A Job Shop's AI Bill Went $60 to $1,900 a Month: July's Spend Ceilings Stop It Without Slowing the Estimator

January's AI bill was $60. June's was $1,900, and nobody in the shop could say why

The first month you turned it on, the assistant that reads incoming RFQ packets cost sixty dollars — a rounding error next to the perishable tooling budget. Five months later the statement showed $1,900 for June alone, and nobody at the estimating desk could say why. The estimator had started feeding whole customer specification trees through it — the 240-page aerospace supplier quality manual, the packaging standard, the eleven referenced sub-specs — because it worked, and because no meter on the wall told anyone to stop.

That is the shape of the problem in a 14-person job shop in the summer of 2026. The AI is not failing. It is working well enough that people use it more than you budgeted for, on a bill that arrives after the money is spent. That is the same problem you had the first year machinists ordered their own end mills from MSC without a purchase order.

AI spend governance means every person and every job in your shop has a dollar ceiling on what they can spend on AI, an alert before they reach it, and a report afterward that shows which quotes the money went to. Until 2 July 2026, that report did not exist in a form an owner could actually read.

How the money leaks in a shop that quotes for a living

Quoting is where AI spend concentrates in precision machining, because quoting is where the documents are. A single RFQ arrives as a zip: a 2D PDF print with the title block, a STEP file, a purchase order terms sheet, and — if the customer is aerospace or medical — a supplier quality requirement document that references six others by number. Reading all of it carefully takes an estimator the better part of an hour. Handing it to an assistant takes ninety seconds, and it comes back with the material callout, the tightest tolerance on the print, the thread specs, the finish note, and a flag on the line buried on page three that says first article inspection per AS9102 is required.

So the estimator does it on every RFQ. Then the lead programmer uses it to summarize customer revision letters. Then the quality manager cross-checks your AS9100D procedures against a customer's new supplier manual. Then somebody discovers it will read scanned 1994 prints and starts running the archive through it. Every one is a reasonable use. Added together with no ceiling, they are why June cost $1,900.

The workaround most shops used through the spring was blunt: one shared login on the estimator's machine, and a rule that nobody else touches it. That controls spend the way locking the tool crib does — it works, and the second-shift setup guy stops asking for the carbide he actually needs.

What arrived on 2 July, in plain terms

The Claude Enterprise governance update on 2 July 2026 added four things that matter to an owner and nothing that matters to a programmer. First, a usage and cost dashboard: what was spent, by whom, on what. Second, spend limits at both the organization level and the individual user level — a hard ceiling per person per month. Third, alerts at 75% and 90% of a limit, so you find out before the ceiling is hit. Fourth, model defaults and entitlements: you decide which model each role gets, and which ones they may reach for at all. There are also Analytics and Admin APIs, which is the part your bookkeeper's software person cares about and you do not.

The 2025 version of this was a bill and a hope. The 2026 version is the control you already apply to shop supplies: a budget per cost center, a signal before you blow through it, and a report of what got made with the money.

flowchart TD
  A["RFQ packet lands in the estimator's inbox"] --> B{"More than 20 pages of customer spec?"}
  B -->|No| C["Fast model reads title block, material and tolerances"]
  B -->|Yes| D["Full spec-tree read on the heavier model"]
  C --> E["Cost posted against the estimator's monthly ceiling"]
  D --> E
  E --> F{"Estimator past 75% of ceiling?"}
  F -->|No| A
  F -->|Yes| G["Alert to owner; remaining RFQs run on the fast model"]
  G --> A

Put it in the burden rate, next to the $9.40 machine hour

You already know how to do this. When you set the shop rate on the VF-4 you did not guess — you took the machine payment, the floor space, the power, the perishable tooling and the coolant, and divided by expected spindle hours. AI belongs in that same calculation, on the estimating side rather than the production side.

The practical setup in a shop of a dozen or so people looks like this. Organization ceiling: a number you can say out loud without flinching — call it $450 a month. Individual ceilings underneath it: estimator $250, lead programmer $120, quality manager $80, office manager $40, with the balance unassigned so you can hand out more during a fourth-quarter surge. Model defaults: the cheap model for RFQ triage, email drafting and revision-letter summaries; the heavier one entitled to the estimator alone, for full spec-tree reads. Alerts at 75% and 90% route to you, not to the user, because the user will simply stop using the thing rather than ask.

Then review it monthly beside tooling spend. Not because the figure is large — it is smaller than what you spend on inserts — but because the trend says something. Rising while quote volume is flat means somebody is running the archive through it. Rising alongside quote volume is what you want.

The number that matters is cost per quote, not cost per month

$1,900 a month sounds alarming. Divided by the work it did, it may be cheap or it may be indefensible, and the only way to know is to divide. Here is a worked example with assumptions you should replace with your own.

At a controlled $310 a month, the AI costs $3.44 per quote against $23.10 of estimator time returned — a clear win, and it also means the quote goes back the same day instead of Thursday. At the uncontrolled $1,900, it costs $21.11 per quote against $23.10 returned, which is a coin flip you are paying to participate in. Same tool, same shop, same month. The difference is entirely which model was doing which job and whether anyone was watching.

The second number worth tracking is the share of RFQs quoted within 24 hours. On jobs under about $8,000, turnaround moves hit rate more than price does.

What a spend cap will not do for you

A ceiling stops a runaway bill. It does not stop a wrong answer. If it misreads a profile tolerance on a datum-referenced feature and your estimator quotes it as a general tolerance, the cap saved eleven dollars and cost you a job you will run at a loss for three years. Every number on a quote still gets read by someone who has stood at the machine.

A cap also will not tell you whether the spend was worth it — it only tells you it stopped. Governance gives you the meter, not the judgment. And it does nothing about the uncomfortable question in a job shop: whether customer prints under export control should be going through an outside service at all. That is a policy decision you make once, in writing, and it sits above the dollar limits.

Finally, if you cut the ceiling too hard, people stop using it, revert to reading the spec manual by hand, and your quote turnaround slips back to Thursday. That failure is quiet and it will not show up on any dashboard.

What to do before Friday's payroll run

Pull the last three statements and divide each month's total by the quotes you sent that month. That number tells you whether you have a problem or a bargain. Then set one organization ceiling, three or four individual ceilings, and route the 75% alert to your own phone. Default everybody to the cheap model and give the heavier entitlement to the estimator alone. Put a line called "quoting software and AI" on the same review page as perishable tooling, and look at it ten minutes a month.

Frequently asked questions

Is $450 a month a sensible ceiling for a 14-person shop?

It is a reasonable start if your main use is quoting and document reading and routine work runs on the cheap model by default. Judge it after two months on cost per quote, not on the monthly total. If quote volume climbs in the fourth quarter while customers burn capital budgets, raise the ceiling for the quarter and put it back in January.

Can I see which estimator or programmer spent what?

Yes. The July governance update reports usage and cost at the user level, which is what makes individual limits enforceable rather than decorative. Tell your people up front that you can see it; finding out sideways breeds the shared-login habit you are trying to end.

Should I let the assistant read customer prints at all, given ITAR?

That is a separate decision from spending, made in writing with your quality manager before you touch the dollar limits. Many shops keep controlled technical data inside systems they control and use the assistant only on commercial work. Whatever you decide, write it into your procedures so your AS9100D auditor sees a controlled process rather than a habit.

What happens when someone hits their ceiling in the middle of a big RFQ?

The 90% alert should reach you first, and raising one user's limit takes a moment. The failure to avoid is an estimator quietly no-bidding work because the tool stopped and he did not want to ask. Say out loud, once, that hitting a ceiling during a live RFQ is a phone call to you and not a problem.

While you are metering AI spend on the quoting desk, the shop phone rings with a buyer chasing a promise date, a mill confirming a heat lot, and a first-time caller with a one-off bracket. CallSphere builds AI voice and chat agents that answer a business phone line and web chat around the clock, capture who called and what they need, and book the callback — so the RFQ that lands at 4:50 on a Friday still gets logged with a name and a part number. It does not price your work or read your prints; your estimator does that.