Bubble or turning point?
What this week says about the future of AI.

A word from Jérôme

JDENIS Consulting · Toulouse

My perspective

I do not know whether we are living through a bubble. I do see several classic mechanisms at work: valuations based on future revenue, suppliers financing their customers, energy and building commitments over fifteen or twenty years. A correction would not mean the technology is useless. The internet survived the year 2000; many companies did not.

For an organisation, the lesson is simple. The tools you use belong to players whose financial situation can change quickly: prices, terms of use, acquisition, closure. That is no reason to wait. It is a reason to keep control of your data, your instructions and your teams' skills.

My perspective

Two stories collide. For some, AI is progressing so fast that kill switches and perhaps a freeze are needed. For others, its promises have been paid for too dearly and markets are correcting. Both can be true at once: a powerful technology can be badly financed.

What strikes me most is that the important decisions are taken far from the teams that use these tools. It is therefore up to organisations to set their own rules: what they delegate, what they check, what they keep in-house. Waiting for the stock market or lawmakers to decide would be a mistake of timing.

My perspective

The study on analysts seems to me the most important of the week for managers. It does not show that AI gets things wrong. It shows that competent professionals adopt its position without realising it, then defend it in its words. In a team, this can lead to uniform, poorly understood decisions.

I would suggest a simple rule for decisions that matter: write down your view before consulting the agent, then explain in writing why you change it. For newcomers, ask "why" for every suggestion. These are modest habits, but they protect what remains the heart of the job: being able to justify a decision.

My perspective

None of these ideas requires technical skills. They require a little discipline and, above all, a named owner. In a small organisation, dependence on a tool builds up without any explicit decision: a trial becomes a habit, then a condition for operating.

I would start with the first and second: export and write down. They protect against a shutdown, a price change and the departure of a colleague all at once. They are also the easiest to forget.

My perspective

We have talked a lot about the agent that helps the employee. Now we need to think about the agent that turns up at reception, on the phone or on the website, on behalf of a customer. United chose to close the door. Others will choose to open it, with rules.

For a small business, I would ask three simple questions. Do we accept an order placed by an agent, and with what confirmation? Who answers when a machine calls? How do we spot complaints produced in bulk? Better to answer them before the first incident.

My perspective

The fall in prices is real, but it says nothing about prices a year from now. Companies that go public or raise funds on tougher terms will one day have to make money on every use. Today's introductory prices are no guarantee.

I therefore advise measuring cost per completed task, not per subscription, and keeping at least one tested alternative. Those who can compare do not suffer price changes; they arbitrate them.

My perspective

AI is often described as an intangible service. This week is a reminder that it depends on power plants, electricity grids and local permits. Residents who vote against a data centre now weigh on chipmakers' market value.

The most solid industrial applications are also the most modest in their promises: a copilot that suggests, an engineer who checks. That is the model I recommend to companies starting out.

My perspective

A tool's reliability is not just the quality of its answers. It also depends on the provider's soundness, what it does with your data and how errors are covered. Fireworks and Cowbell tell the same story from two angles.

mAIjin's small test echoes the study on analysts: the tool does not contradict you of its own accord. I recommend systematically adding a request for criticism to work instructions. It costs nothing and changes the relationship.

My perspective

Rules are moving forward, but at a pace that has nothing to do with that of the tools. Between a New York law that will apply in 2027, a delayed AI Act and a voluntary call signed in Paris, a European company will not find a complete framework for several months.

I see this as one more reason to adopt simple internal rules: which tools are authorised, what data can go into them, who approves sensitive uses such as recruitment or appraisal. These rules will be useful anyway on the day the law applies.

My perspective

Creative work shows on a small scale what awaits many professions. The big players choose their tools based on legal risk as much as quality. Independents sometimes find in them a way to exist outside the usual channels.

For a creative professional, I would keep two habits: knowing what the tool was trained on, and keeping a record of your own contribution. These are the elements that will count in future contracts and disputes.