Start here · Three practical steps
- Collect common questions and their approved answers.
- Start with draft replies that a colleague reviews.
- Define which questions must go straight to a person.
AI customer service can help a small business answer routine questions, prepare email replies and send complicated issues to the right person. The best place to start is one written channel and a short list of questions with reliable answers. Giving an AI assistant access to everything and asking it to handle every customer is a much harder project.
If your team repeatedly explains delivery times, booking changes or how a service works, there may be useful work to automate. If most enquiries need judgement, negotiation or detailed investigation, begin with an assistant that drafts replies for staff to approve.
1. Choose the part of support you want to improve
Look at a recent, representative set of emails and chats. Group them by the customer's actual need, rather than by the words they used. “Where is it?” and “My parcel hasn't arrived” may belong to the same group, but an order-status question and a lost parcel need different handling.
| Customer need | Useful first role for AI | What must be in place |
|---|---|---|
| Understand opening hours or service coverage | Answer from approved information | Current hours, locations and clear exclusions |
| Check an individual order | Retrieve the relevant status | Verified identity and a reliable order-system connection |
| Ask for an exception or refund | Collect context and prepare a staff handoff | An accountable person and a decision policy |
| Explain a complicated problem | Summarise it for the support team | Access to the original conversation for checking |
Estimate how much staff time each group consumes. A frequent question is a stronger starting point when the answer is stable, mistakes are easy to catch and someone owns the process. A useful first scope might be “answer questions about delivery options in website chat”, rather than “automate customer service”.
- Question
A customer asks for help.
- Find
Use approved answers and current information.
- Check
Does the answer fit the question and your rules?
Review and send a useful, grounded reply.
A person takes over with the conversation context.
Illustrative workflow · Adapt the checks and responsibilities to your business.
2. Give the assistant a dependable source of answers
Your knowledge base can begin as a small collection of approved answers. It needs to explain what the business actually does today: delivery areas, cancellation steps, service limits, account access and where customers can get more help.
Remove contradictory versions. A current website policy and an old internal PDF should not both look authoritative. Give each important policy an owner and a review date. Write exceptions explicitly; “usually within three days” is not a promise that every order will arrive within three days.
For example, Intercom documents that Fin can answer over chat and email using a business's support content. That makes the quality and scope of that content a practical buying consideration, not just an implementation detail. Intercom's Fin FAQ.
Separate public explanations from private customer records. A general returns policy can be public; a customer's address or purchase history cannot simply be added to a shared answer library. Connect account-specific information only where the workflow checks who is asking and what they may access.
3. Make the route to a person work properly
A handoff should have an owner, context and a realistic next step. “I'll pass this on” is unhelpful if the conversation disappears into an unmonitored inbox. Decide which team receives it, what happens outside working hours and how the customer will hear back.
Set explicit handoff conditions: a direct request for a person, repeated misunderstanding, missing information, complaints and requests beyond the assistant's authority. Intercom provides configurable escalation rules and guidance; its documentation also makes clear that routing and follow-up need workflow configuration. Intercom's escalation documentation.
Ask the system to pass along the customer's question, relevant details, steps already tried and why it escalated. Staff should see the original messages as well as the summary. In Zendesk's documented messaging handoff, the human agent becomes the responder and the AI stops replying to that conversation. Check how the equivalent transition works in any tool you evaluate. Zendesk's handoff guide.
Tell customers they are speaking with an AI assistant. Keep replies brief, and do not make them complete a long questionnaire before allowing a human handoff.
4. A practical example: an online homeware shop
Hypothetical example: a small shop receives repeated messages about delivery areas, damaged items and order changes. Its first AI project covers delivery-policy questions in website chat. It does not approve refunds or change orders.
When someone asks whether the shop delivers to their area, the assistant uses the current delivery policy. When someone reports a broken lamp, it collects the order reference and a short description, then routes the conversation to staff. It does not tell the customer that a replacement is approved.
The owner tests ambiguous questions such as “Can I return it?” The assistant should ask what the item is and direct the customer to the relevant process, rather than inventing a universal answer. The team checks the early conversations daily and updates confusing policy wording. Only after this limited scope works would it consider connecting live order status.
5. Choose a setup your team can maintain
Start by checking your existing helpdesk. Adding a supported feature to an inbox staff already use may reduce duplicate customer records and training. A separate chat tool can suit a business without a helpdesk, but only if someone will monitor its inbox. A custom integration becomes more relevant when answers depend on several internal systems.
During a demonstration, use your own questions and policies. Ask the provider to show:
- How an incorrect answer is traced back to its source and corrected.
- How the assistant behaves when the answer is missing or systems are unavailable.
- Which actions require staff approval, and how permissions are restricted.
- Where conversations are stored, who can access them and how retention is controlled.
- What the invoice includes: subscriptions, usage, setup and ongoing support.
Begin with staff-reviewed drafts or a limited chat pilot. Include common questions, unusual wording, outdated offers and requests to ignore the rules in your test set. Keep a simple way to turn automation off without closing the support channel.
6. Measure whether customers actually get help
A fast reply is useful only if it moves the issue forward. Compare the pilot with a similar period before launch, accounting for changes in volume or seasonal demand. Review these measures together:
- Answer quality: were sampled replies correct, relevant and complete?
- Repeat contact: did customers return with the same unresolved issue?
- Handoff quality: did a person receive enough context to continue?
- Staff effort: how much handling and review time remained?
- Total cost: what did the service and maintenance cost per resolved issue?
Agree on a review window and what counts as resolved. An abandoned conversation or an automatically closed ticket is not reliable proof of a satisfied customer. A useful pilot leaves you with evidence about where automation helps and which questions should stay with staff.
Want help with your support inbox? Tell Evoogen which questions your customers ask repeatedly, where messages arrive and what your team currently does with them. We can help you identify a manageable first workflow and the human handoff it needs.
From reading to doing
What would this look like
in your business?
Tell us what you’d like to improve or automate. We’ll help you work out where to start.
Prefer email? support@evoogen.comSources & further reading
Primary references used in this guide. Product details and requirements can change.
- Intercom: Fin AI Agent FAQsChecked 2026-09-24
- Intercom: Manage Fin AI Agent's escalation guidance and rulesChecked 2026-09-24
- Zendesk: Managing conversation handoff and handbackChecked 2026-09-24