A private AI system across four automotive service businesses

The operation

Detail Driven Enterprises runs automotive work in Jackson Hole across four shop types: mechanical repair, collision, detailing, and glass. Repair lives in Tekmetric. Collision lives in CCC ONE. Detailing and glass have their own boards, books, and ads. Each line already had software. None of them shared a picture.

Someone still walked the board, chased declined work, rebuilt the morning report from five logins, and hoped ads had booked cars. We did not replace those systems. We installed collectors, a secure operator dashboard, and agents that watch the leftover work so it does not depend on who remembered.

• Systems connected: 12
• Shop types: 4
• Infrastructure: private + cloud hybrid
• Operator dashboard: included

The architecture

Four layers. Hardware and collectors underneath. A secure dashboard in the middle. An orchestrator that assigns jobs. Specialized agents on top. Nothing got a master login.

1. 01Private Linux AI infrastructure

   Dedicated hardware on site. Routine inference stays local. Cloud when a job actually needs a stronger model. The overnight watch and the morning briefing do not have to leave the building.

2. 02Collectors and bridges

   Tekmetric, ads, CRM, and books over their APIs. Collision through a local CCC ONE EMS export. We read the files CCC already knows how to write. We do not scrape private databases.

3. 03Secure operator dashboard

   One board for operations, marketing, finance, and the four shops. Command, inbox, quotes, bookings, money, demand, collision, and a knowledge map. This is where department exceptions and the numbers land in the morning — without five logins.

4. 04Orchestrator and agents

   A manager assigns follow-up, monitoring, and reporting. Agents watch the queues. A person still owns the customer.

The operator dashboard

On top of the agents that run parts of the company, they get a secure dashboard. One board. Operations, marketing, finance, the shops. Pulled from Tekmetric, CCC, HighLevel, QuickBooks, Stripe, Google Ads, Meta, and the rest of the stack they already paid for. Not five logins before coffee. Not a master key to the company.

Agents watch the queues. The dashboard is how the owner sees all of it — exceptions, the numbers, and which system they came from — without rebuilding the morning from exports.

• Least privilege

  Each view and each agent gets the systems it needs. Not a master login to the company.

• Private hardware

  The board runs on dedicated Linux they control. Routine numbers do not have to leave the building.

• Systems stay

  Tekmetric, CCC, QuickBooks, ads — still the systems of record. The dashboard reads them. It does not replace them.

• Command

  Exceptions that need a person today across repair, collision, detailing, and glass. The morning list, not four boards.

• Inbox and calls

  Leads and conversations pulled from GoHighLevel and the phones. Sales follow-up starts here, not in a separate dump.

• Quotes and bookings

  Estimates and schedule in one place. Detailing bookings stay on the booking system they already used.

• Shop views

  Each location keeps its own type, hours, capacity, and management system. Agents watch them. They do not replace them.

• Money

  QuickBooks and Stripe where those shops already ran them. Unbilled work, A/R, collections. No fake finance module.

• Demand

  Google Ads, Meta, GA4, Search Console, and reputation next to what actually booked. Spend without an operator is an exception.

• Collision

  CCC estimates, claims, and finalized bills from the EMS export. Insurance work stays in CCC. Quiet files and uncollected finals surface here.

• Knowledge map

  How the operation is wired, so the next agent has somewhere to look besides a Slack thread.

Sanitized exception list

Example of what the morning board looks like. Identifiers changed. The pattern is real: aging work, quiet files, unbilled jobs, spend without an operator.

Agents on the ground

Installed as three jobs, not a 30-name org chart. Follow-up. Monitoring. Reporting. They read the software the shops already run. A person still sends anything that talks to a customer until the path is trusted. How that shows up inside Detail Driven is by department.

• Follow-up

  Stalled repair orders, declined work, aged estimates, and leads that sat in HighLevel. Drafts a next message. A person still sends it until the path is trusted.

• Monitoring

  Open ROs, collision files that went quiet, bookings that never got confirmed, ads spending with no operator. Exceptions, not a wall of green dashboards.

• Reporting

  The morning briefing from Tekmetric, CCC, ads, and the books. The owner starts from what broke overnight, not from five CSV exports.

By department

Same three jobs. Six places they show up for the people who run the shops.

• Operations

  ~22 hrs/mo

  The board. Repair, collision, detailing, and glass used to meet in Slack.

  • RO watch

    Tekmetric · Repair

    Open repair orders go quiet when the advisor is on the drive. The agent flags aging tickets, missing approvals, and customers waiting too long. Used to depend on someone thinking to look.

  • Collision file watch

    CCC ONE · Collision

    Files that went quiet after the estimate. The agent reads the EMS export CCC already writes — jobs, claims, totals, finals — and flags what stopped moving. Appointments and payments are not in that file. We do not pretend they are.

  • Booking and glass exceptions

    Bookings · Detailing · Glass

    Unconfirmed detailing appointments and glass jobs that cannot start. Exceptions land on the command view instead of a group chat.

• Sales

  ~14 hrs/mo

  The recommendation was logged. The follow-up was not.

  • Declined-work queue

    Tekmetric · Repair

    Declined lines used to leave with the car. The agent keeps the queue and drafts the next text from the ticket. An advisor still sends it.

  • Aged estimates

    CCC ONE · Tekmetric

    Collision and repair quotes sitting after the window. The agent prepares context from the file so the next call is a queue, not a scavenger hunt.

  • Lead follow-up

    GoHighLevel · Detailing · Glass

    Detailing and glass leads that sat in HighLevel. The agent checks history, drafts outreach, updates the CRM, and alerts a person only if judgment is needed.

• Finance

  ~12 hrs/mo

  Work happened. The file did not. No fake general ledger.

  • Unbilled completed work

    Tekmetric · Stripe · QuickBooks

    Glass, detailing, and after-hours repair that finished in the bay and never became an invoice. The agent flags cash sitting in the file before it becomes a Monday pile.

  • Collision finals vs collections

    CCC ONE · QuickBooks

    A job can look done in the booth while the Final Bill export has not hit the books. The agent flags finalized CCC files that are not collected.

  • Daily cash exceptions

    QuickBooks · Stripe

    A/R and unbilled exceptions feed the morning briefing from the systems those shops already ran. The owner does not rebuild cash from three exports.

• Management

  ~16 hrs/mo

  The Sunday job. Four shop types, five logins, one person.

  • Overnight owner briefing

    Tekmetric · CCC · Ads · Books

    Hours, load, comebacks, A/R, ads versus cars that booked. Someone used to build this by hand every morning and again on Sunday. The agent assembles exceptions overnight. The owner starts from what broke. This path is most of the sixteen hours.

  • Cross-shop command view

    Operator OS

    What needs attention today across repair, collision, detailing, and glass. One list instead of four boards and a Slack thread.

• Marketing

  ~8 hrs/mo

  Ads were spending. Nobody was checking them against booked work.

  • Spend vs booked

    Google Ads · Meta · GA4

    Four shop types, two ad platforms, and no operator on the account. The agent puts spend next to cars that actually booked in Tekmetric and detailing that actually confirmed.

  • Reputation watch

    Google Business Profile

    Four profiles. New reviews alert on the briefing instead of a weekly login that did not happen.

  • Lead-source picture

    GoHighLevel · Search Console

    Which inquiries became jobs. Kills the meeting story about where the work came from.

• Customer experience

  ~8 hrs/mo

  The customer used to find out in the driveway.

  • Update reminders

    Tekmetric · CCC ONE

    The RO moved or the collision file moved and nobody texted. The agent drafts the update so the customer is not the last to know. A person still sends it.

  • Booking confirms

    Detailing

    Unconfirmed detailing appointments get a next step overnight instead of sitting on the calendar.

  • Review and issue flags

    GBP · Slack

    A public problem or a thin service file shows up as an exception. The hard conversation stays with a person.

Collision, the hard way

Collision is why the operations and finance hours are not only a Tekmetric story. CCC ONE does not hand you a clean public API for the shop floor. Appointments are not in the file export. Payments are not either. What CCC will give you is an EMS 2.01 estimate export, and a Final Bill export when the repair order is locked.

We put a read-only bridge on the collision machine. It watches the export folder CCC already writes: jobs, claims, customers, vehicles, parts, labor, totals, and attachments a user saved out. It does not open DATA.DAT. That export is enough to watch files that went quiet and finals that never hit the books. It is not enough to schedule. We do not pretend it is.

Systems connected

These are the systems the department jobs actually read. Repair in Tekmetric. Collision in CCC. Leads in HighLevel. Money in QuickBooks and Stripe. Demand in the ad accounts and Google Business Profiles they already paid for.

Named tools are the stack in this deployment, not a certified public integration list. Access is always confirmed in the audit. A configured connection is not the same thing as a live feed.

What we track next

The hours above are estimates of leftover admin. Next we count what we can measure: agent tasks completed, how often a path ran without a person, how often a person had to step in, and how much inference stayed on the local hardware. Those numbers go here when they exist and the operator is willing to stand behind them.