Some time ago, I read about a reporter who was amazed by Tesla's speed in building one of its car factories. While that is no small feat, anyone who has experienced the reality of a plant knows that a much greater challenge is filling it. Ever since the Japanese popularised the JIT method, processes have had to run like clockwork, and recipes are endless. COVID was the final ingredient turning that clock into a weather vane. Amid these uncertainties, we must decide how to improve and automate daily operations. Standing still is not an option.
So today I am sharing a list of 10 key points so that, if you are thinking about automating a process on your line with AI, you can choose where to start wisely. Think of it this way: we have spent 6 years fighting to get that expert model up and running. These experiences were painful to acquire, but sweet to enjoy, and although many may seem obvious, in the heat of development we fall into these mistakes all the same. Ready? Go!
1. The project must have a clear ROI, even if it is small
If it is your first project and you pick any process just to learn the technology, it is very likely you will settle for completing only the algorithm part or programming logic. The reality is that successfully implementing AI means understanding every required operation from start to finish, and without a clear incentive (a real improvement for the plant) it is easy to throw in the towel when progress slows down. There has to be a reward worth fighting for.
2. If it has ROI, look for the quickest and simplest
The other extreme is choosing the project that delivers the highest return, which sometimes comes with high difficulty. It is generally better to take the opposite route: start from the easiest (that has ROI) towards the most difficult. But remember that not everything is measurable money. A process that matters to your customer, which you can showcase as a visible investment, also impacts strategy and is well worth the effort.
3. Do not approach an AI project as "all or nothing"
Speaking of difficulty, do not demand perfection from yourself just to pass. Many projects, with multiple features and developments, are set up with a single end goal, when, for instance, cameras could be used to build a history of passing parts, measure machine uptime, or detect only certain defects rather than all of them. If the reward is 9 months away, early stages tend to be discouraging and we end up with an expensive toy. Reward without intermediate milestones leads to burnout.
4. Start with simple detections
Broadly speaking, the easiest task currently is detecting the presence or absence of objects, or classifying clearly distinct objects (such as a piston and a hammer). For measurements, complex digit detection or difficult classifications (small cracks, dents), the required data volume increases significantly:
Task type | Approx. required images |
|---|---|
Presence/absence, highly distinct objects | 50 to 100 |
Measurements, complex digits, subtle defects (cracks, dents) | 200 to 500, or more |
5. Create test parts
In these learning applications, data is king, and having few samples per week will slow down progress significantly. If you can use parts that can be modified to represent the patterns you want to detect, in a matter of minutes you will have accelerated the project by months. But beware: you cannot take 50 photos of the same part without variation, because AI would not learn the pattern—it would only memorise that specific object.
6. Data that does not allow you to take action is not data
Another obvious point often overlooked. Some systems present relatively easy vision problems (OCR, high-speed product detection), but require complex actuators: a robot for marker marking, special ejection pistons. Sometimes this "dumb" part (the mechanics) can be just as expensive as machine vision and its computing power. Therefore, try to keep an open area with simple mechanical elements.
7. Hardware vs. software
Some projects require a lot of equipment, others many human hours of fine-tuning, and some both. Aim to keep at least 1 of the 2 options simple: a project with 1 camera (even if it demands careful analysis), or multiple cameras looking for the same simple defect. This makes it far more manageable than combining high complexity on both fronts at once.
8. Fast, precise and small: a rather expensive triangle
Typically, AI cycle times range between 200 milliseconds and 1 second. If you want to inspect parts faster than that, or detect defects with high precision (10 pixels/millimetre) across a very large or very small area, you are choosing a combination that rarely comes cheap. Evaluate whether one of the three parameters (speed, precision or scan area size) can be relaxed: you will notice a significant reduction in cost, because easing one usually allows the use of less advanced or more standard technology.
9. Team
An absolute must: there must be a representative from every team or department affected by the system. Production, quality, innovation... each wants a system that might affect the rest of the plant. Ideally, each representative presents their interests, goals and operational expectations from the start. You do not want to spend a month fine-tuning an algorithm only to discover that its defect detection precision is so high it rejects acceptable parts, forcing you to scrap all that effort. If only it didn't happen, but it does.
10. Reusability
The first project is a challenge, but subsequent ones are simplified tremendously, especially if they resemble each other: this reduces costs and improves the viability of the next project. If your candidate shares many aspects with planned future projects, that deserves at least an extra point in its favour when choosing where to begin.
Frequently asked questions
How many images are needed to start an AI project in a plant?
It depends on complexity: from 50–100 for simple detections (presence/absence) to 200–500 or more for subtle defects or precision measurements.
Is it better to start with the project that saves the most money or the easiest one?
Almost always with the easiest one that also offers ROI, rather than the one offering the highest theoretical return. A poorly executed complex project causes more discouragement than learning.
What happens if the system is too precise and rejects good parts?
It is a real risk if all affected departments are not involved from project design: quality, production and innovation must agree on the criteria before fine-tuning the model, not after.
Can you automate without buying new hardware?
In some cases yes, if the process already has usable cameras or sensors, but it is wise to evaluate from the start whether the challenge lies more in hardware, software, or both, so as not to underestimate the effort.
In summary
AI has the ability to improve a large number of processes if applied wisely. And the best way to learn how to "apply it wisely" is to start small but defined, without setting requirements that stall progress, but with milestones that convince the whole team of incremental improvements and actionable insights. That is why we have compiled these 10 concrete points to rate a project and choose a good starting point.
In future articles, we will explore how to develop the chosen project, with tips, rules and tools, to gain a complete end-to-end perspective when tackling projects with this technology. Above all, to empower our working environments and support our teams. They need us!
Which of these 10 points do you still need to address in your project?
Tell us where you are at, and we will help you select the best process to start with.