Questions and answers
Each answer on this page can be read on its own: it starts with the fact, it fits in a few lines and it contains a figure or a rule that can be checked. It is what we say in meetings, written down as is. The answers are written by the Optim-AI team.
Part 01
The definitions a business leader needs to make decisions, without depending on an expert at every meeting.
Generative artificial intelligence is a statistical machine trained to produce the most probable continuation of a text, an image or a line of reasoning. It does not consult a knowledge base and checks nothing: it calculates, from patterns observed during its training, what should come next. This mechanism explains both why it writes a good sales letter and why it is poor at simple arithmetic. It also explains why you need to give it your own documents if you want it to talk about your company rather than an average company.
No. A language model does not understand the meaning of what it produces in the way a human does. It strings together statistical predictions that are good enough to be useful, without intention, without awareness of saying something true or false, and without the ability to signal that it does not know something. It is precisely this absence of doubt that makes human checking mandatory on any output for which the company is accountable.
Because it is designed to produce a plausible answer, never to say that it does not know. Information missing from its training is filled in with whatever statistically resembles a good answer, and it is presented with the same confidence as an accurate fact. This phenomenon, called hallucination, is not a flaw that can be fixed: it is a property of the process, which is managed by connecting the model to your documents and keeping human review on high-stakes outputs.
Conventional automation is deterministic: the same input always produces the same output, and the behaviour can be audited line by line. Generative AI is probabilistic: the same input can produce two different results. This difference determines the use. The probabilistic approach suits loosely defined tasks such as summarising, rephrasing or classifying. It does not suit tasks that require systematic accuracy, such as a payroll calculation or a regulatory filing, where conventional automation remains the right tool.
Eight are enough: language model, generative AI, hallucination, prompt, retrieval-augmented generation, fine-tuning, agent and token. They cover everything a business leader needs to understand to make decisions on a project and judge a sales proposal. Optim-AI teaches them in a dedicated AI awareness engagement, using the company's real cases. A provider who uses more jargon is generally selling complexity rather than competence.
Part 02
What can really go wrong, and what is merely an ill-informed fear.
There are five: data leaking to uncontrolled tools, dependency on a supplier you can no longer leave, invisible errors that travel all the way to the customer, the erosion of in-house skills that are delegated, and a cost that exceeds the gain. Optim-AI deals with them in this order, from the most likely to the most costly. None of them is a reason to do nothing, and none is solved by banning the tool: a ban produces hidden use, and therefore uncontrolled use.
That depends entirely on the contract you have signed. Consumer offerings often provide for conversations to be reused to improve the service, whereas business offerings and access through an application programming interface (API) generally exclude it. It is a clause to check supplier by supplier before any deployment, then to check again whenever the terms of use change. In an Optim-AI audit, this contractual check is systematic and appears in the report submitted to management.
The information leaves the company, with no malicious act, no trace and without anyone having decided it. It is the first risk found in almost every audit, because the action is commonplace: summarising a contract, rephrasing meeting minutes, reviewing code. The effective response is not sanctions but a no-blame inventory, a short and readable usage rule, and above all an internal tool at least as convenient as the unofficial one. Without this last condition, the rule will be circumvented.
Yes. Capable models run on European servers, and some run on your own infrastructure. The choice depends on how sensitive the data concerned really is: not all data requires the same level of protection, and hosting everything in-house is unnecessarily expensive. Optim-AI draws up a target architecture that explicitly separates what stays within your information system from what can be entrusted to a third party, with the reason for each decision set down in black and white.
Up to date as of 10 October 2026. The EU AI Act (Regulation (EU) 2024/1689, amended in July 2026) classifies uses by level of risk. An SME that uses AI tools currently has a general obligation: to take measures so that its teams understand these tools, with no level to guarantee and no certificate to obtain. Certain practices have been prohibited since February 2025, and since August 2026 manipulated content (“deepfakes”) that you distribute must be disclosed. Most office uses entail no other obligation. High-risk uses, such as screening job applications, access to credit or employee monitoring, will be subject to stricter obligations from 2 December 2027 and deserve a dedicated review before any deployment.
Yes, whenever the tool processes their data or influences their work. In France, in companies with at least 50 employees, introducing a system that affects working conditions requires informing and consulting the works council (comité social et économique, CSE). In all cases, no information about an employee may be collected by a system they have not been told about beforehand. Beyond the obligation, transparency is what decides adoption: a tool deployed in silence is circumvented, while a tool that is explained and governed is used. Optim-AI treats this point as a step in the engagement, not as an end-of-project formality.
Part 03
The figures a business leader should have in mind before opening a budget line.
The ranges observed on the French market run from €2,000 to €8,000 for an SME and from €10,000 to €30,000 for a mid-sized company or a multi-site group. The most reliable public benchmark is Bpifrance's Diag Data IA: eight days of expert work, with the quote capped at €10,000 excluding VAT (Bpifrance conditions, October 2026). Optim-AI prices its engagements as a fixed fee, never on a time-and-materials basis, and gives the amount in writing before any commitment.
Yes, subject to conditions. Bpifrance's Diag Data IA funds 40% of an eight-day diagnostic whose quote is capped at €10,000 excluding VAT, leaving at most €6,000 to pay. It is open to SMEs and mid-sized companies with 10 to 2,000 full-time equivalents, annual revenue of at least €1 million and more than one year in existence (Bpifrance conditions, October 2026). The diagnostic must be carried out by an expert approved by Bpifrance. Optim-AI is not approved at this time and therefore promises no funding.
The subscription is the visible part and rarely the heaviest. The real cost includes integration with the information system, data migration and cleaning, team training, maintenance at every change of model version, and above all the human time spent checking outputs. This last item is the one most often left out of estimates, and it is the one that turns a project that is profitable on paper into one that only breaks even in reality.
By choosing the indicator before launching, and taking it from those you already track: processing time, number of cases handled per person, rework rate, cost per request. An indicator created for the occasion proves nothing and convinces nobody in committee. The calculation must include the time spent on human checking. Optim-AI also sets a stopping criterion before the start: without one, a disappointing project drags on through inertia instead of being stopped.
Part 04
What shifts within the company, beyond the tools.
AI replaces tasks, not entire jobs. It first takes over work that is repetitive, structured and involves little decision-making: data entry, sorting, document search, first drafts. According to the International Labour Organization (2025), one job in four worldwide involves tasks exposed to generative AI, and these jobs will more often be transformed than eliminated. The greater the share of these tasks in a role, the more it will change, and the more it needs support. A role that carries responsibility, arbitrates between conflicting interests or maintains a customer relationship does not disappear: it shifts towards checking and decision-making.
No, and putting the question that way leads to the wrong decision. A business leader is not replaced by a machine: they are overtaken by a competitor who structured their data and processes while they waited for things to settle down. The race is won or lost on organisation, not on technology. So the useful question is not “will I be replaced” but “what share of my company's work can be automated, and what do I do with the time freed up”.
By deciding explicitly, and in writing, which skills will never be delegated to a machine, then keeping real practice of those skills even when the tool could do the work. A skill that is no longer exercised erodes quickly: a synthesis of 457 studies published in 2025 estimates that half of the level acquired is lost within six months to a year without practice. The day the machine needs correcting, nobody can judge whether the output is good. Skills development must focus on checking and decision-making, not on handling the tool, which can be relearned in an hour.
Four, observed repeatedly. Roles shift towards checking, because producing a first draft becomes cheap and judging its quality becomes the scarce skill. Traceability becomes a daily concern, since you need to know who approved what. Gaps widen between teams, which becomes a management problem before it becomes a tool problem. And recruitment criteria change: framing a problem counts for more than executing it quickly.
Part 05
The decisions that determine whether you will control your working tools in three years' time.
By not building on the model. Your data, your processes and your teams' skills have a lifespan of several years; today's model lasts a few months. A sound architecture isolates the model's entry point behind a replaceable part, which makes it possible to change model in a day instead of rebuilding the tool. What proves expensive is never having chosen the wrong model, but having built an architecture that prevents you from changing it.
By a single test: can you change supplier within a few days without stopping production? If the answer is no, you do not have a tool, you have a dependency, and nothing else — hosting, contracts, certifications — changes that. Optim-AI runs this switchover test before the end of every engagement, rather than writing a reversibility clause into a contract that nobody will check.
Four things. Your raw and cleaned data, the only asset that gains value over time, which must never be stored solely with a third party. Your business rules, written down and versioned, because if they exist only in the instructions given to a model, they disappear with it. Your supplier accounts, in your name and with your means of payment, never in the provider's name. And at least one person able to judge whether an output is good, without asking the provider.
With four rules that fit on one page. A named owner, one person and not a committee, who does not need to be technical but must be able to say no. Four indicators, not forty: time saved on a measured task, error rate observed, real monthly cost, number of uses stopped. A quarterly review where uses are approved, extended or stopped. And a register of uses, without which none of the other three rules can be applied.
With a high-volume, low-stakes task that is easy to measure, usually in administration or support. It produces a visible result within a few months without a heavy IT project, and lets the company learn where a mistake is not costly. Starting with a large cross-functional programme rarely produces a result within twelve months and uses up the executive committee's patience before anything has been proven.
By checking three points before signing. Does the provider receive a commission on the tools it recommends? If so, its recommendation does not serve you. Will the code, data and supplier accounts be in your name from day one? And will the documentation allow a third party to take over without the provider? A refusal on any of the three indicates that you are being sold a dependency rather than a tool. Optim-AI writes these three commitments into the contract, not just on its website.
Who answers
They come from engagements and executive committees, not from press monitoring.
Optim-AI is a French artificial intelligence consulting and development firm for leaders of SMEs and mid-sized companies, working throughout mainland France. The answers on this page are written by the firm's team. They draw on the backgrounds of its experts: leadership of engineering and purchasing in the automotive industry, data science and AI product development, AI audit and governance. See the team.
Next step
Send it to us. If other business leaders ask it too, it will be added to this page.
