kakinang INSIGHTS · 2026-09-20
Which five questions should a time-strapped startup let AI answer first?
A founder often makes the product and replies to customers alone. AI Help me prioritize five startup customer questions can start with repeated questions backed by approved answers, while uncertain decisions need boundaries and a human route.
Pick five categories from real questions
First, explain what you offer, whom it fits, where, and what you decline. Second, give opening, response, production, and delivery expectations without turning a usual lead time into a confirmed slot. Third, explain fixed or starting prices and quote variables; a custom amount goes to a person. Fourth, describe the actual inquiry, booking, confirmation, cancellation, and payment steps, distinguishing a lead from a completed transaction. Fifth, show a human contact and exception path for complaints or urgent needs. These are starting categories, not a mandatory taxonomy.
Take twenty recent messages, remove personal details, and tag their intent. 'Can I collect it tomorrow?' can mix lead time, capacity, and rush work. If there are no messages, ask someone who knows the audience for ten plausible questions; label which have public answers and which need live state. Prioritize frequency, risk, and owner effort rather than a popular keyword alone.
Separate frequently asked from safe to answer automatically. Daily stock questions do not justify a certain answer without live inventory; offer a check route or gather item and quantity for the owner. A delivery area asked twice a week can still deserve a direct answer when a stable public table exists. Map frequency, change rate, and cost of error: automate frequent stable scope, give changing prices or slots stopping rules, and hand high-risk cases to people. Do not force medical advice, custom rights, or disputes into five neat categories. Wholesale and retail can have different minimum quantities or shipping fees, as can local and international delivery times. State conditions: 'for retail ...; for wholesale, leave quantity and destination.' Do not treat one arrangement as universal. Trial operations should label area, hours, and prototype specifications as changing facts. For unfamiliar questions, state the missing facts, decision owner, and reply time instead of filling space with polished vagueness.
Give each answer an owner and a stopping rule
Each card needs common wording, a short approved answer, a source, last review date, owner, and human handoff trigger. A fictional bakery may say custom work usually needs five days' notice; it cannot assert Saturday availability without checking a schedule. 'You may inquire' is not 'your order is accepted.' Allergens, health, safety, legal issues, and refund disputes call for the right person, not an individual judgment invented by AI.
Put approved text into available kakinang knowledge settings with source and version; if the feature or role is unavailable, keep the fact sheet until access is confirmed. Keep answers aligned with the public page and actual staff practice. Do not only edit a prompt while leaving a stale page. Test formal phrasing, shorthand, and a false premise. 'You take orders 24 hours, right?' should correct the hours instead of agreeing.
A colleague should be able to verify each card. 'It depends' needs missing inputs, owner, and reply time. 'Free' needs included items, limits, validity, and extras. Before uploading a public policy PDF, remove internal rates, personal information, and revision comments from the file itself; do not assume automatic redaction. AI may draft a card, but the owner checks numbers, address, area, legal conditions, and negations. Use change triggers as well as calendar reviews. A move changes address and pickup directions that day; a supplier change prompts material and allergen review; a new plan prompts rates and eligibility checks. Update page, knowledge, and team guide together, then rerun confusing questions. Hold an unapproved language version or route to a maintainable one; Chinese cannot allow refunds while English denies them by accident. Handoff needs a named owner and reasonable time. 'Passed on' means nothing if nobody checks it.
Measure saved effort and failures
Before release, run fifteen to twenty cases covering five direct categories, paraphrases, missing information, requests for immediate certainty, and attempts to coax an unauthorized promise. Record expected and actual answers and whether handoff succeeded. A delivery-area question gets a direct public answer; an unusual surcharge needs an address and owner review; 'Have I paid?' cannot be guessed without a formal payment-state check. Give the human enough necessary context so the customer need not restart.
Each week review unanswered forms, stale answers, and decisions AI wrongly called final. Reports or inquiry records can show categories and usage, but a chat is not a sale. If the owner still asks for specifications, improve intake fields. If pricing errors cluster, narrow the answer and fix the source. External search or AI citation is never guaranteed; a public FAQ first keeps customer and team information consistent.
Acceptance means a visitor can take the right next step, not just that an answer sounds smooth. Delivery to Kaohsiung needs a verifiable area or quote route. Timing distinguishes standard production from an actual slot. Cancellation points to the formal process and terms rather than hiding a missing policy. Classify failures as missing source, retrieval error, ambiguous language, missing live state, or unavailable permission; a longer prompt is not always the repair. Verify critical dates and quantities when users make typos. A five-minute weekly review asks whether the top five changed, sources expired, handoffs were answered, and false promises corrected. If delivery becomes easy but cancellation disputes rise, reprioritize. Small samples cannot support a fictional 80% saving; check work records for the repeated explanations actually removed.
Questions teams ask
Must I follow this category order?
No. Order by actual frequency, risk, and human effort; route sensitive matters first.
What if I have no message history?
Use audience-informed sample questions, then revise against real inquiries. Do not invent customer quotes or outcomes.
Can AI confirm a custom quote or slot?
Only through granted features, current facts, a formal flow, and explicit confirmation. Otherwise gather needs for a person.
Does a FAQ remove the need for people?
No. Exceptions, disputes, and unapproved commitments need a responsible person.
Sources and further reading
- Self Customer Service: What It Is and How It Works — Shopify
- Creating Helpful, Reliable, People-First Content — Google Search Central
- AI Risk Management Framework — NIST