Your AI Bill Went From $180 to $2,600 a Month and Nobody Can Say Which Quote It Priced

"Nobody approved this and I cannot tell you what it bought"

That is the objection, and it is fair. Eighteen months ago your AI spend was a $20 subscription your NPI engineer expensed. Then the quoting desk used it on BOM scrubs because it worked, program managers drafted customer updates, quality wrote up corrective actions, shipping did export paperwork. Last month the card statement showed $2,600 and your controller asked a fair question: which of that priced which quote?

The instinct is to shut it off, or pick one tool and ban the rest. Both are mistakes — the quoting throughput is real and visible in your RFQ log. What you have is a cost you cannot attribute and an access problem you have not examined. As of 2 July 2026 there are proper controls for both.

Budgeting AI in a contract assembly shop means treating it like line time: a cost per unit of work you can name — per quote packet, per BOM line, per corrective action closed — with a hard monthly ceiling, per-person limits, and an alert before you hit the wall rather than after.

Where the money actually goes in a board shop

Before capping anything, find out what you are buying. The spend lands in four places:

• The quoting desk. The largest share by far: scrubbing a 400-line BOM against distributor pricing, checking lifecycle and change notices, reading fab drawings for process flags. This spend earns money, and it is the one people panic about because the number is big.
• Program management correspondence. The weekly status to a medical OEM, an engineering change summary, the shortage report. Small per use, high volume, easy to forget.
• Quality documentation. Corrective action write-ups, audit prep before AS9100D surveillance, first-article narrative. Bursty: it spikes six weeks before an audit, then goes quiet.
• The experiment nobody told you about. One engineer trying something clever on test data, running the same job forty times. This is where a surprise bill comes from.

Three of the four tie to a countable event: quotes, programs, corrective actions. That is what makes budgeting possible — you are not capping a mystery, you are pricing work you already count.

What the 2 July governance update actually gave you

Claude Enterprise's governance release on 2 July 2026 added the controls an owner needs: a dashboard showing cost and usage by person and by model, spend limits at organisation and individual level, alerts at 75% and 90% of a limit, a default model setting, and entitlements — who may use which capabilities at all. Reporting and admin tools came with it, so your controller pulls the numbers into her own spreadsheet.

The difference from 2025 is not subtle. Before, an owner's options were a shared login and hope, or personal subscriptions with no visibility. Now it behaves like any other utility you buy: a meter, a ceiling, a warning light, per-person switches. That makes it a budget line instead of a surprise.

flowchart TD
  A["Month opens: shop-wide ceiling set at $3,000"] --> B["Per-person limits: quoting desk higher, everyone else lower"]
  B --> C["Dashboard shows spend by person and by model"]
  C --> D{"Alert fires at 75 percent"}
  D -->|Before the 20th| E["Controller checks quotes returned against spend"]
  D -->|After the 20th| F["Normal month, review at close"]
  E --> G{"Cost per quote packet climbing?"}
  G -->|Yes| H["Move routine BOM lookups to the cheaper default model"]
  G -->|No| F
  H --> A

Caps that follow the quoting calendar, not the accounting calendar

Here is where owners get it wrong: one flat monthly ceiling, which strangles them in the weeks it should not. This trade has a shape. October through December, OEMs push RFQs out to lock next-year pricing and quote volume can double. January and February, everyone quotes pre-buys ahead of Lunar New Year shutdowns at the bare-board fabricators. Summer is quieter. Your AS9100D surveillance audit lands on a fixed date, and quality's documentation spikes for six weeks before it.

So set the ceiling seasonally. A $2,000 shop-wide limit from May through September and $3,500 from October through February costs less over the year than a flat $2,800, and will not stop your estimator on 12 December — the day stopping costs you a program. Set individual limits the same way: buyer and estimator get room, program managers a moderate number, everyone else a small allowance that covers real use and caps the experiment.

Send alerts to two people: the 75% to whoever runs the desk that is spending, the 90% to your controller. If the 90% arrives on the 14th, that is a conversation, not a shutdown — sometimes you quoted eleven extra jobs and the money was well spent. Set the cheaper model as everyone's default while you are in there; letting only the quoting desk step up often takes 30% off a bill with no difference in output.

Entitlements: the export control problem you have to solve first

If any of your work is ITAR-controlled or falls under EAR, this matters more than the cost section. Gerbers, fab drawings, netlists and assembly documentation for a defense program are controlled technical data. Who may access them, and where they may be processed, is not a preference — it is a legal obligation with penalties that dwarf your AI bill. If you hold or are pursuing CMMC Level 2, your assessor will ask how that data is handled in every tool your people use.

Entitlements are the practical answer: decide by person who may use these tools at all, and write a rule that packets flagged as controlled do not go into a general-purpose tool without export-control review. In a 60-person shop that is six names and one line in the quality manual. Write it before the budget — a spend cap on a control violation is not much comfort.

Do the same for customer confidentiality: NDAs with medical and consumer OEMs often say designs are not shared with third parties. Most customers are fine with these tools when told and almost none are fine finding out later, so send your top ten a paragraph on what you use and how their files are handled.

The arithmetic: what a quote packet actually costs

Stop looking at the monthly total and look at the unit. Illustrative assumptions: 24 quote packets last month averaging 380 BOM lines each, 41 program status documents, 9 corrective action write-ups, plus general shop use.

Now the comparison that matters. Your estimator's loaded cost is $86 an hour. If the scrub saves three and a half hours per packet, that is $301 of desk time against $59 of spend, and the quote goes out two days earlier. At a 24% win rate, $1,420 a month buys the work behind roughly five and a half won programs. Written that way, the argument stops being about the bill.

The number to watch month over month is cost per quote packet. If it climbs from $59 to $95 with no change in BOM size, something changed on the desk, and that is a five-minute conversation. If total spend rises because you returned 34 quotes instead of 24, that is not a cost problem. That is a good month.

What a spend cap will never tell you

A dashboard shows what was spent, not whether it was any good. It cannot tell you Tuesday's exceptions list missed a product change notice. Output quality stays a human judgement, checked by spot-checking ten priced lines on every quote by the estimator who signs it.

A cap will not manage the political side either. Announce a ceiling and give nobody an allowance, and people go back to personal subscriptions on expense reports — losing you the visibility you just bought. Hidden spend comes from rules that leave people no honest way to work.

And do not confuse the meter with the value. Deloitte's 2026 survey found 84% of organisations investing in AI report positive returns and two-thirds of US small businesses now use it, but roughly 70% of owners say their people need more training. The gap between a $2,600 bill that earns and one that does not is usually training, not tooling. Budget three hours per person to sit with your buyer and estimator.

The Monday step: pull last month's spend, split it into the four buckets above, and divide the quoting bucket by the quotes you returned. That number — dollars per quote packet — is what you manage by for the next two years.

Frequently asked questions

What is a sane starting ceiling for a 60-person assembler?

Anchor it to the work, not a round number. If you return roughly 24 quotes a month and the scrub is the main use, a shop-wide ceiling of $2,500 to $3,000 with fourth-quarter headroom is a reasonable start. Watch cost per quote packet for two months and adjust. A percentage of revenue is a worse anchor — it has nothing to do with what drives spend.

How do I stop one engineer running up the bill on an experiment?

Per-person limits, which is exactly what the governance update added. Give the engineer a real allowance, not zero, and an alert at 75% that lands in his inbox before yours. Most runaway spend is not misconduct — it is somebody looping the same job, and a warning ends it that day instead of at month close.

Does any of this touch the EU AI Act or the state laws?

Cost governance itself does not, but the same controls answer the questions that do. The EU AI Act's high-risk and transparency obligations carry a 2 August 2026 compliance date and reach US companies whose systems affect EU users; Texas TRAIGA and California SB 53 took effect 1 January 2026. For a contract assembler the exposure is narrow — hiring and employee monitoring, not board assembly — but knowing who uses which tools is the first thing those questions ask.

A note on the calls this generates

One thing worth budgeting separately: the phone. Faster quotes mean more callbacks, more status chasing, more inbound from buyers who want a number by end of day. CallSphere builds AI voice and chat agents that answer business phone lines and web chat, capture who called and which program they are asking about, and book the callback — a fixed monthly cost on a budget line like any other, rather than a headcount decision you make in a busy quarter and regret in a slow one.