7 questions to ask before starting an AI project in your factory

Artificial intelligence is on every meeting agenda now. Vendors, consultants and trade fair booths all promise to "boost efficiency with AI". But as a factory owner or production manager, where should you actually start? Most failed projects do not fail because of a technical mistake. They fail because the right questions were asked badly, or never asked at all.

In this guide we have gathered, in plain language, the seven questions to put on the table before you set aside a budget for an AI project. Behind them are public resources such as the AI Risk Management Framework published by the US National Institute of Standards and Technology (NIST) and the OECD AI Principles. These documents were written with large organizations in mind, but their core logic adapts easily to a small or mid-sized factory.

1. What exactly is the problem we want to solve?

"Using AI" is not a goal; it is a tool. A good opening sentence looks like this: "Defects missed during quality control on line B are causing customer returns." That sentence names a place, a loss and a consequence.

When you write down the problem, make three things clear:

• Who is affected? Operators, the sales team, customers?
• How is it solved today? By hand, by experience, or not at all?
• What changes if it is solved? Less scrap, faster quotes, less downtime?

If you cannot describe the problem in one sentence, the project is not ready yet.

2. Does this job really need AI?

Many problems can be solved with simple automation, a well-designed spreadsheet or a process change. When the rules are clear ("raise an alarm if the temperature passes a set value"), classic automation is both cheaper and more predictable.

AI usually makes sense in situations like these:

• Work where the rules are too complex to write down (for example, spotting defects in images),
• Work that requires reading and summarizing large numbers of documents,
• Forecasting work that depends on finding patterns in past data.

Ask for the simplest solution first. Bring AI to the table only when the simple solution is not enough.

3. Do we have enough clean data?

AI systems are fed by data. In a business where data is scattered, incomplete or wrong, even the most advanced model will produce poor results. The NIST framework also stresses that knowing the quality and origin of data is one of the foundations of a trustworthy system.

Ask yourself:

• Where is the relevant data kept? On the machine, in a notebook, on one person's computer?
• Is it recorded regularly, or only when something goes wrong?
• Who can access it, and does it contain personal data?

Any project that touches personal data needs a plan for compliance with local data protection law from day one. Do not move forward without your legal advisor's opinion.

4. Which number will tell us we succeeded?

Measure where you stand before the project begins. You should be able to say not "it will get better" but "the return rate will fall from this level to that level". If you do not know the starting value, nobody will be able to tell you after three months whether the investment worked.

A good measure:

• Is simple enough for everyone to understand,
• Can be pulled easily from existing systems,
• Connects directly to money or time.

5. What happens when the system gets it wrong?

Every AI system makes mistakes from time to time. What matters is knowing the cost of a mistake in advance. A typo in a draft quote and a wrong decision by a safety sensor do not carry the same level of risk.

The NIST framework recommends mapping, measuring and managing risks. Translated into factory language:

• Map: Which decisions does the system influence?
• Measure: How often is it wrong, and what is lost when it is?
• Manage: Who makes the final call, and what happens when an error is spotted?

Especially at the start, letting a person have the final say on critical decisions is the safest path.

6. Is our team ready for this change?

Even the best system creates no value if nobody uses it. If operators see the new tool as a threat, they will neglect data entry and ignore alerts. The OECD principles likewise emphasize that AI should be human-centred and transparent.

In practice, this means:

• Design the project together with the people who work on the shop floor.
• Show clearly why the system made a given decision.
• Plan short, hands-on training sessions.
• Name an internal owner for the project; do not hand everything to an outside firm.

7. How will we start small and grow?

Trying to transform the whole factory at once is one of the most common mistakes. Instead, start with a single line, a single process or a single department. Set a short trial period, measure the results, then decide.

A good trial plan includes:

1. A clear start and end date,
2. A success measure agreed in advance,
3. A way to roll back if it fails,
4. A plan for which processes it will spread to if it succeeds.

A short checklist

When you evaluate an AI proposal, ask your vendor these seven questions as well:

• Can we describe our problem in one sentence?
• Has a simpler solution been tried?
• Where is the data, how clean is it, and who owns it?
• Which number will we use to measure success?
• What does a mistake cost, and who makes the final call?
• Is the team involved in the process?
• How long will the trial last, and what comes next?

If you can answer these questions clearly, your investment rests on a plan rather than on luck. If you are unsure about the answers, the Atomaxis team would be glad to help you define the first step.