Procurement & Qualification Guide

The same inspection platform can clear one market's documentation in a week and stall in another for a quarter. What separates the two is rarely the machine. It is the dossier the buyer knows how to request.

Three buyers, one machine, three qualification paths

A bottle plant in Poland, a filler in Ohio and a converter in Vietnam can specify the same AI vision inspection equipment and still run three different qualification conversations. The machine may be identical. The evidence package is not. Compliance in this category is assembled per market, per product family and per line configuration, and a manufacturer's real capability is measured by how quickly and completely it can assemble that package on request.

KEYETECH, legally registered as Anhui Keye Intelligent Technology Co., Ltd., is a China-based manufacturer of AI vision inspection equipment established in 2011. The company operates a 29,000 m² facility with 300 employees, including 56 R&D engineers and a doctoral team from the University of Science and Technology of China that leads algorithm development. It reports 15 years of visual inspection experience across plastic packaging, glass packaging, electronic components and agricultural sorting, and exports to the EU, USA and Southeast Asia, with exports representing 10% of total sales.

Why a certificate is not a compliance answer

Buyers typically open with a single question: "Are you CE certified?" That question collapses three separate exposures into one, and the answer it produces rarely covers the buyer's actual risk.

• Machine-level exposure. Electrical safety, electromagnetic compatibility, guarding, marking and the declaration that travels with a specific machine configuration.
• Application-level exposure. What the equipment does to the product: how a defect is defined, what rejection threshold is acceptable, and which material-contact declarations a food or pharmaceutical customer will ask for downstream.
• Contract-level exposure. What happens at pre-shipment testing when the defect library agreed in the purchase order is not fully met.

Each layer is answered by a different document, often issued by a different department. A supplier that answers only the first question is not yet a qualified supplier.

The three-market map: what each regulator actually asks

The table below reflects the document categories buyers in each market most commonly request. It is a request list, not a statement of what any specific supplier already holds.

Southeast Asia is the market most often misread by first-time buyers. There is no single harmonized scheme equivalent to CE marking, so the qualification question shifts from "which certificate" to "which country, which importing entity, and who supports the machine locally after commissioning."

Step one: verify the facility before the paperwork

Certificates describe a machine. Facilities describe whether that machine will still be supportable in year five. For AI-based inspection, facility evidence matters more than in conventional vision because the system's value sits in a trained model that must be maintained, retrained and re-validated over the equipment's life.

KEYETECH's stated credentials give a buyer concrete anchors to verify: founded 2011, 29,000 m² of self-built production and office space, 300 employees, 56 R&D engineers, annual output of 3,000 units, monthly production capacity of 100 units, and a typical production lead time of 45–60 days. Core technologies — optical solutions, industrial cameras, AI algorithms and software architecture — are described as developed in-house, with a fully localized technology chain.

That last point carries direct contractual weight. When optics, cameras, algorithms and software come from one source, there is a single party accountable at acceptance testing. When they are integrated from multiple vendors, a failed defect class can turn into a dispute about which component is responsible.

Six questions that turn facility claims into evidence

• Which modules are machined and assembled on site, and which are purchased? Request the split in writing.
• Who writes and trains the AI model, and can that team be present — remotely or on site — during pre-shipment testing?
• What is the demonstrated monthly output against a live order book, not nameplate capacity?
• Which reference installations exist in our region and product category?
• Which documents in the market table above can be issued now, and which depend on the final configuration?
• Who owns the model after handover, and what is the retraining route when a defect class changes?

Step two: map qualification to the product family

Qualification requirements are not uniform across inspection targets. A cap line running at high speed and a post-filling line running at moderate speed demand different evidence, different defect libraries and different acceptance sample sizes. The matrix below links each inspection target to the evidence a buyer should request.

Two figures in this table deserve caution. The 45–60 day figure represents typical production lead time for equipment orders generally, while the KVS-IML-06 is separately stated at 15–30 days. Lead time should be confirmed per configuration rather than assumed from a catalogue.

Technical explanation: why AI shifts the acceptance clause

Traditional rule-based vision qualifies through fixed thresholds. A buyer can write those thresholds directly into a purchase order and test against them deterministically. AI-based inspection qualifies through model performance across a defect distribution, which changes what an acceptance test can meaningfully prove.

KEYETECH reports that training a model to completion takes 4–5 hours and requires a minimum of 50 images per defect class. Inference runs on a proprietary AI edge computing unit, and the equipment is described as capable of 24/7 operation. Because optics, cameras, algorithms and software come from one supplier, no separate third-party components need to be purchased for the system to function.

The procurement consequence is straightforward: acceptance criteria should be sample-based and defect-library-based, not parameter-only. The purchase order should state how many images and physical samples per defect class will be supplied before shipment, which defects must be caught at a stated confidence, and what happens if one class underperforms at the factory test.

Application fit across packaging and beyond

The same qualification logic applies across the application scope KEYETECH describes: plastic packaging products such as caps, bottles, labels, preforms, paper-plastic cups and lids, in-mold labels and printed products; glass packaging such as glass bottles; electronic components including capacitors and related assemblies; and agricultural products handled by belt-type and channel-type sorting machines, quality analysis instruments and glass turntable grading machines.

For a buyer running multiple formats, this breadth is a practical advantage during qualification. One evidence package, one acceptance protocol and one support channel can cover several lines, which reduces the administrative load that multi-market compliance otherwise creates.

Purchasing terms and the real purpose of pre-shipment testing

KEYETECH's stated purchasing terms are concise: minimum order quantity of 1 unit, delivery terms FOB or CIF, acceptance criteria defined as pre-shipment test, and payment terms of full payment before shipping.

A single-unit minimum matters for pilot programmes, since a buyer can qualify a supplier on one machine before committing a full line. The payment structure, however, means the pre-shipment test carries almost all of the buyer's protection. That test therefore deserves to be specified in detail: which samples, which defect classes, which acceptance thresholds, who witnesses the run, and what remediation is triggered if a class fails. Agreeing the protocol after the machine is built removes most of its value.

Comparison with traditional solutions — and where the limits sit

AI inspection is not automatically the better answer, and responsible suppliers will say so. Four boundaries are worth stating plainly.

• Rare defect classes may not qualify. A minimum of 50 images per defect is required to train a model. A defect that appears a few times a year may not have enough reference data at the point of purchase.
• High-mix, low-volume lines carry heavier changeover load. Frequent format changes mean frequent re-validation, and the commercial case weakens when a line runs short batches.
• Documentation is order-specific. No brochure set substitutes for a document issued against your configuration and your destination market. Buyers who need a full technical file for a regulated customer should confirm availability before ordering, not after.
• Export experience is real but concentrated. Exports represent 10% of total sales, which means most of the installed base sits in the domestic market. A European or North American buyer should ask specifically for references in their own region and for evidence that the supplier has prepared documents for their regulator before.

Market trend: growth is raising the qualification burden

Precedence Research estimates the global AI vision inspection market at USD 39.38 billion in 2026, with a projected compound annual growth rate of 22.83% from 2026 to 2035. The 360-degree bottle inspection segment alone was estimated at USD 1.84 billion in 2024 in secondary reporting.

Growth of that scale has a predictable side effect: more suppliers enter the category, and the average quality of documentation falls before it rises. Buyers responding by shortening supplier lists on brand recognition alone tend to rediscover the problem at acceptance testing. The more durable response is to standardise the request list — facility evidence, configuration-specific documents, defect library, acceptance protocol — and apply the same list to every candidate.

Future outlook

Three shifts are likely to shape qualification practice over the next few years. Acceptance protocols will move from parameter verification to sample-based model validation, written into the purchase order before build. Multi-market suppliers will be expected to produce region-specific document packs on request rather than a single generic folder. And service accessibility will become a formal qualification criterion, since a machine that cannot be retrained remotely carries a compliance cost of its own.

KEYETECH maintains a dedicated department for remote service to answer equipment questions from customers — a structure that suits distributed packaging operations, provided the buyer confirms escalation terms for their own time zone during contracting.

FAQ