Local AI vs cloud AI for service businesses: what actually belongs on each

Local versus cloud is a bad argument if you treat it like a religion. Some work should never leave the building. Some work needs a frontier model. Most service businesses need both, routed on purpose. The operators we sit with do not wake up wanting a position on GPUs. They want the morning report to land, the $1,800 estimate to get a second touch, and the ads account to stop being a Sunday job.

NIST published the AI Risk Management Framework in January 2023 around four functions: govern, map, measure, manage. That is a better starting point than a vendor war. Map the jobs. Decide where each job is allowed to run. Measure whether it actually ran. Keep a human on anything that looks like a promise to a customer. The framework is federal guidance, not a shop manual, but it is the cleanest public language we have found for the conversation owners actually want to have.

Where local wins

Local wins when the job is repetitive, high-volume, and sits next to records you do not want to ship off-site every hour. Owner reports. Ticket watches. Drafting a follow-up from a job file you already store. Dedicated hardware you control also makes the cost of a run more predictable than renting tokens every time a bay opens. At Detail Driven, overnight watches and the daily briefing are the obvious local jobs. The records already live in Tekmetric, CCC exports, QuickBooks, and the CRM. Sending them on a tour of someone else’s cluster so a small model can say “this RO is old” is a waste of a trip.

Local also wins when the shop’s internet is the constraint. Independent repair is not a downtown office. The Bureau of Labor Statistics still places a large share of technicians in automotive mechanical and electrical repair shops, many of them in markets where the upload is the first thing that dies on a snow day. A watcher that has to phone home to notice a stalled ticket will miss the morning that mattered.

Where cloud wins

Cloud wins when the job is messy, rare, or needs a model that is still moving fast. A weird customer email. A long estimate that does not match a template. A one-off summary of a messy file. Paying for that occasionally is cheaper than pretending a small local model can do everything. It is also more honest. We have not met an owner who wants to be in the business of training foundation models.

Cloud also wins when the system of record is already in the cloud and the vendor’s API is the only legal door. Tekmetric, GoHighLevel, Shopmonkey, ServiceTitan, Jobber, these are not files on a desktop. You are already trusting a vendor with the customer list. The question is not “does data leave the building.” The question is “what new inference are we adding, under whose keys, with what log.”

Rules still beat models

If a repair order has no approval and no customer update after a set number of hours, you do not need a language model to notice. You need a watcher, a permission, and a person who still owns the customer. The same is true for a collision file with no activity, a detailing appointment that never confirmed, or a Google Ads campaign that spent yesterday and produced no booked car in the shop system. Those are thresholds. Thresholds are automations or agents with a very small brain. Mixing them up with ChatGPT is how the category got a reputation for theater.

• Local: overnight watches, daily briefing, anything that reads records you already store on-site.
• Cloud: messy writing, rare exceptions, jobs that need a model you will not run yourself.
• Rules: aging tickets, missing approvals, spend with no booking, appointments that never confirmed.
• Human: money, promises, comebacks that need a conversation, anything that would embarrass you if it were dumb.

What we actually installed

Detail Driven runs hybrid. Private hardware for the operator OS and the routine watch. Cloud when a job needs a stronger model. Collision estimates arrive as EMS files CCC already knows how to write. We did not invent a third way to log into CCC. That constraint is the design. The Automotive Service Association can talk all day about process; the process only holds if the watcher still works when the writer is in a meeting with an insurer.

Operators care about the morning report landing, not about where the tokens ran.

How we decide in the first thirty minutes

We ask where the records live, who is allowed to send a customer a message, and what would embarrass the owner if it were wrong. Those three answers usually settle the routing. A shop that already stores everything in a cloud SMS and is comfortable with that will not get a lecture about a box. A shop that wants overnight watches next to collision files will. The audit writes the decision down so it is not a vibe.

The practical answer for most shops and trades is the same shape even if the box never shows up. Local or VPC for the overnight watch and the briefing. Cloud when the agent has to write something that would embarrass you if it were dumb. The audit is where we decide which jobs belong where. It is not a slide about GPUs.