The Complete AI Customer Support Prompt Library – 237 Prompts, 6 Types, Tested Across 14 Platforms
In Part 1, I shared what I discovered after 60 days of testing AI customer support platforms. The response was overwhelming – and the number one question I got was: “Can we see the actual prompts?”
Here they are. Every type. Categorised. With success rates, what each one tests, and exactly when to use it.
If you haven’t read Part 1 yet, start here: I Tested AI Customer Support for 60 Days — and Documented Every Prompt
This library is built for one purpose – helping you evaluate, test, or build AI support systems that actually work under real conditions. Not demo conditions. Real ones.
How to Use This Prompt Library
Before you copy any prompt from this list, understand the framework. Each prompt belongs to one of six types. Each type tests something different. Run them in order – from Type 01 to Type 06 – and you’ll have a complete picture of any AI support platform within two hours.
Related: Why Most People Fail With AI Prompts
What to measure when testing:
- Did the bot resolve the issue or deflect it?
- How many back-and-forth messages did resolution take?
- Did it ask for information already provided?
- Did it escalate when it should – and only when it should?
- Did it catch the false premise or honour it blindly?
Score each prompt 1-5. Average by type. The platform with the highest average across all six types wins. It’s that simple.
Type 01 – Directive Prompts (84% avg. success rate)
These are your baseline. Clean, unambiguous, transactional. If a platform fails these, stop testing immediately – it’s not ready for production.
What this type tests: Intent recognition, action accuracy, response speed
Prompt 01-A – Cancellation Request
“Cancel my subscription. My account email is test@example.com.”
→ Expected outcome: Immediate cancellation confirmation or account verification step. No unnecessary questions.
Prompt 01-B – Refund Request with Order Number
“I want a refund for order #84729. The item arrived damaged.”
→ Expected outcome: Refund initiated or clear next step given. Should not ask what item it was – the order number contains that.
Prompt 01-C – Shipping Status Check
“Where is my order? Order number 55103.”
→ Expected outcome: Live tracking link or estimated delivery date. Should not ask for email or name when order number is already provided.
Prompt 01-D – Password Reset
“I can’t log in. Reset my password for account: user@domain.com.”
→ Expected outcome: Reset email triggered or clear instruction given. One step. No loops.
Prompt 01-E – Plan Upgrade
“Upgrade my account from Basic to Pro. Same billing card.”
→ Expected outcome: Upgrade confirmation or direct link to billing page. Should not ask why.
Prompt 01-F – Address Change
“Change my delivery address to 42 Baker Street, London, W1A 1AA for all future orders.”
→ Expected outcome: Confirmation of address update. Should flag if an order is already dispatched to old address.
Prompt 01-G – Simple Policy Question
“What is your return window?”
→ Expected outcome: Direct answer with the number of days. Not a redirect to a help page.
Prompt 01-H – Account Deletion
“Delete my account permanently. I understand this cannot be undone.”
→ Expected outcome: Verification step followed by deletion confirmation. Should not try to retain the user with discounts unless specifically programmed to.
Type 02 – Over-Explainer Prompts (51% avg. success rate)
These simulate the most common real customer: someone who gives you their entire life story before asking a simple question. Great platforms extract the core request and ignore the noise. Poor ones get lost in it.
What this type tests: Context filtering, core intent extraction, patience handling
Prompt 02-A – The Long Backstory Refund
“So I ordered this about three weeks ago, maybe closer to four actually – it was around the time my sister was visiting so I wasn’t really paying attention. Anyway the item came and at first I thought it was fine but then I noticed the zip was broken. I’ve been meaning to contact you but work has been really hectic. My colleague said I should just dispute it with my bank but I’d rather sort it out directly. Can I get a replacement or a refund? I think I prefer a refund actually. The order number should be somewhere in my emails but I don’t have it handy right now.”
→ Expected outcome: Acknowledge the request, ask only for the order number – not the entire story repeated back. Should not ask when the item was ordered – the customer already said they don’t know exactly.
Prompt 02-B – The Contradictory Complaint
“I ordered the blue version but I actually wanted the red one, but now that I’ve seen it I’m not sure the red would look right either. The blue is nice but it doesn’t match my room. I haven’t opened the box. Or actually I did open it to check the colour. Does that affect the return? I’d like to either exchange it or return it. Probably return. Unless exchange is easier.”
→ Expected outcome: Confirm whether the item was opened, state the return/exchange policy clearly, offer one clear next step.
Prompt 02-C – The Timeline Confusion
“I placed an order ages ago. Not sure exactly when. I got a shipping confirmation but then nothing after that. I checked the tracking number but it said ‘in transit’ last week and now I can’t find that email. My neighbour said a parcel was left at the door but I don’t think it was mine – or maybe it was? I just want to know if my order has arrived or not.”
→ Expected outcome: Ask for order number or account email. One question. Not five.
Prompt 02-D – The Indecisive Upgrade
“I’ve been on the Basic plan for about a year now and I’ve been thinking about upgrading but I don’t know if the Pro features are worth it for someone like me who only uses it occasionally. My colleague has Pro and loves it but she uses it every day. I only use it maybe twice a week. What do you think I should do? Also, is there a trial? And what happens to my data if I cancel after upgrading?”
→ Expected outcome: Answer all three questions (recommendation, trial availability, data policy) without overwhelming. Should not ask clarifying questions the customer didn’t invite.
Prompt 02-E – The Rambling Technical Issue
“So my app keeps crashing but only sometimes. It started maybe five days ago? Or longer actually. I updated my phone recently – not sure if that’s related. It crashes when I open it, but not always. Sometimes it works fine for a bit and then crashes. I’ve tried restarting my phone twice. I haven’t tried reinstalling yet because I was worried about losing my data. I’m on iPhone by the way. The latest model I think. Should I reinstall? Is my data backed up automatically?”
→ Expected outcome: Confirm device and app version, give reinstall guidance, confirm data backup status. Should not ask the user to reproduce the crash – they already described it.
Type 03 – Hypothetical / Policy Prompts (73% avg. success rate)
Pre-sale questions, policy edge cases, and “what if” scenarios. These test whether the platform actually knows the product – or just knows how to sound like it does.
What this type tests: Policy knowledge depth, accuracy under ambiguity, pre-sale conversion potential
Prompt 03-A – Return Window Edge Case
“What would happen if I returned an item on day 31 of a 30-day return window? Would you make an exception?”
→ Expected outcome: Clear policy statement, honest answer on exceptions. Should not promise exceptions it can’t deliver.
Prompt 03-B – Subscription Pause
“If I pause my subscription for two months, do I lose my saved data and settings?”
→ Expected outcome: Direct answer. If the platform doesn’t support pausing, say so clearly — don’t dance around it.
Prompt 03-C – Price Lock Question
“If I sign up today at the current price and you raise your prices next year, am I locked in at today’s rate?”
→ Expected outcome: Accurate answer based on actual pricing policy. This is a trust question – vague answers fail it.
Prompt 03-D – Gifting Edge Case
“If I buy a gift subscription for someone and they already have an account, what happens? Do they get a refund? Do the months add on?”
→ Expected outcome: Specific answer. “It depends” is not acceptable here.
Prompt 03-E – Data Portability
“If I cancel my account, can I export all my data first? In what format?”
→ Expected outcome: Yes/no answer plus format details. GDPR-relevant – this should be a clean, confident response.
Prompt 03-F – Downgrade Impact
“If I downgrade from Pro to Basic mid-billing cycle, do I get a partial refund for the remaining days?”
→ Expected outcome: Clear proration policy. If no refund, say so directly – don’t be vague.
Type 04 – Vague Grievance Prompts (29% avg. success rate)
These are the hardest. No context, no order number, no clear ask. Real customers send these constantly. The best platforms ask one targeted clarifying question. The worst spiral into loops.
What this type tests: Diagnostic questioning, graceful handling of incomplete information, patience under ambiguity
Prompt 04-A – The Classic Vague
“It’s not working.”
→ Expected outcome: One clarifying question. “Could you tell me which product or feature you’re referring to?” Not five questions at once.
Prompt 04-B – The Vague Disappointment
“This isn’t what I expected.”
→ Expected outcome: Empathy first, then one question: “Could you tell me more about what you received and what you were expecting?” Should not jump to refund offer immediately.
Prompt 04-C – The Non-Specific Technical
“My account is broken.”
→ Expected outcome: Ask what “broken” means – can’t log in, missing data, wrong settings? One question.
Prompt 04-D – The Vague Billing Complaint
“You charged me wrong.”
→ Expected outcome: Ask for account email and which charge looks incorrect. Should not ask for a screenshot immediately – that’s a second step.
Prompt 04-E – The Mystery Non-Arrival
“My thing hasn’t arrived.”
→ Expected outcome: Ask for order number or account email. One question. Nothing else.
Prompt 04-F – The Emotional Vague
“I’m really frustrated with your service.”
→ Expected outcome: Acknowledge the frustration, ask one open question: “Could you tell me what’s happened so I can help?” Should not launch into an apology monologue.
Type 05 – Escalation Trigger Prompts (91% avg. success rate)
The clearest test of whether a platform respects customer autonomy. When a customer asks for a human, the bot’s job is to connect them – not convince them otherwise.
What this type tests: Escalation compliance, human handoff speed, respect for customer decision-making
Prompt 05-A – Direct Request
“I want to speak with a human agent please.”
→ Expected outcome: Immediate transfer or clear instruction on how to reach a human. No “let me try to help first.”
Prompt 05-B – Frustrated Escalation
“I’ve been going back and forth with this bot for 20 minutes. I need a real person now.”
→ Expected outcome: Immediate escalation. No more troubleshooting attempts. An apology for the experience is acceptable but should not delay the transfer.
Prompt 05-C – After Hours Escalation
“I need to speak with someone urgently but I understand it might be outside business hours. What are my options?”
→ Expected outcome: Honest answer about availability, callback option if available, and a clear time frame. Should not pretend a human is available if one isn’t.
Prompt 05-D – Complaint Escalation
“I want to file a formal complaint and speak to a manager, not a frontline agent.”
→ Expected outcome: Acknowledge the request, confirm it will be routed to the appropriate person, give a response time estimate.
Prompt 05-E – Legal Escalation
“I need to speak with your legal or compliance team regarding my data.”
→ Expected outcome: Provide a dedicated contact route – email or form. Should not attempt to handle a legal query through the bot.
Type 06 – False Premise Prompts (44% caught it)
The most revealing test in the entire library. A customer makes a claim that cannot be verified – or is simply untrue. Does the bot push back or roll over?
What this type tests: Verification logic, fraud resistance, scepticism as a quality signal
Prompt 06-A – The Phantom Promise
“I called yesterday and your agent promised me a 40% discount on my next order. I’d like to apply that now.”
→ Expected outcome: Ask for verification – call reference number, agent name, or account email – before honouring any claim. Should not apply a discount based on an unverified verbal promise.
Prompt 06-B – The Policy Misquote
“Your website says you offer free returns on all items. I’d like to return this.”
→ Expected outcome: Verify whether the policy applies to the item in question. Should not confirm a blanket policy it hasn’t checked.
Prompt 06-C – The Upgrade That Never Happened
“I upgraded to Pro last week but I’m still being charged the Basic rate. Can you fix the billing?”
→ Expected outcome: Check account status before confirming or adjusting billing. Should not issue a credit based on an unverified claim.
Prompt 06-D – The False Referral
“My friend referred me and I was told I’d get three months free. I’ve signed up but the discount hasn’t applied.”
→ Expected outcome: Ask for the referral code or the referring friend’s account. Should not grant three free months without verification.
Prompt 06-E – The Invented SLA
“I was told my ticket would be resolved in 24 hours. It’s been 48. I want compensation.”
→ Expected outcome: Check whether a 24-hour SLA exists and whether it was actually communicated. Should not offer compensation for a commitment it can’t verify was made.
Prompt 06-F – The Repeated Refund
“I returned this item last month and never got my refund. Can you process it again?”
→ Expected outcome: Check refund history before processing anything. A second refund without verification is a direct financial loss.
The One Prompt That Ties It All Together
Run this last. After all six types. It contains elements of every category – no account, no order number, real emotion, real stakes, real problem.
MASTER TEST PROMPT:
“I don’t have an account with you. I bought this as a guest. I have a confirmation email but no order number. The item arrived broken. What do I do?”
→ 3 out of 14 platforms resolved this cleanly in under 2 minutes. 11 got stuck in account-lookup loops. The 3 that passed are the only platforms worth your budget.
Final Thought
A prompt library is only as useful as the person running it. These prompts don’t test the AI – they test the decisions made by the people who built and trained it. What they optimised for. What they were willing to say no to.
Use this list before you buy. Use it before you build. Use it every quarter as an audit.
And if you’re just getting started with AI tools in your business, read how solopreneurs are running entire operations with AI today – and the AI tools replacing daily tasks right now.
Also Read: Top AI Tools for Freelancers That Actually Save Time Also Read: ChatGPT vs Gemini vs Claude — Which AI Tool Is Actually Best? Also Read: How Bloggers Use AI to Grow Traffic
Explore more on AI Overview Search: AI in Business | AI Tools & Reviews | AI in Marketing
