Blog/AI
AI

AI in the enterprise: why your teams use it badly

The problem is almost never the tool or the model. It is that nobody has set the framework of use. What it costs, and what to decide before deploying.

Premaccess30 July 20264 min read

We run artificial intelligence training courses in companies, for executive committees as well as for technical teams. The finding is the same everywhere: the subscription is paid for, the tool is installed, and the real gain is well below what was expected. Almost never because of the tool.

What we observe when we arrive

  • A single mode of use. Assistants offer several ways of working, with very different levels of risk. Most users know only one of them, and it is rarely the right one for a subject that involves money or a business system.
  • No trace of the reasoning. The result arrives quickly, it looks correct, and nobody knows by which path it was produced. On a transfer order or a consolidation, that path is precisely what you need to be able to read back.
  • Zero capitalisation. A case that worked in March is redone from scratch in April. Time saved once is never saved twice.
  • The skill resting on a single person. Whoever got started alone becomes one more single point of failure.

What it actually costs

The figures that participants themselves put on the table at the start of a session are systematically higher than management imagined: three days a month of re-entering data at a manufacturer, around an hour a day of bank reconciliation at a group, several hours a week reformalising decisions taken verbally. Those hours already exist. They are simply not counted anywhere.

The four questions to settle before deploying

These are the only ones that really block a general roll-out, and they are not technical:

  1. Which data never passes through an external tool.
  2. Who is entitled to what, with which access levels.
  3. Who validates before an action leaves the company.
  4. Who bears the responsibility in the event of an error.

As long as they are not settled, a broad deployment is a gamble. Once they are, the rest is execution.

The principle that never moves

The assistant prepares, the human validates, nothing goes out without review. At the start: no automatic sending of email, no direct writing into a business tool. This single rule avoids most of the incidents we see elsewhere.

We set this framework and work on your own cases over three days on site, then two days of follow-up. See the AI training and the executive committee programme.

Going further

Does this concern you? Have a look at our AI and Data services, or talk to an expert (reply within 24 business hours).

Frequently asked questions
Why AI training rather than a simple subscription?

Because the subscription says neither what to entrust to the tool, nor how to review what it produces, nor what must never leave the company. It is the framework of use that produces the gain, not the licence.

Which questions have to be settled before generalising?

Four: which data never leaves, who is entitled to what, who validates before an action leaves the company, and who bears the responsibility in the event of an error. They are organisational, not technical.

How do you measure the gain?

By first quantifying the existing workload, role by role, before showing anything at all. Without that starting measurement, no gain is demonstrable afterwards.

P
PremaccessCloud experts · Franco-Swiss since 2007
/ Also worth reading