Production AI Is Fashion’s Largest Open Opportunity

This spring my company surveyed 250 fashion professionals in more than 90 countries. Each rated AI’s impact on their own business, and the scores split by layer inside the same companies. Marketing and content tools scored 3.8 out of 5. Production and technical tools scored 2.1, and The State of AI Fashion 2026 calls that 1.7-point difference the production gap.

The gap matters to margin because the two layers pay out on different terms. Every competitor rents the same image models at the same price, so a saving on campaign imagery drifts to whoever holds the buying power, while production AI runs on the blocks, materials, tolerances and graded specs a brand and its factories already hold, and no rival can rent that history.

The scores by workflow

Adoption follows the same split. Seventy percent of respondents use AI for sketch and concept work, and 25 percent have tried it on a tech pack. On a five-stage maturity scale, creative direction averages 3.2 and merchandising and launch 2.8. Production readiness sits at 1.9, and production handoff, the step where an approved spec becomes a factory’s job, comes last at 1.6.

Table 1. Marketing layer and production layer in the State of AI Fashion survey

Measure	Marketing and creative layer	Production layer
Impact rating out of 5	3.8	2.1
Workflow maturity out of 5	Creative direction 3.2, merchandising and launch 2.8	Production readiness 1.9, production handoff 1.6
Adoption	70% use AI for sketch and concept	25% have tried AI for tech packs
Year the capability became usable	2022	2024 and 2025, as drafts for technical review

Both columns come from the same brands, so a brand at stage three on campaign imagery can still send its factory a stage-one document.

The order of purchase explains most of the gap. Creative AI became usable in 2022, and technical-design AI became reliable enough to draft a factory-ready spec for review in 2024 and 2025, about three years later. Brands bought creative tooling first, and their budgets, vendor lists and training plans still run in that order.

Sampling rework starts with the spec

On a recent factory tour I watched how much of sampling is back and forth. A sample passes its fit session only when the body fit, the garment fit, the way the fabric behaves and the skill on the sewing line all agree, and a miss on any one of them sends it back for another round.

The cause sat upstream of the cutting table. Each extra file, version and person in the handoff gave the record another place to drift, and the sampling and fit iterations alone added 10 to 15 percent of overhead and about two weeks to the calendar.

Survey respondents report a hand-built tech pack takes 4 to 16 hours depending on the garment. Forty-five percent spend more than four hours on each pack and 20 percent more than eight, and the 10 percent who outsource the work pay $125 to $1,900 a style, about $300 on average.

The hours compound at every handoff. A collection runs on nine tools on average, and someone retypes the data each time it moves between the line sheet, the PLM, the costing spreadsheet and the factory’s planning system. I track this as the re-key count, the number of times one style’s specification is typed by hand before the factory cuts it. Every retype is a point where a measurement can read 101 in Bengaluru and 98 at the brand, and the disagreement stays hidden until the fit session.

The same record problem runs to the end of a garment’s life. In March 2023 I wrote for Newsweek on Web3 product passports, as EU rules brought in Digital Product Passports that carry a product’s lifecycle data, and a passport is only as accurate as the spec the factory cut from.

Brand A cut its sampling spend by 58 percent

The clearest numbers I have come from one brand my company measured as its software vendor, with the results reviewed with the brand’s finance team for fiscal 2025. Brand A stays anonymous here, as it does in the book, and I am the vendor in this case, so weigh the figures with that in mind.

Table 2. Brand A sampling and spec measures, before and after, fiscal 2025

Measure	Before	After
Hours per released tech pack	9.2	1.1
Sample rounds per style	3.8	1.6
Physical samples	228	96
Annual sampling spend	$186,960	$78,720
Average calendar slip per style	4.1 weeks	0.6 weeks

Sampling spend fell by $108,240 in one year, and calendar slip fell by 3.5 weeks per style.

The calendar line carries more weight than the cash line on a seasonal range. A style that slips four weeks can miss its delivery window. It then reaches stores with less of the season left to sell at full price, while a style that slips less than a week still lands on its planned floor set.

The cost of one sample round

The sample round costs more than the pack. A round costs $200 to $1,500 in the survey, with a median of about $750. Forty-five percent of respondents need four to five rounds to approve a style, and respondents who work from AI-generated specs cluster at one to two.

Take one style at the median as an illustrative case. Cutting it from four rounds to two saves 2 × $750, or $1,500. Against a $300 development package the brand keeps $1,200 per style before it counts any calendar gain, and a 40-style program at the same rates keeps $48,000.

Fund the factory handoff first

Sequence decides the return. A brand that adds a fourth image tool gets a saving its competitors get the same week, while a brand that fixes the handoff to its factory cuts rounds on styles only it makes. The book calls this second body of work the production base: fewer sample rounds, one approved tech pack at the factory, and written limits on what an AI agent may change without a person signing.

Fund the pilot so claimed hours stay apart from cash. Brand A’s sampling spend fell by $108,240 as physical samples fell from 228 to 96, and that saving shows up on invoices. Its 8.1 hours saved per tech pack count only when the brand stops paying for them or uses them to ship more styles, and a board that sees the two numbers on separate lines sees the cash trough before it arrives.

The move fails when the factory never works from the approved spec. A brand can produce a clean tech pack in an hour and still lose the gain if the factory retypes it into its own planning system, so the re-key count has to fall at the factory as well as at the brand. Ask each factory how many times it types your spec before cutting, and add that number to the baseline.

Factories can sell development as a product

The opening sits on the factory side as much as the brand side. In the book, Kaveri Garments, a composite factory of about 1,200 people near Bengaluru, prices development as a separate $300 package. The fee covers pattern work, the first proto and fit comments through to an approved fit sample, and sample making and production are billed separately. Starting each style on a block the line has already sewn brings variable cost down to about $170 a package, so the factory earns a margin on judgment it used to give away.

The book’s pivot score rates a move on six items from 0 to 3. Kaveri’s package scores 7, inside the pivot band of 5 to 13, because it changes how the factory sells and what it owns while stopping short of a transformation, which starts at 14.

Limits of the evidence

Every survey figure here comes from one source. The survey is self-reported and drawn from my company’s user community. Production and sourcing staff made up 19 percent of the sample and design and technical design another 34 percent, so the people closest to the tech pack wrote more than half the answers. The report also finds that brands with a positive return on production AI had first loaded their own approved blocks, materials and tolerances into the tools. The finding comes from self-selected platform users, so it shows correlation rather than cause.

Count your baseline first

Count the five Brand A measures on your own styles before you sign for a tool. They are hours per released tech pack, sample rounds per style, physical samples, annual sampling spend and calendar slip per style. Add the re-key count for one jacket or one dress, and the comparison will exist before the first invoice does.

The AI Pivot for Fashion Brands works through this decision with two composite companies, Ashcombe, a $60 million tailoring brand, and Kaveri, and scores 15 public fashion pivots on the same pivot score.

Questions

What is the production gap in fashion AI?

The production gap is the 1.7-point difference between how fashion professionals rate marketing AI, 3.8 out of 5, and production AI, 2.1, in The State of AI Fashion 2026 survey of 250 professionals. Nitin Kumar argues the cash a brand can keep sits on the production side.

How long does it take to make a tech pack by hand?

Respondents to the survey behind Nitin Kumar’s The AI Pivot for Fashion Brands report 4 to 16 hours per hand-built tech pack, depending on the garment. Forty-five percent spend more than four hours on each pack, and brands that outsource the work pay $125 to $1,900 a style, about $300 on average.

How much does a fashion sample round cost?

A sample round costs $200 to $1,500 in the survey, with a median of about $750. Cutting one style from four rounds to two saves about $1,500, and Nitin Kumar's worked example keeps $48,000 on a 40-style program.

Where does AI save money in fashion production?

In the tech pack and the sample round, according to The AI Pivot for Fashion Brands by Nitin Kumar. One brand cut sampling spend by 58 percent, $186,960 to $78,720 in fiscal 2025, as physical samples fell from 228 to 96. The book works through the decision with two composite companies.

Sources

The F* Word, The State of AI Fashion 2026: The Production Gap, Edition One, June 24, 2026. Survey of 250 fashion professionals in more than 90 countries, fielded February 15 to April 30, 2026.

Nitin Kumar, Web3 Product Passports: The Regulatory Use Case, Newsweek, March 24, 2023.