When you don't need AI.

Almost every week, a company asks us for "AI." When we ask why, the answer, in more than a third of cases, is vague. "Everyone is using it." "The competitor announced a project." "The board asked for a plan." No specific pain, no data, no definition of success. Just enthusiasm.

We turn down about a third of these requests. Not because we do not need the money, but because a project that begins in enthusiasm usually ends in disappointment. That disappointment poisons our reputation, and poisons AI's reputation in the market as a whole. This piece is about the cases in which we say: no.

One: you have a problem, not data.

AI without data is magic without a trick. Before you ask "which model?", ask: where is my data, is it clean, is it accessible? Most projects that fail do not fail at the model. They fail in the first weeks, when the team discovers the data is scattered across spreadsheets, locked in a legacy system, or does not really exist.

If you have not started collecting and organizing your data, the first invoice you should pay is not to an AI company. It is for a data-infrastructure project. That work is less glamorous, but it is the foundation.

Two: the problem is deterministic, not probabilistic.

A model answers in probabilities. Your accounting cannot tolerate probabilities. If you want to sum numbers, apply a clear rule, or compute a tax — a spreadsheet formula is faster, cheaper, and auditable. Use AI where there is real ambiguity: language, images, long text, behavior. Do not use it where certainty is required.

AI is not an answer to every question. It is a good answer to very specific ones.

Three: volume does not justify the cost.

If you receive fifty customer messages a month, a good employee is cheaper than a bot. AI has a floor cost that does not disappear: training, maintenance, monitoring, retraining. That cost is justified by volume, not by the idea. We tell clients plainly: if your interaction volume is not in the hundreds per week, stay with humans. You will be happier, and your customer will be happier.

Four: trust matters more than speed.

In contracts, medical advice, legal opinions — a slow human beats a fast machine. The customer does not want an answer in a second. They want someone who takes responsibility. AI can assist that person — classify documents, summarize, suggest — but it cannot replace them. Whoever tells you otherwise is selling you a fantasy.

Five: your problem is organizational, not technical.

Sometimes a client comes to us complaining that their team is slow to respond to customers. They ask for a bot. We sit down, ask questions, and find that the problem is not speed, but that the team receives messages on three different channels, with no system, and no clear owner.

A bot here will not fix anything. It will only add a new layer on top of an existing mess. The fix: repair the process first. Then, if you still have a volume or language problem, call us. Do not lay technology over a broken process and then blame the technology.

When do you actually need AI?

When the conditions above do not apply and a real need remains, the project becomes a serious candidate. In our experience, the cases that pay off always share:

• Large volumes of language or text (customer service, search, summarization, classification).
• Patterns in large, unstructured data no human can read end-to-end.
• Repetitive, boring work with a reasonable error margin.
• Personalization per user at a scale manual tailoring cannot reach.
• Data that exists, is relatively clean, and is legally accessible.

Closing.

The easiest sale in tech today is selling "AI" to a company that does not need it. Demand is high, scrutiny is low, budgets are open. But we would rather lose a deal than sell a solution that does not fit.

The client we tell today, "you don't need AI for this," is the client who comes back a year later, when they actually do. Honesty is not just ethics. It is a longer-lived business model than hype.