AI Went From $40 a Month to $610 in February at a 12-Bay Truck Shop. The Cap Nobody Turned On.

Forty dollars in November. Sixty-one in December. Six hundred and ten in February.

That is a real shape of bill for a twelve-bay heavy-duty shop that started letting its service writers and its warranty administrator use an AI assistant, and it is the shape that makes owners cancel the whole thing in a temper. Which is the wrong move, because February is also the month the shop closed 380 repair orders instead of its usual 240. The spend did not run away. The work ran away, and the spend followed it — nobody had put a number on the account, and nobody was watching.

Why February is the month it happens in this trade

Every truck shop north of the Mason-Dixon line knows the January and February pattern. DEF gels, doser injectors plug, air dryers freeze and cartridges dump water into the tanks, batteries that were marginal in October will not turn a cold 15-liter, and the road-call board never clears. Road calls triple. Your writers are opening repair orders one-handed while the other hand is on the phone.

That is exactly when an assistant earns its keep, and exactly when usage goes vertical. The writer dictates a two-minute voice note from the tech and gets a clean complaint-cause-correction written up. The warranty administrator drafts eleven claim narratives before noon instead of four. Somebody drops a sixty-page service bulletin in and asks what it actually says about the doser. Every one of those was worth doing. The bill was not a mistake. It was unmanaged.

Where the money actually goes in a twelve-bay shop

Before you cap anything, know what you are capping. In a heavy-duty service business the usage clusters in four places, and they are not equal:

• Repair order write-ups. Turning a tech's voice note or three lines of scrawl into a complaint, cause and correction a fleet customer will actually pay against. High volume, low cost each.
• Warranty claim narratives. The single most valuable use and the most expensive one, because a rejected claim gets re-argued. Your warranty administrator will use more than anyone else in the building, and should.
• Fault code triage. Pulling the codes off JPRO or INSITE into something the shop foreman can read and prioritize before he assigns a bay.
• Everything else. Translating notes for a Spanish-speaking tech, drafting the follow-up on an unapproved estimate, writing the customer letter about a comeback. Small individually, and it adds up quietly.

Budgeting AI in a service business means treating it exactly like a fuel card: a monthly ceiling for the business, a smaller ceiling per person, an alert before you hit either one, and a review that ties the spend to the work that caused it.

What the 2 July governance update actually gave an owner

On 2 July 2026, Claude's enterprise governance update added the four things that were missing for anyone signing the check: a dashboard that shows usage and cost, spend limits you can set for the whole organization and for individual people, alerts that fire at 75% and 90% of the limit, and control over which model people default to and what they are allowed to use at all. It also added reporting and administration screens so your office manager can pull the numbers without asking anyone technical.

The last one matters more than it sounds. In 2025 the answer to "who spent this" was a support ticket. In 2026 it is a screen your office manager opens on the first of the month next to the shop supplies line. That is the difference between a mystery charge and a budget line.

flowchart TD
  A["Owner sets the monthly shop ceiling"] --> B["Service writers seat limit"]
  A --> C["Warranty administrator seat limit"]
  A --> D["Foreman and parts counter seat limit"]
  B --> E["Alert fires at 75 percent"]
  C --> E
  D --> E
  E --> F{"Is spend tracking with repair order count?"}
  F -->|Yes| G["Raise the ceiling before the 90 percent alert lands"]
  F -->|No| H["Office manager pulls the usage report and finds the seat"]

Set the ceiling per repair order, not per month

A flat monthly cap is why shops choke in February. Your work is seasonal and your cap should be too. Price it the way you price shop supplies: as a small charge attached to each repair order, then multiply by the repair orders you expect.

Those dollar figures are an illustration for a shop that size, not a quote. The discipline is what transfers: at seventy-five cents per repair order, a February with 380 repair orders should cost about $285, not $610. When the alert fires and the repair order count says the work is really there, you raise the ceiling on the spot and move on. When the alert fires and the repair order count is flat, you have found something to look at.

Two ways to save money that do not involve saying no

Model defaults are the cheaper lever and almost nobody uses them. Most of what a shop does — turning a voice note into a write-up, tidying a customer email, summarizing fault codes — does not need the most expensive model in the catalog. Set the default to the mid-tier one and let people ask for the heavy one when they are arguing a denied warranty claim or reading a two-hundred-page contract from a national account. That alone typically takes a chunk out of the bill without anyone noticing a difference in the work.

Entitlements are the second lever, and it is really a security control wearing a cost-control hat. Your parts counter does not need access to everything. Your night writer probably does. Deciding who has what, once, in an admin screen, is a twenty-minute job that also answers the question your insurance carrier will eventually ask about who can see customer data.

What not to cap, and the review that takes twenty minutes

Do not put a hard stop on the warranty administrator in the last week of the month. That is when claims have to be filed inside the manufacturer's window, and a claim that misses the window is worth far more than the entire month's AI bill. Give that seat a soft limit and an alert, not a wall.

Do not cap anything attached to the after-hours line or the road-call board. A capped assistant at 11pm in February means a call goes unanswered, and an unanswered breakdown call is a tow that goes to the shop down the road plus a fleet that starts wondering about you.

And do not use spend limits as a substitute for training. The 2026 surveys keep saying the same thing — about two-thirds of small businesses now use AI in some form, and roughly seven in ten owners say their people need more training on it. Capping an untrained user saves you forty dollars and costs you the improvement. A ninety-minute session with your writers on what to hand it and what not to is worth more than any limit you set.

The review itself is short. First of the month, with your office manager, three numbers: total spend, spend divided by repair orders closed, and the top three seats. If the per-repair-order number is stable, you are done. If it jumped, look at the seats. If one seat is running away, that person has either found something genuinely valuable that you should copy across the shop, or they are pasting entire manuals in when a page would do. Both are worth five minutes of conversation, and neither is worth cancelling the account over.

Frequently asked questions

Our shop management system has AI built in now. Do I still need limits?

If the AI is bundled into your Fullbay, Karmak or Excede subscription at a flat price, that piece is already capped for you — that is what flat pricing means. Limits matter for the accounts you buy directly, which is usually where the warranty administrator and the writers end up, because the bundled features rarely cover claim narratives well. Know which of your uses is flat-rate and which is metered, and put ceilings only on the metered ones.

What is a sane starting number for a shop my size?

Take your repair orders per month, multiply by a dollar, and set that as the ceiling for the first sixty days. It will be too high, which is what you want at the start — you are trying to learn the real shape of the usage, not throttle it. After two months you will have your own per-repair-order number and you can tighten to it.

Who should own this — me, or the office manager?

The office manager owns the screen and the monthly report. You own the ceiling and any decision to raise it. That split works because raising the ceiling is a business decision about whether the work is really there, and reading the report is bookkeeping. If you own both, the review stops happening around March.

If we hit 90% and everything stops, what breaks first?

Write-ups, which means invoicing slows, which means cash slows. That is why you set the alerts and act on the first one instead of the second. Treat 75% the way you treat the low-DEF light: it is not an emergency, it is a reminder that you have a decision to make in the next couple of days.

Where the phone fits in the budget

CallSphere builds AI voice and chat agents that answer business phone lines and web chat, book appointments and capture leads around the clock — and it is priced as its own line, not as part of whatever your writers spend on write-ups. That separation is useful when you are budgeting: the phone line is a fixed cost you set against missed calls and after-hours tows, while assistant usage inside the shop moves with your repair order count. Two different lines, two different reviews, and neither one should be a surprise on the February statement.