Online stores answer the same questions all day. Where is my parcel? Can I change my address? How do I return this? What does the warranty cover? The answers are in systems the store already owns: the order system, the carrier tracking feed, the returns policy. The work is retrieving and presenting them, which is exactly where a well-grounded assistant earns its place.
What to automate first
- Order status — look up the order, read the carrier's tracking, and explain it plainly, including what a delay means and what happens next.
- Returns and exchanges — check eligibility against the policy, generate the label or instructions, and update the order record.
- Pre-purchase questions — sizing, compatibility, delivery times and stock, answered from the product catalogue rather than from memory.
- Address and detail changes — before dispatch, within rules the store sets.
Where people should stay involved
- Damaged or wrong items — often need photos, goodwill decisions and carrier claims.
- Refunds above a threshold — approved by a person, with the case summarized.
- Upset customers — recognize frustration and escalate quickly rather than answering politely in circles.
- Fraud signals — unusual patterns routed to a person, never settled by the assistant alone.
Customers forgive a clear 'let me hand you to a person' far more readily than a confident wrong answer about their money. Make escalation easy, and pass the full context so nobody asks twice.
Rules for sounding trustworthy
- Quote only live data — delivery dates and refund status come from the systems, never from guesswork.
- Apply the policy as written — and cite it. If the policy is ambiguous, escalate instead of choosing.
- Say it is automated — and make it easy to reach a person.
- Log every action — refunds issued, labels created and addresses changed, so the team can audit them.
What to track
Measure the share of questions resolved without staff, first-response time, customer satisfaction on automated and human conversations separately, and the rate of escalations that turned out to be unnecessary or too late. If satisfaction on automated conversations is lower than on human ones, the fix is usually better policy handling or a faster handoff, not a different model.