.svg)
100 million users in three months. That’s ChatGPT’s tally at launch. It took mobile 16 years to reach that figure. The internet, 7 years. Facebook, 4 and a half years. No company had anticipated this kind of acceleration. And in the meantime, your employees are using it—with their personal accounts. On data that belongs to you. Without a framework, without governance, and without you really knowing it.
The question is no longer whether you will adopt AI. It’s already here. The question is whether you will be at its mercy or take the reins.
This article is based on a webinar hosted by Samir Amara. Prefer to watch the video? Watch the full session below (30 minutes).
Why AI Caught Everyone Off Guard
An unprecedented adoption
Previous waves of technology gave people time to get organized. It took nearly a decade for the cloud to gain significant traction in businesses. SaaS took about the same amount of time. Now, we’re talking about a matter of months. Employees figured it out on their own. They tried it, saved time, and kept using it. No one asked for their input, and they didn’t ask for yours either.
This scenario is reminiscent of the arrival of the first personal versions of OneDrive and Google Drive fifteen years ago. Employees were uploading company files to share them more easily, without IT oversight. Except this time, the scale is different. So is the speed. And above all, the nature of what’s happening: it’s not just data leaving the company—it’s know-how being injected into tools that don’t belong to you.
The budget reflex of leaders
According to an IDC study published in 2024, 45% of organizations are making AI a budget priority in 2025. Two years ago, it was cybersecurity. Five years ago, it was collaboration. The shift has begun, and it is massive. A Capgemini study from the same period indicates that 71% of French executives plan to integrate AI agents into their organizations in the coming months.
Watch out for this pitfall: simply signing up for a business subscription isn’t enough. Many people still think it is. It’s the same mindset as “I bought Microsoft 365, so I’m in the cloud.” Owning the tool isn’t the same as deploying it. Deploying the tool isn’t the same as managing it.
The Myth of the In-House LLM
The Dream Fades Away
Two or three years ago, a lot of companies wanted their own AI—models trained on their own data. Law firms, accounting firms, and mid-sized industrial companies. At IT Systèmes, we supported several of these projects. We messed up. Honestly.
Gartner predicts that 70% of projects to develop proprietary AI will be abandoned in the coming years. Why? Because it’s too cumbersome, too expensive, and too technical. OpenAI, Anthropic, and Mistral take years and require substantial resources to achieve acceptable results. An SME or mid-sized company that tries to do the same with a team of three people is headed for disaster.
The AI agent: the real gateway
An AI agent is different. You take an existing AI and configure it to meet a specific need within your company. You give it a role, data sources, and a scope. You don’t reinvent the engine. You build the car around it.
And that’s where it gets interesting, because agents aren’t tied to a single provider. You can build an agent based on GPT, Claude, Gemini, Mistral, or several of them at once. The most powerful AI engine today won’t be the most powerful in six months. Your agents, however, will remain. It’s your intellectual property that builds on top of the engine, not the engine itself.
The Three Stages of AI Maturity in the Enterprise
Phase 1: Humans and Their Assistants
That’s how everyone knows it works. You open ChatGPT, ask a question, and the AI responds. You ask it to draft an email, summarize a text, or help you structure a presentation. It’s useful and saves time, but it has its limits. The AI does whatever it wants, has no stable memory, can hallucinate, and you have no control over where your data goes.
At this stage, the value remains individual. Every employee has their own little secret, their favorite shortcut, their own trick. The company, however, doesn't capitalize on any of it.
Phase 2: The Human-Controlled Agent
That’s where things change. You configure an AI for a specific task. You feed it your documentation, your processes, your collective bargaining agreement if it’s an HR agent, and your catalogs if it’s a sales agent. You define its scope of action. The agent won’t search just anywhere on the internet; it operates within the parameters you’ve set for it.
This is the phase where most companies would be wise to position themselves by 2026. The value is immediate, the risk is manageable, and you begin to build equity that belongs to your company. The HR department drafting a new contract calls the agent. The agent is familiar with the collective bargaining agreement, company-specific details, and approved templates. The agent makes a proposal, and the human approves it. The sales department preparing a response to a request for proposals calls another agent, who is specialized for that task.
Phase 3: The Autonomous Agent
You set the strategy; the AI executes it. Hiring someone? The agent handles the pre-employment background check, prepares the contract, initiates onboarding, verifies the documents received, and follows up if they’ve expired. And it adapts in real time: if a document doesn’t meet the requirements, it rejects it and requests a new one. No more rigid workflows. Operational intelligence that runs continuously.
This phase is powerful but demanding. It requires your company to have a thorough understanding of its processes, legal framework, and documentation. If any one of these three areas is unclear, the autonomous agent will only amplify the confusion. At IT Systemes, we start by helping clients with simple tasks: scheduling appointments, qualifying leads, and administrative follow-up. As for the rest, most companies aren’t ready. And it’s better to acknowledge that than to jump in unprepared.
Your AI will never be better than you
The mirror, not the savior
Many people hope that AI will save poorly organized companies. Wrong. AI simply reflects your level of organization—period. If your processes are clear, your agents will be clear. If your processes are confusing, redundant, and poorly documented, your agents will be confusing, redundant, and poorly documented.
It's not a bug; it's a feature. The AI works with what you give it. It doesn't invent discipline where there isn't any. It doesn't invent expertise where it's lacking. It amplifies what already exists—for better or for worse.
The new value of your business
In the medium term, what will drive your company’s value will no longer be just your revenue or your customer base. It will be your ability to deploy AI agents that fully understand your business DNA. A company without well-designed AI agents will, in five years, resemble a company without an information system today. It may still function, but it will be significantly less valuable when it comes time to sell.
This issue of knowledge transfer is still underestimated. The baby boomer generation is set to leave companies without their expertise. AI could become the repository for expertise that would otherwise be lost to retirement—provided we start feeding it that knowledge now.
The Invisible Trap: Your Employees' Personal AI
The same old story
Take Marc, your sales rep. The company hasn’t rolled out anything. He signs up for a personal ChatGPT subscription at 23 euros a month. He uses it every day. He trains it on your product offerings, your sales pitches, your common objections, and your pricing. He refines his prompts for months. He becomes incredibly effective.
And then Marc leaves. To a competitor, or somewhere else. He takes his data with him—that’s nothing new; it happened even before AI. But he also takes his history of prompts, his accumulated learning, and the optimization built up over months. Everything he’s fed into his personal AI stays in his personal account. The company has to start from scratch.
An invisible debt that keeps piling up
This phenomenon is already evident. We have seen companies lose their project leaders, who take months’ worth of agent configuration work with them when they leave. The loss isn’t immediately apparent, because nothing disappears overnight. But it takes the replacement months to restore the same level of performance.
The longer you wait to establish guidelines for AI use in your company, the more this debt grows. And this isn’t a debt you can pay off—it’s a debt you have to absorb. The expertise embedded in personal accounts is lost forever.
Choosing Enterprise AI: The Key Criteria
The market is flooded with options. Microsoft is promoting Copilot. Google is promoting Gemini Workspace. Anthropic offers Claude for Enterprise. OpenAI has ChatGPT Enterprise. Mistral offers European sovereign solutions. How can you make the right choice?
Data sovereignty
This is the number one criterion—and often the most misunderstood. When you use consumer-grade AI, your data is sent to servers located overseas, most often in the United States. For anonymized or trivial data, it doesn’t matter. But for sensitive customer data, HR data, or strategic information, it’s a problem.
Here are some questions to ask your supplier:
- Where is the data hosted?
- Under which law is the contract signed?
- Does the data train the public model, or does it remain isolated?
- Is there a European processing zone, and if so, which law governs it?
On this last point, be wary of empty gestures. A "buffer zone in Europe" governed by U.S. law remains subject to the same extraterritorial regulations. If GDPR compliance is a real concern for you, the origin of the contract matters just as much as the location of the servers.
Integration with existing rights
If you deploy AI in an environment where document access rights are poorly defined, the AI will expose everything it can. An employee might request the CEO’s pay stub, and if nothing technically prevents access, the AI will pull it up. This is what’s known as latent exposure: the vulnerabilities already existed; the AI simply makes them visible on a large scale.
Solutions integrated into an existing office environment have the advantage of reusing existing permissions. This is a particular strength of Microsoft Copilot, which leverages the security policies already in place in Microsoft 365 (Purview, Entra, Defender). Standalone solutions require everything to be reconfigured in parallel. This is no minor detail; it is often what determines the success or failure of a deployment.
The ability to create agents
Enterprise AI that cannot create custom agents remains a Phase 1 tool. You’re paying more for the same chatbot. Make sure the tool offers a studio or agent creation environment that allows you to integrate your documentation, connect to your business databases, and apply access policies.
At IT Systèmes, we primarily work within the Microsoft ecosystem because it currently offers the most robust combination of agent creation (via Copilot Studio), native governance (via Purview), and integration with the tools our clients already use. But we also support companies that have made different choices, and we remain platform-agnostic when the context warrants it.
The ecosystem, not the engine
An AI engine that’s top-notch today won’t be in six months. Claude was ahead a year ago. GPT has regained the lead in certain applications. Gemini is making rapid progress. Mistral is surprising us in other areas. If you tie yourself to a provider solely because of its engine, you’ll have a problem in twelve months.
Instead, base your choice on the ecosystem: governance, security, integration with the tools you already use, and the ability to connect multiple engines depending on your needs. The best deployments today are hybrid: one engine for sensitive internal agents, another for creativity, and another for code.
How to Deploy Without Messing Up
Start with a governance audit
Before rolling out any new system, take a look at your current setup. How are your document access permissions structured? Are your sensitive documents labeled? Are your processes documented? Do your teams already use AI, and if so, which ones?
This audit takes two to four weeks, depending on the size of the organization. It often reveals surprises: personal Google Drive accounts filled with customer data, haphazard naming conventions, and permissions granted six years ago that have never been reviewed. All of this needs to be addressed before deploying AI on a large scale. Otherwise, the AI will just make the mess worse.
Start with business pain points
The most important thing—and this is where most projects go wrong—is to start with the business. Not the tool. Not the technology. The business.
What specific frustrations do your teams face on a daily basis? Which repetitive tasks take up a disproportionate amount of their time? Which errors occur regularly? Which documents have to be redone over and over again? This is where frontline staff need to weigh in—not on theoretical use cases selected by the IT department simply because they’re technically interesting.
A well-chosen POC that addresses a real pain point can drive adoption. A poorly chosen POC—one that’s technically brilliant but useless in day-to-day operations—can undermine the project’s credibility for two years.
Measure, adjust, enlarge
Once the first agent is up and running, measure the results. How much time is saved per user per week? What is the actual usage rate after three months? What kind of qualitative feedback have you received? The numbers are important, but informal feedback is just as important. If users are spontaneously talking about it with their colleagues, you’ve got a hit on your hands.
Then expand to other services and use cases. Gradually build out your agent catalog. Document their configurations. Train your teams to use and improve them. This is a long-term effort, not a sprint.
What nobody tells you about the real cost
The visible cost
A Copilot, Gemini Workspace, or ChatGPT Enterprise license costs around 30 to 60 euros per user per month, depending on the provider. That’s the visible cost—the one you include in the business case.
The hidden cost
The real cost lies in the support. This includes governance audits, ensuring compliance with legal requirements, designing initial agent configurations, training teams, and measuring and adjusting performance. Expect between three and six months of internal or external work for a full-scale deployment at a mid-sized company. Without this support, licenses cost three times as much, yet usage rarely exceeds 20%.
Many companies underestimate this cost because it doesn’t look like a traditional IT investment. It’s not servers; it’s not software. It’s human time, consulting, and organizational change. That’s precisely why it works when done right—and why it fails spectacularly when people try to skip this step.
The right time to start
If you haven't launched anything yet, you're not behind—but you're starting to fall behind. Your competitors who started a year ago have already built up a track record, developed effective tools, and assembled teams that know how to use them. That gap isn't going to close on its own.
You don’t need a five-year plan. You need a decision for the next six months: an audit, one or two business POCs, a governance framework, and a primary vendor chosen with full knowledge of the facts. The rest will fall into place as you go, because the market is changing too quickly to set a long-term strategy in stone.
One thing is certain: your employees aren’t waiting around. They’re already using AI. The only question is whether that AI belongs to your company or to their personal accounts.
Want to take it a step further?
At IT Systèmes, we support companies in their AI deployment, starting with their initial Copilot projects. From governance audits and the design of business agents to compliance and team training, we provide support throughout the entire process.
Schedule a 30-minute consultation with our team to review your situation and identify your top priority use cases.
Our conviction, turned into a tool
Everything you’ve just read is something we experience every day with our clients. And through our work supporting businesses on the ground, we’ve identified a gap: conversational AI solutions can respond, but they can’t take action. Yet it is action that creates value.
To address this gap, we developed FlexFlow, our business process orchestration and automation platform. Hosted in France and compliant with GDPR and ISO 27001, it connects your existing tools, automates your workflows, and enables your AI agents to turn conversations into action.
If data sovereignty and control over your processes are important to you, FlexFlow is worth checking out.
About the author
Samir Amara is the president of IT Systèmes. For more than 15 years, he has been helping companies navigate their successive IT transformations: digitalization, cloud computing, cybersecurity, and now artificial intelligence. His belief is that AI is not just a subscription service—it requires governance.
.svg)