conversation history.

Design the handover deliberately:

1. Offer a human option on every screen, for example a "Talk to a person" button or a keyword like "agent".
2. Trigger automatic escalation on signals: repeated questions, negative language, words like "refund" or "complaint", or two failed answers in a row.
3. Pass a summary of the conversation, the customer's details and what the bot already tried to the agent.
4. Set honest expectations: "A colleague will reply within 2 working hours" is better than silence.
5. Route by topic and language, so a Pashto-speaking customer reaches someone who speaks Pashto.
6. Let the agent return the chat to the bot for routine follow-ups such as sending a receipt.

What About Dari and Pashto Support?

Modern AI models handle Dari reasonably well and Pashto less consistently, so test carefully with real customer messages before going live. Customers often mix scripts, write Dari or Pashto in Latin letters, or switch to English mid-sentence. Your test set should include all of these.

Prepare approved answer templates in each language for your most common questions, and have a native speaker review the bot's replies during the first weeks. If quality in one language is not good enough, it is better to route that language to a person than to frustrate customers. Our article on AI chatbots for websites in Dari and Pashto goes deeper on this.

How Do You Measure Whether AI Support Is Working?

Measure whether customers' problems got solved, not just whether the bot replied. A bot can "close" thousands of chats while customers leave annoyed.

Track a short set of numbers:

• Resolution rate: conversations resolved without a human, checked by sampling, not assumed.
• Escalation rate and reasons: what the bot could not handle tells you what content to add.
• Time to human: how long escalated customers wait.
• Customer feedback: a simple thumbs up or down after the chat.
• Error review: read a sample of conversations weekly and correct wrong answers at the source.

Start with one channel and your top twenty questions, review weekly, and expand only when the numbers and the transcripts both look good.

Work With Sheen

Sheen builds AI support assistants for websites and WhatsApp that answer from your own content in English, Dari and Pashto and hand over cleanly to your team. Explore our automation services or contact us to map which of your questions are worth automating. Email info@sheen.af or call +93 784 966 018.