How we installed a private AI workforce inside a real automotive business

Jackson Hole is not a software demo. Detail Driven Enterprises runs mechanical repair, collision, detailing, and glass as four different businesses that happen to share an owner. Each line already had a system of record. Repair lived in Tekmetric. Collision lived in CCC ONE. Detailing and glass had their own boards, books, and ads. Omega EDI moved parts. GoHighLevel held the conversations. QuickBooks and Stripe held the money. Google Ads, Meta, Google Business Profile, GA4, and Search Console held demand. Slack held the exceptions that never made it into a ticket. A person was still the integration.

The Bureau of Labor Statistics Occupational Outlook Handbook counted about 825,800 automotive service technicians and mechanics in the United States in 2025, with roughly a quarter of them in independent mechanical repair. The Auto Care Association, in the 2026 Auto Care Factbook, put U.S. light-duty aftermarket sales at $413.7 billion in 2024 and projected further growth as the light-vehicle fleet aged past twelve years. None of that shows up as a clean API. It shows up as an advisor on the drive, a writer in CCC, and an owner rebuilding the morning report from five logins.

Four shop types, four systems of record

The first honest fact of this build is that “the automotive business” was never one business. Mechanical repair produces repair orders, approvals, declined lines, and comebacks. Collision produces estimates, supplements, insurance files, and Final Bills. Detailing produces bookings that either confirm or ghost. Glass produces a different cycle time and a different customer. Treating those as one CRM with extra fields is how operators buy software they then work around.

We mapped the leftover work per shop type before we talked about models. Aging tickets in Tekmetric. Quiet files in CCC. Unconfirmed detailing appointments. Ads spending without anyone checking them against cars that actually booked. A/R sitting in QuickBooks while the shop floor thought the job was closed. The Automotive Service Association and I-CAR both spend a lot of ink on process discipline for a reason: the work is already defined. The gap is that nobody is watching the process when the people who own it are busy.

• Repair: Tekmetric remains the system of record for repair orders, approvals, and declined work.
• Collision: CCC ONE remains the estimating system. Insurance work stays in CCC.
• Detailing and glass: existing boards and booking tools stay. Agents watch them. They do not replace them.
• Demand: Google Ads, Meta, Google Business Profile, GA4, and Search Console are read, not re-platformed.
• Money: QuickBooks and Stripe stay. No fake finance module.

The architecture is boring on purpose

Private Linux hardware sits at the bottom. Collectors pull from the APIs Tekmetric, GoHighLevel, QuickBooks, Stripe, and the ad platforms already publish. Collision is the honest hard part: CCC Intelligent Solutions documents an EMS export (CIECA EMS 2.01 estimates, workfile copies, part price changes) written to a folder the shop already configured. The bridge reads that folder. It does not scrape DATA.DAT. It does not pretend CCC will hand us the schedule as a clean feed, because in this install it would not.

In the middle sits an operator OS: command, inbox, quotes, bookings, shop views, money, demand, collision, and a knowledge map of how the place is wired. A manager (the orchestrator) assigns work. Specialized agents watch follow-up, monitoring, and reporting. None of them get a master login to the company. Least privilege is not a slogan we borrowed for a slide. It is how you keep an agent from becoming a new kind of office manager with the keys to payroll.

The dashboard is included in the same build. It is not a second product. It is how the owner sees exceptions without becoming the integration.

Hybrid inference, not a religion

Routine work stays local: overnight watches, ticket aging, assembling the morning briefing from records that already live in the building. When a job needs a stronger model (a messy email, a long estimate that does not match a template) it can go to the cloud. Operators care about the morning report landing, not about where the tokens ran. NIST’s AI Risk Management Framework is the language we use when an owner asks about risk: govern, map, measure, manage. Map the jobs. Measure whether the watcher is actually watching. Keep a person on the customer.

Local hardware also makes the cost of a repetitive run more predictable than renting a frontier model every time a bay opens. That is an operations argument, not a privacy brochure. Some shops will never want a box in the office. The audit is where we decide. We do not lead with GPUs.

What the workforce actually does

Follow-up: stalled repair orders, declined lines, aged estimates, and leads that sat in GoHighLevel. The agent drafts a next message from the file. A person still sends it until the path is trusted. Monitoring: open ROs, collision files that went quiet, bookings that never confirmed, ads spending with no operator. Exceptions, not a wall of green. Reporting: the morning briefing from Tekmetric, CCC, ads, and the books, so the owner starts from what broke overnight instead of five CSV exports.

We will not put hours saved, dollars recovered, or an owner quote on this page yet. Those go here when they are measured and the operator is willing to stand behind them. Until then, the facts are the story: four shop types, dedicated hardware, hybrid inference, Tekmetric and CCC still in place, daily reporting automated as a job, a secure dashboard across the connected stack, and a person who still owns the customer conversation.

What this is not

It is not a chatbot on the website. It is not a pack of agents we drop on every shop. It is not a replacement for the advisor, the writer, or the owner. The Auto Care Association can tell you the aftermarket is large and getting larger because cars are older and more complex. That complexity is exactly why a general-purpose assistant fails in a bay. The work is specific. The permissions are specific. The software is already paid for.

What the audit asks before we talk models

Which shop system is actually the board. Whether declined work is coded or just a note. Whether ads conversions are a click, a call, or a car. Whether CCC is exporting EMS at all. Whether GoHighLevel and Tekmetric share a customer id or a hope. Those questions take thirty minutes. They save you from buying a chatbot while the $1,800 estimate sits. They also save us from designing an agent for a door that does not exist.

If you want the same shape inside your shops, start with the audit. We will tell you if the stack and the volume of leftover work actually justify a build. If they do not, that is the audit doing its job.