The core idea, and where my services are going

How to build a self-learning local service business.

The core idea is simple: a local business can be built to learn from its own customers, week over week, instead of guessing. Right now I'm doing it inside the marketing services I already provide for clients. The bigger play is a standalone service that builds these loops across the whole business — using the five-layer model below.

Author Joey · Local Pro Solutions

Domain Local Services

Status In active deployment

Wedge Marketing → Operations

If you just landed here from a social post —

Hey, I'm Joey. I run Local Pro Solutions, a marketing agency for local service businesses — tree services, plumbers, HVAC, landscapers, that whole world. Over the last couple of years I've been embedding AI into how my clients actually operate. Not as a productivity hack. As the brain that runs the business.

This paper is me writing out the framework. The diagram below is the picture of the whole thing. The rest is the why and how. If any of this hits, come find me.

§ The picture

The customer intelligence loop, drawn out.

Every time around, the brain in the middle gets a little richer. Patterns get clearer, decisions get sharper, ads get smarter. An agency walking in next year starts from scratch. A self-learning business starts with three years of intelligence about itself sitting right there.

§ Where my services are going

Two phases. One's already happening. The other is what I'm building toward.

Phase 01 Happening now

Embedded inside the marketing I already do.

When a client hires Local Pro Solutions for Google Ads, Meta, websites, local SEO, GBP, or lead nurturing — the loop is already running underneath the work. Every call gets captured. Every pattern feeds into the next decision. Ads get rewritten with the language customers actually use. Landing pages answer the questions customers actually ask. The marketing gets smarter every week because it's listening, not guessing.

If you're a client today, you're already in this phase. This is what you're paying for.

Phase 02 Building toward

A standalone service. The whole business becomes the loop.

Marketing's just the wedge. The bigger version of this is a service where I help a business build self-learning loops across the whole operation — pricing, dispatch, capacity, sales follow-up, retention, hiring, even which jobs to take. Same five-layer model. Same customer-intelligence core. The whole business becomes legible to itself.

Same architecture, different surface. Marketing pays for the foundation. The foundation runs the company.

The most valuable thing any service business owns is what their customers are telling them — and almost every business is letting it die on the call.

— The thesis

The Static Business

What most local businesses look like today.

Everything the owner knows lives in their head. The day they sell or retire, it's gone.
The phone rings, a customer tells you exactly what they want, you hang up, and the info disappears.
The ads say what the agency thinks sells. Not what real customers actually say.
Same decisions get made next week, next month, next year. Nothing compounds.
Marketing, ops, and sales are basically running on three different versions of who the customer is.

The Self-Learning Business

What's actually possible now.

The owner's know-how gets pulled out and stored somewhere the system can actually use it.
Every call gets recorded, transcribed, scored, and looped back into how the business runs.
Ad copy and landing pages get written using the actual words customers used on the phone last week.
The system gets a little smarter every week. Last year's data makes this year's decisions sharper.
One source of truth feeds everything — marketing, ops, sales, retention. All running off the same brain.

§ 02 — Why Now

There's so much signal. Nobody's capturing any of it.

Calls per year, per business

~2,000

A typical tree service does about 40 calls a week. That's a customer literally telling you what they want, 2,000 times a year.

Captured & structured today

≈0%

Pretty much nobody records, transcribes, and looks at their own inbound. It dies on the call and that's it.

Lift on lead quality + conversion

$100k+

If you bump lead quality 20% on a $1–5M business, that's six figures of new revenue every year. Easy math.

§ 03 — Architecture

The five layers that make this work.

Here's how I think about the stack. Each layer feeds the next one. What you know becomes policy. Policy drives the tools. Tools run through quality gates. Outcomes loop back as learning. Do that on repeat and the whole thing compounds.

Layer 01

Domain Knowledge

This is the foundation. You take what the business already knows — services, pricing, capacity, the owner's gut judgment — and combine it with what the customers are saying on calls and emails. Then you store it somewhere the AI can actually read it. That's the brain.

→ Deep dive

Sources

Owner interviews
Call transcripts
Email & SMS threads
Invoice & job data
Reviews & GBP Q&A

Layer 02

Policy Layer

Now you take the gut-feel decisions and actually write them down. How do we price this kind of job? When do we qualify a lead vs. pass? When do we pause an ad? You turn the owner's judgment into rules the system can follow. Versioned. Updateable. Auditable.

→ Deep dive

Examples

Pricing rules by service
Lead qualification logic
Negative-keyword triggers
Capacity & dispatch rules
Objection handling scripts

Layer 03

Tool Layer

This is where the system actually does stuff. Pushes ad changes. Updates the CRM. Sends the follow-up. Posts to GBP. Edits the landing page. Every action runs through a policy first, so nothing goes rogue. The tools change all the time. The protocol — MCP — is the part that matters.

→ Deep dive

Surfaces

Google Ads / Meta Ads
CRM & invoicing
Email & SMS
GBP & landing pages
Reporting dashboards

Layer 04

Quality Gates

This is the layer most AI setups skip, and it's why most of them blow up. Before anything ships, you've got checks. Human approval for the bigger stuff. Tone and fact checks on anything customer-facing. Escalation for weird situations the system hasn't seen before. This is where the owner actually trusts the system.

→ Deep dive

Gate Types

Human approval thresholds
Tone & brand checks
Factual verification
Novel-case escalation
Audit log of all actions

Layer 05

Learning Mechanism

Every action has an outcome. Quotes get accepted or they don't. Leads close or they ghost. The owner approves or overrides what the system proposed. All of that flows back into Layer 1. Next week the system knows a little more about the business than it did this week.

→ Deep dive

Feedback Signals

Lead → job conversion
Quote acceptance rate
Override frequency
Ad variant performance
Customer language drift

Deep Dive · Layer 01

Domain Knowledge

Get what the owner knows out of their head. Get what the customer says off the phone. Put them in the same place so the system can actually use them.

How the data flows in

01 · INGEST

Raw signals

Call recordings, emails, form fills, owner interviews, invoices, reviews. Anything a customer or owner touches.

02 · TRANSCRIBE

Speech → text

Calls become transcripts. Voicemails become text. Everything becomes searchable.

03 · EXTRACT

LLM scoring

Claude reads each interaction, extracts structured fields, scores against the schema.

04 · STORE

Knowledge base

Lands in Notion / Supabase as the canonical Business Brain. Queryable, versioned, retrievable.

Business Brain · Example schema

// Each call becomes a structured record
call_id: "call_8412",
service_requested: "tree removal",
customer_language: ["leaning toward the house", "storm damage"],
location: { zip: "97214", in_service_area: true },
urgency: "this_week",
fit_score: 4, // 1–5
likelihood_to_book: 4,
objections_raised: ["wants_other_quotes"],
outcome: "booked", // updated when known
revenue: 2400

Tools & databases in this layer

Call capture

CallRail

Dynamic call tracking, recording, transcription, source attribution.

Transcription

Whisper / AssemblyAI

High-accuracy speech-to-text when CallRail's native transcription isn't enough.

LLM extraction

Claude API

Reads transcripts, extracts structured scoring against the schema.

Knowledge store

MCP-reachable workspace

A database the AI can read and write. Notion, Supabase, Airtable — whatever speaks MCP and fits the team.

Email & docs

Inbox & drive MCPs

Customer email threads and historical documents feeding the knowledge base.

Reviews

Google Business Profile

Review content and Q&A pulled in as customer-language signal.

Deep Dive · Layer 02

Policy Layer

All the stuff the owner does on autopilot, written down as rules the system can actually run. Updateable. Versioned. Easy to roll back.

Anatomy of a policy

Policy · Geographic Targeting

IF call originates from outside core service zip codes

THEN add originating keyword to negative list within 24h; flag for review.

Policy · Lead Qualification

IF fit_score ≥ 4 AND urgency = "this_week"

THEN trigger same-day follow-up; route to senior estimator.

Policy · Ad Spend Trigger

IF keyword CPA exceeds target by 30% for 5+ days

THEN pause keyword, alert account manager, queue replacement variant.

Where policies live

Policy storage

MCP-reachable workspace

Owner-readable. Each policy has trigger, conditions, action, version history.

Programmatic store

Structured database

Same policies, machine-readable for agent execution. SQL or JSON, doesn't matter.

Decision engine

Claude + MCP tools

Reads context, consults policies, proposes or executes the action through MCP servers.

Versioning

GitHub

Every policy change tracked. Roll back, A/B test, audit on demand.

Deep Dive · Layer 03

Tool Layer

The hands of the whole thing. Don't get hung up on which specific tool — they all change in 6 months anyway. What matters is the AI can talk to them. That's the whole game.

The principle

Why MCP matters

Here's the way I think about it. An AI is only as useful as the tools it can actually reach. MCP (Model Context Protocol) is the standard that lets an AI talk to anything — your ad platform, your CRM, your inbox, your billing, your repo, your knowledge base. I change tools all the time. What I don't change is this: I pick tools the AI can plug into. If something better comes out next quarter, swap it in. The system doesn't care.

Build & execute environment

Agent runtime

Claude Code

The agent that reads context, consults policies, and operates the tools. Calls MCP servers to do real work.

Code & versioning

GitHub

Where the system lives. Scripts, policies, prompts, infrastructure — all versioned, all reviewable.

Hosting & deploy

Vercel

Where landing pages, dashboards, and ephemeral software get deployed — instantly, on demand.

Reach surfaces (via MCP)

Paid media

Ad platform MCPs

Read performance, write changes — pause keywords, adjust bids, swap creative, shift budgets.

Local presence

GBP & directory MCPs

Post updates, respond to Q&A, manage service info — driven by what customers actually ask.

Customer comms

Inbox & messaging MCPs

Email, SMS, chat — follow-ups, alerts, win-backs sent on policy, drafted in customer language.

Billing & revenue

Payments MCP

Invoice creation, payment status, subscription state. Closes the loop on real ROAS, not just leads.

Scheduling

Calendar & booking MCPs

Books estimates, dispatches crews, manages availability — gated by capacity rules from Layer 2.

Knowledge & ops

Workspace MCPs

Notion, docs, project management — where domain knowledge lives and where work gets coordinated.

The framing that matters

Wrong question

Which tool should I use?

Tools change every quarter. If you build the whole thing around one vendor, you've tied your own hands.

Right question

Can the AI reach it?

If there's an MCP server, or an API I can wrap in one, the AI can talk to it. That's the only question I actually care about.

Deep Dive · Layer 04

Quality Gates

Not every action carries the same risk. Some stuff is safe to just run. Some needs a one-tap approval. Some always needs a real conversation. This is the layer that lets an owner actually sleep at night.

Three risk tiers

Tier 1 · Auto Adding negative keywords, small bid adjustments, GBP review responses.

→

Gate Pattern-match check + audit log entry.

→

Action Executes immediately. Owner sees in weekly digest.

Tier 2 · Approval New ad copy, budget shifts under $X, landing page edits.

→

Gate Tone check, factual check, one-click approval queue.

→

Action Owner taps approve. Executes within minutes.

Tier 3 · Strategic Campaign restructures, major budget changes, new offers.

→

Gate Full review with proposed rationale + risk model.

→

Action Conversation. Then decision. Then execution.

How gates are implemented

Approval UI

Messaging MCP + workspace

Proposed actions appear in a chat channel. One-click approve, deny, or comment.

Tone & brand

Claude

Brand-voice checker reads every piece of customer-facing copy before it ships.

Factual verification

RAG against Business Brain

Claims in copy are verified against Layer 1 knowledge before publishing.

Audit log

Structured store

Every action — auto or approved — logged with full context. Full provenance.

Deep Dive · Layer 05

Learning Mechanism

This is the part that makes the whole thing actually compound. Every outcome — closed, lost, overridden, objected to — gets fed back into Layer 1 and makes next week's decisions sharper.

What signals get captured

Conversion

Lead → booked job

Closes the ROAS loop. Stripe invoice or CRM status confirms revenue, not just call volume.

Override

Owner says no

Every rejected proposal reveals a gap in the policy layer. Logged for refinement.

Variant test

Ad / page winner

A/B results feed back into the language model — winning phrases get amplified.

Drift

Customer language shift

New phrases or objections trending? Surfaced for review. The market is changing.

Acceptance

Quote outcomes

Price acceptance rates by service & segment refine the pricing logic in Layer 2.

Escalation

Novel cases

Anything the system flags as out-of-distribution becomes a candidate for new policy.

Tools that close the loop

Outcome data

Payments · CRM · Calendar MCPs

Did the lead become a booked, paid job? Truth source for revenue attribution.

Analytics

Reporting MCPs + dashboards

Tracks variant performance, channel mix, conversion funnel over time.

Synthesis

Claude

Weekly: reads everything that happened, proposes policy updates, drafts owner digest.

Digest delivery

Inbox & messaging MCPs

Owner gets a one-page summary: what changed, what learned, what to decide.