Eight agents. One pipeline.
Zero missed leads.
Real LLM pipelines — not chatbot toys. Built on OpenAI function calling, Claude, n8n, and Twilio — integrated into your booking system, CRM, and Google Business Profile.
trigger → classify → LLM call → structured action → human escalation path
The first reply wins the customer
more likely to qualify a lead contacted within 1 hour vs. 1 hour later
more likely vs. waiting 24 hours or more
typical first-reply time for a Lead Responder agent
Source: Oldroyd, McElheran & Elkington, “The Short Life of Online Sales Leads,” Harvard Business Review, March 2011 (2,241 U.S. companies). The <60s figure is the design target of the build, not a third-party statistic.
Local service businesses lose most leads in the gap between “inquiry sent” and “owner free to answer.” The owner is mid-haircut, behind the espresso machine, or asleep. An agent closes that gap: it answers, qualifies, offers real slots and only then involves you.
How every agent is structured
8 agents I build for local businesses
Each card shows the real trigger, model, and function schema. These are deployed pipelines — not demos.
Fires the moment a lead fills a form, messages your GBP, or clicks a Facebook Lead Ad. Pulls context from your service list and sends a personalised reply that sounds like you — in under 60 seconds.
function schema›
{
"name": "qualify_lead",
"parameters": {
"name": "string",
"service_interest": "enum[haircut,colour,blowout]",
"preferred_time": "string",
"urgency": "enum[today,this_week,flexible]"
}
}Monitors Google Business Profile, Yelp, and Facebook for new reviews. Generates a reply that references the specific service mentioned, thanks by name, and matches your brand tone. Negative reviews route to you first.
function schema›
{
"name": "draft_review_reply",
"parameters": {
"reviewer_name": "string",
"rating": "integer(1-5)",
"service_mentioned": "string",
"sentiment": "enum[positive,neutral,negative]",
"reply_tone": "enum[warm,professional,empathetic]"
}
}A multi-turn conversational agent that handles the full booking flow — asks about service, stylist preference, and date, checks real-time availability via Calendly/Acuity/Square API, and confirms the appointment.
function schema›
{
"name": "create_booking",
"parameters": {
"service": "string",
"staff_preference": "string | null",
"date": "ISO8601",
"duration_mins": "integer",
"customer_phone": "string",
"send_reminder": "boolean"
}
}When a call goes unanswered, Twilio fires a webhook. The agent sends a personalised SMS within 30 seconds, offers a direct booking link, and starts a two-way SMS conversation to capture the lead before they call a competitor.
function schema›
{
"name": "handle_missed_call",
"parameters": {
"caller_number": "string",
"call_time": "ISO8601",
"business_hours": "boolean",
"prior_customer": "boolean",
"message_template": "enum[new_lead,returning,after_hours]"
}
}A multi-step sequence: 48h before → SMS reminder with one-tap confirm. 2h before → final nudge. No reply? Agent sends a reschedule offer and marks the slot potentially free — reducing dead time on your calendar.
function schema›
{
"name": "send_reminder",
"parameters": {
"booking_id": "string",
"customer_name": "string",
"service": "string",
"appointment_time": "ISO8601",
"reminder_stage": "enum[48h,2h,1h]",
"channel": "enum[sms,email,whatsapp]"
}
}RAG-powered chat agent trained on your services, pricing, hours, and policies. Uses vector search over your content to give accurate, specific answers. Unknown questions route to you with full conversation context.
function schema›
{
"name": "query_knowledge_base",
"parameters": {
"query": "string",
"context_window": "integer",
"top_k_chunks": "integer(1-5)",
"confidence_threshold": "float(0-1)",
"fallback_action": "enum[escalate,capture_email,suggest_call]"
}
}Queries your booking system for clients who haven't visited in 60+ days, segments them by last service type, and sends a personalised re-engagement message with a tailored offer. Tracks opens and re-bookings automatically.
function schema›
{
"name": "reactivate_client",
"parameters": {
"client_id": "string",
"last_service": "string",
"days_since_visit": "integer",
"lifetime_value": "number",
"offer_type": "enum[discount,gift,priority_booking,none]",
"channel": "enum[sms,email]"
}
}After a completed appointment, the agent sends a satisfaction check via SMS. Happy customers (4–5 stars) get a direct link to your Google review page. Unhappy customers get a private feedback form — protecting your public rating.
function schema›
{
"name": "route_review_request",
"parameters": {
"customer_id": "string",
"service_completed": "string",
"internal_rating": "integer(1-5)",
"google_review_url": "string",
"private_form_url": "string",
"send_delay_mins": "integer"
}
}Monitors your booking system, new reviews, and seasonal events — then writes and publishes on-brand content to Instagram, Facebook, and Google Business Profile. Before/afters for salons, daily specials for cafés, class schedules for gyms. Each platform gets native content, not copy-paste.
function schema›
{
"name": "generate_social_post",
"parameters": {
"business_type": "enum[salon,cafe,gym,pet_groomer]",
"post_type": "enum[promotion,before_after,review_highlight,seasonal,class_schedule]",
"platform": "enum[instagram,facebook,gbp]",
"tone": "enum[warm,professional,playful,urgent]",
"include_cta": "boolean",
"cta_type": "enum[book_now,call_us,visit_link,dm_us]"
}
}Tools I use to build and deploy agents
Lead Responder — full pipeline config
This is an abbreviated version of the n8n workflow + OpenAI function definition used in a real hair salon deployment. The agent handles 40–60 inbound leads per month, responds within 45 seconds, and books ~28% into consultations.
{
"workflow": "lead_responder_v2",
"trigger": {
"type": "webhook",
"method": "POST",
"path": "/leads/inbound",
"auth": "bearer"
},
"nodes": [
{
"id": "classify_source",
"type": "switch",
"conditions": [
{ "field": "$.source", "equals": "gbp_message", "next": "enrich_gbp" },
{ "field": "$.source", "equals": "facebook_lead", "next": "enrich_fb" },
{ "field": "$.source", "equals": "web_form", "next": "enrich_form" }
]
},
{
"id": "call_llm",
"type": "openai_chat",
"model": "gpt-4o-mini",
"system_prompt": "You are a warm, concise assistant for {business_name}...",
"tools": [
{
"name": "qualify_lead",
"description": "Extract lead intent and urgency",
"parameters": {
"type": "object",
"properties": {
"service_interest": { "type": "string", "enum": ["haircut","colour","blowout","other"] },
"urgency": { "type": "string", "enum": ["today","this_week","flexible"] },
"contact_preference":{ "type": "string", "enum": ["sms","email","call"] }
},
"required": ["service_interest", "urgency"]
}
},
{
"name": "check_availability",
"description": "Query Calendly API for open slots",
"parameters": {
"type": "object",
"properties": {
"service_type": { "type": "string" },
"preferred_date": { "type": "string", "format": "date" },
"duration_mins": { "type": "integer", "default": 60 }
}
}
}
],
"fallback_on_error": "escalate_to_human"
},
{
"id": "send_reply",
"type": "twilio_sms",
"to": "{{trigger.phone}}",
"body": "{{llm.generated_reply}}",
"timeout_ms": 8000
},
{
"id": "create_crm_contact",
"type": "gohighlevel_api",
"action": "upsert_contact",
"data": {
"name": "{{trigger.name}}",
"phone": "{{trigger.phone}}",
"tags": ["ai_lead", "{{nodes.call_llm.service_interest}}"],
"pipeline": "new_leads"
}
},
{
"id": "escalate_to_human",
"type": "notify",
"channels": ["sms_owner", "slack"],
"message": "⚠️ AI handoff: {{trigger.name}} — low confidence or error. Conversation: {{context_url}}"
}
],
"sla": {
"max_response_ms": 60000,
"escalation_on_timeout": true
}
}* Simplified for clarity. Production config includes retry logic, PII scrubbing, and audit logging.
What one agent can be worth — an example
Illustrative salon scenario. Every input is an assumption — swap in your own numbers.
| Input | Value | Basis |
|---|---|---|
| Inbound leads / month | 40 | Assumption |
| Lead → booking, manual replies | 30% (12 bookings) | Assumption |
| Lead → booking, instant agent reply | 38% (15.2 bookings) | Assumption: +8 pts |
| Extra bookings / month | +3.2 | 15.2 − 12 |
| Value per new client (first year) | $255 | Assumption: $85 × 3 visits |
| Extra revenue / month | ≈ $816 | 3.2 × $255 |
| Running cost / month | ≈ $100 | Mid-range of $50–$200 |
| Net gain / month | ≈ $716 | $816 − $100 |
| Payback on a $1,500 build | ≈ 2.1 months | $1,500 ÷ $716 |
If the true uplift is only +4 points, payback stretches to roughly 5 months; at +12 points it falls under 1.5. That sensitivity is why the audit week measures your real baseline first. Not sure of your numbers? Use the AI Agent Builder or the budget calculator.
From first call to live agent in 4 weeks
- 1Week 1
Audit & baseline
I map every place a lead can come in (forms, calls, DMs, GBP messages, ads) and measure today's reply time, lead-to-booking rate and no-show rate. This baseline is what we judge results against.
- 2Weeks 2–3
Build the first agent
Prompt, tools, knowledge base and escalation rules for one agent, wired into your CRM and booking system. Tested on 30–50 real past conversations before it ever messages a customer.
- 3Week 4
Shadow mode → live
The agent drafts replies and you approve them for a few days. Once approval rate is high, it goes live with a daily digest of everything it sent.
- 4Weeks 5–8
Add agents & tune
Second and third agents (e.g. No-Show Reducer, Reactivation) reuse the same plumbing, so they ship faster. Prompts are tuned weekly from real transcripts.
- 5Day 30+
Review & handover
Results vs baseline, cost per lead, and a recorded walkthrough. 30 days of post-launch support are included.
What agents will — and won't — do
Will
- Reply to every new lead in under a minute, at 2 a.m. too.
- Book, reschedule and remind through your existing calendar.
- Reply to every Google review in your voice, flagging the angry ones to you.
- Wake up past clients with a personal, relevant nudge.
- Show you the numbers: response time, bookings, cost per lead.
Won't
- Replace your team — it handles the repetitive 70%, people handle the relationship.
- Take payments or change prices on its own.
- Give medical, legal or veterinary advice — those messages escalate to you.
- Send marketing messages to people who haven't opted in.
- Answer from the open internet — it only uses content you approved.
Where agents fit your business
Hair & beauty salons
DM and call inquiries during appointments, no-shows, rebooking gaps
Read the guideCoffee shops & cafés
Review replies, catering and event inquiries, loyalty reactivation
Read the guidePet groomers
Same-day slot requests, vaccination checks, recurring-visit reminders
Read the guideFitness studios
Trial-class leads going cold, lapsed members, class waitlists
Read the guideWhat it costs — no surprises
Missed-call text-back, FAQ bot, review replies. 1–2 weeks.
Lead responder + booking + no-show + reactivation. 3–5 weeks.
LLM API, n8n hosting, Twilio. Paid directly to vendors, no markup.
With a small model (GPT-4o-mini / Claude Haiku class); varies with length.
Vendor accounts are in your name. Build is a one-time fee; there is no retainer unless you want ongoing tuning.
✓ What you get
This engagement is best for:
Technical questions answered
What's the difference between a workflow automation and an AI agent?
A workflow automation (Zapier, Make) follows a fixed if/then path — it can't reason or adapt. An AI agent uses an LLM to understand unstructured input, decide what action to take, and call the right tool. For example: a workflow can auto-send a confirmation email. An agent can read an ambiguous 'I want to come in Thursday but maybe Friday' message, extract the intent, check availability, and respond naturally.
Do you use GPT-4o or Claude? How do you choose?
It depends on the task. GPT-4o with function calling is best for structured data extraction, booking flows, and tool use. Claude 3.5 Sonnet/Haiku is better for customer-facing text where tone consistency matters — it's harder to make sound robotic. For most setups I run a hybrid: Claude for the customer-facing layer, GPT-4o-mini for the background data work.
Can agents access my existing booking system or CRM?
Yes, if it has an API (which most do — Calendly, Acuity, Square, Mindbody, GoHighLevel, HubSpot, Salesforce all do). I map the API endpoints into OpenAI function definitions or n8n HTTP nodes, so the agent can read availability, create bookings, and update contact records in real time.
How do you handle hallucinations or wrong answers?
For customer-facing agents I use RAG (retrieval-augmented generation) — the agent answers only from your approved content, not from general training data. I set confidence thresholds: if the retrieval score is below the threshold, the agent escalates to a human instead of guessing. Critical actions (booking creation, refunds) always require explicit confirmation steps.
What does a typical build cost and how long does it take?
A single-agent setup (e.g. missed call text-back + FAQ bot) typically takes 1–2 weeks and runs $800–$2,000 one-time depending on integration complexity. A full multi-agent pipeline (lead responder + booking + no-show reducer + reactivation) is 3–5 weeks, $2,500–$5,000. Monthly tool costs run $50–$200 (API calls, n8n hosting, Twilio). I provide full documentation so you can manage and update it yourself.
What happens when the AI gets something wrong or a customer complains?
Every agent has a human escalation path. The fallback triggers when: confidence score is low, the customer uses escalation keywords ('speak to a person', 'this is wrong'), or the conversation exceeds a set turn limit without resolution. You get a Slack/SMS alert with full conversation context so you can step in immediately. Nothing gets auto-resolved without a clear success condition.
How fast do agents actually reply, and does speed matter that much?
A Lead Responder typically sends its first message in 10–60 seconds. Speed matters: a Harvard Business Review study of 2,241 U.S. companies (Oldroyd, McElheran & Elkington, 2011) found firms that contacted web leads within an hour were about 7x more likely to qualify them than firms that waited just one more hour, and over 60x more likely than firms that waited 24 hours or more.
Will customers know they're talking to AI?
I don't hide it. The agent opens as your business's assistant, answers in your tone, and hands over to a human the moment a customer asks or the topic is sensitive (complaints, refunds, medical or legal questions). Honest framing performs better than pretending, and in several jurisdictions disclosure is expected for automated messaging anyway.
What if I only have 10–15 leads a month?
Then a full pipeline is overkill. Start with one agent — usually Missed Call Text-Back or Review Monitor — which pays back on a handful of recovered customers. The free AI Agent Builder tool estimates which agent fits your volume before you spend anything.
Who owns the system after it's built?
You do. Accounts (n8n, OpenAI/Anthropic, Twilio, CRM) are created under your business, not mine. You get workflow exports, prompts, documentation and a recorded walkthrough. If you stop working with me, everything keeps running.
How do you protect customer data?
Agents only receive the fields they need (name, service, time — not payment details). API keys live in environment variables, never in prompts. Business-tier LLM APIs don't train on your data by default, and message logs are retained only as long as you choose. Payments always stay inside your booking or POS system.
How do I know it's working?
Every build ships with a dashboard tracking response time, replies per lead, bookings attributed to the agent, escalation rate and cost per lead. After 30 days we compare against your baseline — and if an agent doesn't move a number, we switch it off or change it.
Explore related services
Ready to brew up better results?
Let's take a look at your current setup — totally free, no pressure.
Get Your Free Marketing Audit