Choose an AI visual inspection vendor by testing it on your own parts, on your own line, before you sign. Brochure accuracy figures, demo videos and customer logos tell you little about how a system will handle your defects, your lighting and your cycle time. The vendors worth shortlisting are the ones willing to prove it on your parts and to put the result in writing.
This guide is for plant heads, quality heads and automation engineers in India comparing AI visual inspection systems. It lists what to check, what to ask, and the red flags that should end a conversation early.
Start with the problem, not the vendor list
Before you call anyone, write down four things. Vendors who are a good fit will ask for them anyway.
- The defects you need to catch, with photos of good parts and of each defect type, and how often each defect occurs.
- Where the check happens: the station, the line speed or cycle time, and the space available for cameras and lighting.
- What happens to a reject: a PLC signal, an air jet, a stopped line or an alert to an operator.
- How you judge success: how many defects you can afford to miss, and how many good parts you can afford to reject.
The last point matters most. A system that catches every defect but rejects too many good parts will be switched off by the line within weeks. Agree both numbers before the trial starts.
What to check in every vendor
| Check | What a good answer looks like | Why it matters |
|---|---|---|
| Trial on your parts | A paid or free trial on your line or with your parts, with a written result | Accuracy depends on the part, the defect and the lighting, not on the software alone |
| How accuracy is measured | A stated method: a baseline set checked by your quality team, a parallel run against manual inspection, and both miss rate and false-reject rate reported | A single “accuracy” number can hide a high false-reject rate |
| Named references | Customers you can call, in a similar industry, with the system running in production | Pilots that never went live are common |
| Training data needed | A clear number of good and defect images, and who collects and labels them | Some systems need thousands of labelled defect images you may not have |
| PLC and MES integration | Named PLC families supported, the signal used for pass/fail, and how results reach your MES or quality system | Integration is where many projects stall |
| Where data lives | On-premise processing, or a clear statement of what goes to the cloud | Many plants do not allow production images to leave the site |
| Changeovers and new models | Who retrains the system for a new part or variant, how long it takes, and what it costs | Product changes are frequent; retraining should not need the vendor every time |
| Service in India | Engineers who can reach your plant, response times in writing, and spares held locally | A stopped inspection station can stop the line |
| Total cost | Hardware, software licence, installation, training, yearly support and retraining, all listed | Low upfront prices are often recovered through support and change fees |
Questions to ask in the first meeting
- Can you run a trial on our parts before we buy, and what will you put in writing at the end?
- How do you measure accuracy? Will you report both missed defects and false rejects?
- Which customers in our industry can we speak to, and is the system running in production there?
- How many images do you need to train the system, and who labels them?
- Which PLCs have you integrated with, and how does a reject signal reach our line?
- Does any image or data leave our plant?
- When we add a new part or variant, who retrains the system and how long does it take?
- Where are your service engineers based, and what response time will you commit to?
- What is the full cost over three years, including support and retraining?
Red flags
- A fixed accuracy figure before seeing your parts. No one can know it yet.
- No trial, or a trial only on the vendor’s own sample parts.
- Only a miss rate, no false-reject rate, or the other way round.
- References you cannot contact, or logos without a running system behind them.
- Retraining that always needs the vendor and is billed per change.
- Vague answers about where data is processed and stored.
Running a fair comparison
If you shortlist two or three vendors, give each the same set of parts: good parts and each defect type, checked by your own quality team. Keep a sealed set back that no vendor sees during training, and use it only for the final test. Compare the results on that sealed set, along with integration effort, service terms and three-year cost. Our accuracy methodology page shows how we run this kind of baseline and parallel run.
Where Indus Vision fits
Indus Vision is a Bengaluru company building DeepVision AI visual inspection. It can be trained with as few as 200 images, connects to cameras and PLCs in under 30 minutes, inspects in under 100 ms per part, works with Siemens, Allen-Bradley and Mitsubishi PLCs, and runs on-premise. Named results include 100% on the JBM Auto body-in-white line and 98.2% combined across two automotive OEM lines (1,073 samples), 98% or better on flat-panel boards, and 98% or better on SWIR inspection at Tata Steel; see our case studies. We also build DeepVision Tunnel for inline gap and flush measurement and DeepVision Warehouse for dock carton counting. We would rather be tested on your parts than chosen from a brochure.
Frequently asked questions
How long should an AI visual inspection trial take?
Long enough to see the defects you care about at their normal rate, and to run the system in parallel with manual inspection. For rare defects, collect samples in advance so the trial does not depend on them turning up.
Is a higher accuracy figure always better?
Not on its own. Ask for the miss rate and the false-reject rate separately, measured on your parts. A system that rejects too many good parts costs money and loses the trust of the line.
Should inspection run on-premise or in the cloud?
Most production lines need decisions in milliseconds and many plants do not allow images to leave the site, so on-premise processing is the usual choice. If a vendor uses the cloud, ask exactly what is sent and where it is stored.
Can we switch vendors later?
Ask who owns the trained models and the labelled images before you sign. If you own the images and labels, moving to another system later is much easier.