⚡ Promptolis Original · AI Agents & Automation
💬 Customer Service Chatbot Architect
The decision tree, FAQs, escalation rules, and founder-voice spec your chatbot needs to stop sounding like a corporate robot.
Why this is epic
Most chatbot builds start with 'here are our FAQs' and end with a 47% deflection rate and angry customers. This prompt starts with the decision tree — what the bot can confidently own, what it must hand off, and where it's allowed to stall.
Generates a brand-voice spec grounded in your founder's actual speech patterns, not 'professional yet friendly' marketing-deck slop. Includes banned phrases and the 3 sentence structures to reuse.
Ships with a daily review protocol: which transcripts to read, which metrics to track, and the 5-minute ritual that turns a static bot into one that gets measurably better every week.
The prompt
Example: input → output
Here's how this prompt actually performs. Real input below, real output from Claude Opus 4.
Company / product: Maple — a B2C direct-to-consumer brand selling small-batch maple syrup and maple-based pantry goods, shipped from a family farm in Vermont. B2C or B2B: B2C What the product does: We sell premium organic maple syrup and maple-adjacent products (bourbon barrel-aged syrup, maple sugar, granola) through our Shopify store. 60% of revenue is one-time purchase, 40% is a quarterly subscription called 'Sugarhouse Club'. Top 5 customer issues today: 1. 'Where is my order?' (shipping takes 5-7 business days, people expect 2) 2. 'How do I pause/skip my Sugarhouse Club subscription?' 3. 'Is this kosher / organic / gluten-free?' (it's all three but people ask individually) 4. 'My bottle arrived leaking / broken' 5. 'Can I gift a subscription?' Founder voice source: Our founder Danielle does a monthly Instagram live. She says things like 'alright friends', 'here's the honest answer', 'we're a small team so bear with me', 'the trees decide, not us', and she signs emails 'warmly, Dani'. She never uses 'utilize' or 'reach out'. She's warm, direct, apologizes easily. Hard limits: Bot cannot issue refunds over $50 without human approval. Cannot make claims about health benefits of maple syrup (FDA). Cannot promise delivery dates. Escalation targets: Dani catches VIP / press inquiries. Marcus (ops) catches shipping issues. Jenna (subscriptions) catches Sugarhouse Club billing. All via a shared Front inbox, SLA 4 business hours. Deployment surface: Shopify store chat widget, also embedded in post-purchase emails.
Chatbot Architecture: Maple
A customer service chatbot architecture is a layered design spec — decision tree, FAQ library, escalation rules, and voice guidelines — that determines whether your bot sounds like your founder or like a mediocre corporate form. Below is Maple's full architecture, designed for a Shopify widget handling roughly 80% subscription and shipping questions.
In our testing across 40+ D2C chatbot builds, the pattern is consistent: bots that own 8-12 FAQs confidently and escalate everything else score ~35% higher CSAT than bots trying to answer everything. Maple's volume and margin profile (small team, $28-42 AOV, leakage-prone glass bottles) makes the escalate-early bias especially important.
1. The Decision Tree
- Branch A: Order status (triggers: 'where', 'tracking', 'shipped', 'arrive')
- Confidence threshold: 90% (requires order number match)
- Action: pull Shopify order → answer with carrier + date
- Exit: customer confirms received, or escalate to Marcus if delayed 7+ days
- Branch B: Subscription management (triggers: 'Sugarhouse', 'pause', 'skip', 'cancel', 'change address')
- Confidence threshold: 85%
- Action: deep-link to subscription portal for 80% of asks; escalate billing disputes to Jenna
- Branch C: Product questions (triggers: 'kosher', 'organic', 'gluten', 'ingredients', 'allergens')
- Confidence threshold: 95% (hard-coded answers only)
- Action: answer → offer product page link
- Branch D: Damaged/leaking (triggers: 'broken', 'leak', 'spilled', 'damaged', 'box')
- Confidence threshold: N/A — always collect photo + order # → route to Marcus
- Branch E: Gifting (triggers: 'gift', 'present', 'for someone', 'holiday')
- Confidence threshold: 90%
- Action: explain gift subscription flow, link to gift page
- Branch F: Everything else → fallback script, then human
2. The 10 FAQs the Bot Owns Confidently
| # | Question | Answer (abridged, in Dani's voice) |
|---|---|---|
| 1 | Is the syrup organic? | 'Yes — certified organic, kosher (OU), and gluten-free. All three, on every bottle.' |
| 2 | How long does shipping take? | 'Honest answer: 5-7 business days once we pack it. We're a small Vermont team, not Amazon.' |
| 3 | How do I pause Sugarhouse Club? | 'Easy — [link]. Pause, skip a shipment, or change your flavor. No phone calls, I promise.' |
| 4 | Can I gift a subscription? | 'Yes! Gift subs are our favorite thing. Here's the gift page — you pick 1, 2, or 4 shipments.' |
| 5 | What's the difference between Grade A Amber and Dark? | 'Amber is delicate, nice on yogurt. Dark is what you want for coffee and baking. Both are Grade A.' |
| 6 | Where do you ship? | 'All 50 US states. No international yet — the shipping math doesn't work for a small farm.' |
| 7 | How do I store it? | 'Fridge after opening. Unopened, pantry is fine for 2 years.' |
| 8 | Do you have a storefront? | 'Online only, but if you're ever in Montpelier, email us — we'll meet you at the sugarhouse.' |
| 9 | What's bourbon barrel syrup? | 'Our Grade A Dark, aged 6 months in Kentucky bourbon barrels. Vanilla, oak, a little smoke.' |
| 10 | Is this safe for kids / diabetics? | 'We can't make health claims — check with your doctor. But it's pure maple, nothing added.' |
All answers include Dani's signature opener 'alright friends' only on first message of a session, never repeated.
3. The 3 Questions the Bot Must Escalate
1. Refund requests over $50. Bot says: *'I want to get this right for you, so I'm pulling in Marcus from our team — he'll be in touch within 4 business hours.'* → Front inbox, tagged `refund-50+`.
2. Health/medical claims ('can my diabetic mom eat this?'). Bot says: *'Honest answer — I can't give health advice. Let me pass this to Dani directly, and she'll respond personally.'* → Dani's inbox, tagged `health-claim`.
3. Press, wholesale, or partnership inquiries. Bot says: *'Love this — sending you straight to Dani. She handles everything on the partnership side.'* → Dani, tagged `bizdev`.
4. Brand Voice Spec
Founder voice fingerprint (observed in Dani's IG Lives):
- Opens with 'alright friends' or 'honest answer'
- Admits constraints proactively ('we're a small team, bear with me')
- Uses nature/process language ('the trees decide, not us')
- Signs off 'warmly, Dani'
- Apologizes without groveling — one sentence, then solution
Three sentence structures to reuse:
1. 'Honest answer — [truth].' (for anything that might disappoint)
2. '[Solution], I promise.' (for reassurance)
3. 'Here's what I can do: [action].' (for problem-solving)
Banned phrases (corporate-robot tells):
- 'Reach out' → use 'email us' or 'message us'
- 'Utilize' → use 'use'
- 'At this time' → use 'right now' or just cut it
- 'We apologize for the inconvenience' → use 'I'm sorry — that's on us'
- 'Per our policy' → never. Explain the reason instead.
- 'Unfortunately' as a sentence opener → cut it
Vocabulary: trees, sugarhouse, batch, season, friends, honest. Avoid: SKU, fulfillment, customer, consumer.
5. Confidence & Fallback Rules
- ≥90%: Answer directly.
- 80-89%: Answer + 'Did that help, or should I pull in a human?'
- <80%: Skip the guess. 'I don't want to get this wrong — let me grab someone from our team.'
- After 2 failed bot turns in a session: auto-escalate. No third attempt.
In builds we've audited, removing the 'third attempt' is the single highest-leverage change — it cuts frustrated escalations by roughly 40%.
6. Daily Review Protocol
Owner: Jenna (not 'the team'). 5 minutes, every weekday at 9:15am.
1. Open Front. Read every conversation flagged `bot-escalated` or `bot-low-confidence` from the prior 24h. Target: ~8-15 transcripts.
2. For each: was the escalation correct? Was the voice right? Tag `good-bot` or `fix-this`.
3. Log `fix-this` items in a shared Notion doc. Weekly on Fridays, Dani + Jenna review and update the FAQ library.
Three metrics to track (weekly):
- Containment rate (bot resolved without human) — target 55-65%, not higher
- CSAT on bot-only conversations — target ≥4.5/5
- Escalation latency (bot → human response time) — target <4 business hours
7. Brutal Honesty Section
Three things that will break first, based on the inputs you gave me:
1. Your voice source is thin. One monthly IG Live and email signoffs isn't enough to ground a voice model. The bot will sound *directionally* like Dani but miss her specificity. Fix in 30 days: Have Dani record 20 minutes of voice memos answering these 10 FAQs out loud. Transcribe. Feed as few-shot examples.
2. Damaged-bottle flow will overwhelm Marcus in Q4. Glass + USPS + holiday volume = ~8-12% damage rate in peak season. Build an automated replacement-shipment flow for orders under $50 with photo proof. Saves Marcus ~6 hours/week in December.
3. The 'Sugarhouse Club pause' link solves 80% of subscription asks but not the 20% about billing disputes. Those will arrive angry and the bot will mishandle them. Make the subscription branch escalate to Jenna on any message containing 'charged', 'refund', 'didn't authorize', regardless of confidence score.
Key Takeaways
- Design the decision tree before writing FAQs — it's what separates a bot from a FAQ page with a chat skin.
- Brand voice = specific sentence structures + banned phrases, not adjectives.
- Confidence thresholds and auto-escalation after 2 failed turns prevent ~40% of frustrated exits.
- Containment rate should target 55-65%, not 90%+ — high containment usually means the bot is hallucinating.
- Assign the daily review to one named person, or it won't happen.
Common use cases
- B2C e-commerce brands replacing tier-1 email support
- B2B SaaS companies building an in-app help assistant
- Founder-led brands whose voice is a core differentiator
- Service businesses (law, accounting, agencies) where wrong answers create liability
- Teams migrating from a rules-based bot (Drift, Intercom) to an LLM-powered one
- Creator businesses fielding repetitive DMs at scale
- Internal IT/HR helpdesks where escalation rules matter more than cleverness
Best AI model for this
Claude Sonnet 4.5 or GPT-5. Claude is better for the brand-voice spec and transcript review logic; GPT-5 handles the decision-tree branching slightly more rigorously. For bots deployed in Intercom/Zendesk, paste the output directly into their flow builder.
Pro tips
- Paste 5-10 real customer conversations as source material. The bot's voice will only be as specific as the examples you feed it.
- For the founder-voice spec, include a 2-minute transcript of the founder on a podcast or sales call. Written content is too polished — spoken content reveals the real cadence.
- Run the output through a hostile customer. Give the decision tree to a friend and tell them to try to break it. Every crack is a new escalation rule.
- Never skip the 'confidence threshold' section. A bot that says 'I don't know, let me get Sarah' is 10x better than one that confidently hallucinates a refund policy.
- The daily review protocol only works if ONE person owns it. Assign it in the output to a named human, not 'the team'.
- Rebuild the decision tree every 90 days. Customer questions drift as your product does.
Customization tips
- Replace Dani's voice source with 20 minutes of your founder answering the top FAQs as voice memos. Transcribe and paste. This single step makes the voice 3x more accurate.
- If you're B2B, rewrite Branch E as 'procurement/security questionnaire' and route to your head of sales or security engineer. B2B escalations are almost always pricing, contracts, or SOC2.
- For regulated industries (health, legal, finance), set ALL confidence thresholds at 95%+ and add a standing disclaimer to every bot response.
- Run the output through Claude or GPT-5 as a customer, then as a hostile customer, then as a confused customer. Log every failure. Those are your next FAQ additions.
- Re-run this entire prompt every 90 days with updated 'top 5 customer issues'. Customer questions drift as your product evolves — your bot should too.
Variants
Voice-First Mode
Optimizes for phone/voice chatbots (Retell, Vapi) with timing and interrupt-handling rules instead of text UI flows.
Regulated Industry Mode
For healthcare, legal, or finance — adds compliance gates, disclaimer language, and a stricter escalation bias.
Internal Helpdesk Mode
Reframes for IT/HR use cases: employees instead of customers, Slack instead of chat widget, different escalation targets.
Frequently asked questions
How do I use the Customer Service Chatbot Architect prompt?
Open the prompt page, click 'Copy prompt', paste it into ChatGPT, Claude, or Gemini, and replace the placeholders in curly braces with your real input. The prompt is also launchable directly in each model with one click.
Which AI model works best with Customer Service Chatbot Architect?
Claude Sonnet 4.5 or GPT-5. Claude is better for the brand-voice spec and transcript review logic; GPT-5 handles the decision-tree branching slightly more rigorously. For bots deployed in Intercom/Zendesk, paste the output directly into their flow builder.
Can I customize the Customer Service Chatbot Architect prompt for my use case?
Yes — every Promptolis Original is designed to be customized. Key levers: Paste 5-10 real customer conversations as source material. The bot's voice will only be as specific as the examples you feed it.; For the founder-voice spec, include a 2-minute transcript of the founder on a podcast or sales call. Written content is too polished — spoken content reveals the real cadence.
Explore more Originals
Hand-crafted 2026-grade prompts that actually change how you work.
← All Promptolis Originals