AI as cost of goods sold — recovery in custom manufacturing.

Manufacturing is the cleanest version of the cost recovery story, because the answer isn't "bill it back to the client" — it's "absorb it into COGS, where it already belongs." For a custom job shop, contract manufacturer, or anyone building to customer order, AI spent on a production run is part of the cost of producing that run. It rolls into per-unit cost, gets priced into the quote, and never touches SG&A.

COGS vs SG&A — why the difference matters

In a manufacturer's financial statements, gross margin is the headline number. Investors, banks, and the owner all look at it. Gross margin = Revenue − COGS. Anything that lives in COGS reduces gross margin; anything that lives in SG&A reduces operating margin but leaves gross margin untouched.

When AI gets paid for on the corporate Amex and dumped into SG&A — Software & Subscriptions, the gross margin on the underlying jobs looks artificially healthy — because the AI cost that helped produce them is sitting on a different page of the P&L. You're overstating gross margin and understating operating margin. The total profit number is the same, but you're flying blind on per-job profitability.

When AI is attributed to the job at the moment of use, it absorbs into COGS — labor, materials, machine time, and now AI. Gross margin reflects the real cost of production. Quoting next month's work gets more accurate. The CFO can answer "are we making money on this customer?" honestly.

Where AI shows up on the shop floor

In custom manufacturing, AI is increasingly used for:

• Custom engineering — interpreting customer drawings, generating BOM line items, validating tolerances against incoming specs.
• Vendor sourcing — researching alternate suppliers, comparing datasheets, drafting RFQs to the supply base.
• Quality documentation — drafting inspection plans, FAI documentation, PPAP narratives, root-cause writeups.
• Process planning — generating router sheets, identifying op sequence options, writing setup sheets.
• Customer-facing technical documentation — assembly instructions, O&M manuals, certificates of conformance.

Each of these activities is tied to a specific work order, job number, or production run. The shop floor system already knows which job is which.

Per-job and per-unit attribution

Job-shop accounting already absorbs indirect costs into jobs using burden rates — a shop overhead rate per machine hour, a labor burden rate per direct labor dollar. AI can be handled the same way, except more cleanly: instead of an averaged burden rate, the actual AI usage on the job rolls into the actual job cost.

For repeat parts, AI cost per unit becomes a quotable line in the standard cost. For one-off custom jobs, it becomes an actual charge against the work order. Either way, it stops being invisible.

The numbers for a small job shop

Consider a small custom job shop running about 12 production jobs in a typical month. Across custom engineering, vendor sourcing, spec review, and QA documentation, AI usage averages roughly $80 per job:

• 12 jobs × $80/job = $960/month
• $960 × 12 months = ~$11,500/year

That $11,500 either rolls into COGS on each job (where it gets recovered through the price you charged) or sits in SG&A where it erodes operating margin without ever showing up in your gross-margin analysis. The dollar amount is the same; the financial-statement implications are not.

The piece that was missing

Job costing in manufacturing has always required attribution at the moment of work — a timecard charged to a job number, a material issue tied to a work order, a machine cycle posted against an op. AI usage needs the same posting discipline: every query tagged to the job that triggered it.

Without that, the AI bill lands in SG&A and the job cost reports lie. With it, AI becomes one more cost factor that gets priced, quoted, and absorbed exactly the way every other direct cost is handled.

The accounting pattern is general — see the Cost Recovery overview for the full framework, or Interchange for the platform that makes per-job AI attribution operational.