Frequently asked questions about AI in companies

This post gathers real questions I get in consulting work and talks. No jargon, no promise of digital transformation in 30 days. If you lead a team or a company, you have probably asked at least one of them.

Will AI replace my team?

Not in the short term, and almost never the way the fear suggests. What changes is the mix of tasks inside each role. The repetitive, predictable parts tend to be absorbed. The parts that require judgment, context and relationships gain weight. In practice, the person who uses AI well produces more and spends less energy on busywork. The real risk is not the machine replacing your team. It is a competitor with the same team using better tools.

Do I need to train a model from scratch?

For the overwhelming majority of business cases, no. You combine ready-made models with your data and your process. Training from scratch is expensive, slow and rarely justified outside companies with a very specific need and a lot of volume. Start with what already exists. Only consider building your own once you have proven the value and hit a concrete limit.

Where do I start if my data is not organized?

Pick a critical, repetitive process. Document the decisions it requires today. Run a small, measurable pilot. Perfect data is an excuse to never start. Sufficient data is already enough to learn. You organize the rest along the way, with a clear goal pulling the cleanup, instead of organizing everything first and never reaching the part that creates value.

How much does an AI project cost?

It depends less on the technology and more on three things: scope, integration with the systems you already have, and how much risk you want to tolerate in the first cycle. Serious pilots fit modest budgets when the problem is well defined. What blows up budgets is usually not the model. It is the lack of clarity at the start, which turns into rework in the middle.

How do I know it is working?

Define the metric before turning on the model. Time, unit cost, error rate, customer satisfaction. Pick one main metric, record the baseline and track it weekly. Without a baseline, any result looks good and any vendor looks competent. With a baseline, the conversation stops being about impressions and becomes about numbers.

What about the security of my data?

It is a legitimate question and should be asked early. Define what can and cannot leave the company, prefer vendors that do not use your data to train their own models, and treat sensitive data with the same care you already apply in any other system. Security is not a reason to avoid starting. It is a reason to start with the right case and clear rules.

What is the biggest mistake when starting out?

Starting from the tool instead of the problem. The question is never "where do I use AI". The question is "which decision in my operation is expensive, slow or inconsistent, and could get better". When you start there, the technology becomes an execution detail, not the center of the project.