AI agents at J and A Automotive: catching the calls, running the back office

J and A Automotive is a repair shop in Chicago, and its owner runs it largely from a distance. The experiment is to find the work an AI can reliably carry, hand it over in scoped pieces, and keep a person in control of anything with real consequences. Two parts of that are already running: a receptionist on the phones, and an agent layer that takes on the owner’s back-office work.

The AI receptionist

The receptionist answers every call, day or night. When the front desk is slammed, when a call comes in after hours or on a weekend, or when a caller would otherwise drop into voicemail, the AI picks up and holds a natural, human-sounding conversation in English or Spanish. It captures the vehicle and the problem, answers from the shop’s own hours and pricing, books the job, and texts a reminder so the appointment actually shows up. A missed call at a repair shop is usually a booked job that went to the next shop down the road, which makes catching those calls some of the most valuable work in the building. It stays scoped: it handles the routine front-desk work and hands off anything past it, like a tricky diagnosis or an upset caller, to a person.

It also chases the work the shop already earned and lost track of. The same assistant calls and texts lapsed customers to book them back in, sends service and maintenance reminders before a car is due, and follows up after a job to ask for a review. It is steady outreach that quietly fills slow days, the kind of follow-through that rarely happens when the people up front are already busy fixing cars and answering phones.

An agent layer for the back office

The second part runs in the back office. The owner’s own work used to be 20 to 30 hours a week of scattered spreadsheets and email threads: hiring and payroll oversight, budgets and cash flow, insurance and license renewals, vendor terms, pricing, and the strategic calls that pile up on weekends. That work now runs through a layer of focused agents, one each for people, finance, risk and renewals, vendors, and strategy. An orchestrator ties them together and produces a single morning briefing the owner can read in under 3 minutes.

Back there, the boundary is strict: the agents draft, recommend, and escalate, but they do not act. Nothing moves money, signs a contract, makes a commitment for the business, or deletes a record on its own. Every hire, renewal, price change, and vendor agreement lands in an approval queue with the draft prepared and the reasoning attached, and the owner approves, edits, or rejects it. A few rules never bend: no money-moving actions, no binding commitments, no public statements, nothing destructive. The finance agent is read-only on money by design.

How the system keeps learning

Both parts are built to get better over time. When the owner edits a draft or rejects a recommendation, that correction becomes the example the agent learns from, so the next draft needs less fixing. The receptionist sharpens the same way on the calls and bookings it sees most. The guardrails do not loosen as the system learns. It earns a wider scope of work, never the freedom to act without a check.

What autonomous really means here

Autonomous, here, does not mean a business that runs with no people in it. The shop still has the humans who fix the cars, and the owner still holds every decision that moves money or carries real weight. What becomes autonomous is the work around those decisions: catching calls, winning back customers, drafting choices, watching the numbers, chasing renewals. The aim is to take the owner’s load from 20 to 30 hours a week down toward 5, and to spend those 5 on judgment instead of administration.

Every draft, escalation, booking, and decision is logged, so the experiment leaves a record of where an agent could be trusted and where it could not. That map, what an agent can own and what a person keeps, is the point, more than any claim that the shop runs itself. It is ongoing, and what holds up gets written down here.