Updated July 2026 · 7 min read · Adente Vision Engineering Team
Why does a demo hide the numbers that matter most?
A demo is staged to succeed. The vendor picks the part, the lighting and the defects, and shows the system passing and failing exactly as scripted. What the demo rarely shows is how the system behaves on the defects nobody staged, how much of your own data and time it needs, and what happens after the sales engineer leaves. Those are the numbers that decide whether the system holds up on your line.
The way to see past a demo is to ask questions whose answers are checkable. A good vendor answers with a measured figure and the context around it. A weak vendor answers with an adjective, a single headline percentage, or a promise to sort it out later. The five questions below are the ones that most often separate the two, and each one has a good answer and a red flag you can listen for.
Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and the answers below are written the way we would expect any serious vendor, us included, to answer them.
What is the false-negative rate, not just the accuracy?
Ask for the false-negative rate, the share of real defects the system lets through, not just a headline accuracy number. Accuracy alone can look excellent while the system still misses defects, because on a line where most parts are good, a model that passes everything scores high on accuracy and catches nothing.
The false negative, also called an escape, is the expensive error: a bad part that reaches your customer. A good answer gives a measured false-negative rate with the context of where it was measured. On a live cap-inspection line, for example, Adente Vision recorded a 0.69% false-negative rate alongside a 99.65% F1-score at about 30 ms per part, and that context, the line and the conditions, is what makes the number checkable rather than abstract. A red flag is a vendor who quotes only a single high accuracy figure and cannot give you a false-negative rate at all. For the full explanation of why this metric matters more than accuracy, see the sibling post on the false-negative rate.
How many images and how long to a working model?
Ask exactly how much of your data the system needs and how long it takes to reach a working model. This question exposes the real setup cost, because a system that needs thousands of labelled defect images before it works carries a data-collection project you will pay for in time.
A good answer is a small, specific number. Adente Vision trains on good parts only: about 20 reference images to learn a part, with the model ready in under 48 hours. That is possible because the anomaly mode learns what a correct part looks like and flags deviation, rather than requiring a labelled catalogue of every defect first. A red flag is a vendor who needs a large labelled dataset of every failure mode before the first inspection runs, or who cannot commit to a time-to-model at all.
Where does the image data live, and who else can see it?
Ask where the images are processed and stored: on the line, or in the cloud. The answer tells you who controls your production imagery, what happens if the network drops, and whether a per-image or per-seat charge is buried in the running cost.
A good answer is on-device. Adente Vision processes every image on a fanless edge board inside the enclosure, so raw part imagery stays on the line, and model updates arrive by USB stick for air-gapped operation. That keeps sensitive production data inside the plant and removes a dependency on connectivity and cloud egress. A red flag is a system that must upload every image to a vendor cloud to run, which turns your inspection into a data-export arrangement and a recurring bill. For the wider trade-off, see the pillar guide on AI visual inspection.
Which protocols, I/O and enclosure rating ship out of the box?
Ask for the named protocol list, the discrete I/O count, and the enclosure IP rating for your actual environment, with a certificate on request. These are the specifics that decide whether the system drops into your cell or needs a bridge, a gateway or a rebuild.
A good answer names things. Adente Vision speaks PROFINET, EtherNet/IP, Modbus TCP, EtherCAT and OPC UA, and carries 4 inputs and 4 outputs at 24V for triggering and reject actuation. On the enclosure, it offers IP54, IP65 and IP66 ratings across its variants, declared against CE, IP and IEC standards, and a serious vendor will hand you the specific certificate and datasheet when you ask. A red flag is "we can add that protocol later," a vague "industrial rating," or an IP claim with no document behind it. The IP rating itself is defined by IEC 60529, so it is a checkable standard, not a marketing term. For the full connectivity and enclosure options, see the system overview.
The five questions, good answers and red flags
Take this table into any vendor conversation. The middle column is what a checkable answer sounds like; the right column is what should make you pause.
| Question to ask | A good answer sounds like | Red flag |
|---|---|---|
| False-negative rate, not just accuracy? | A measured FNR with context, e.g. 0.69% on a named cap line | Only one high accuracy number, no FNR |
| How many images and how long to a working model? | About 20 good images, model in under 48 hours | Thousands of labelled defects, timeline unclear |
| Where does the image data live? | On-device, on the line, USB updates, air-gap supported | Every image must upload to the vendor cloud |
| Which protocols and I/O out of the box? | PROFINET, EtherNet/IP, Modbus TCP, EtherCAT, OPC UA; 4-in/4-out 24V | "We can add that later" |
| Enclosure IP rating and certificate? | The IP rating plus certificate and datasheet on request | A rating claimed with no document |