AI integration: Claude, ChatGPT, or Copilot in your tools.

Not with the tool. With the use case. The right starting point is a specific task your team does often, one that involves looking things up or writing, and whose result is easy to check: preparing a meeting, answering a recurring request, summarizing a call, finding a clause.

Three questions are enough to choose. Where is the useful data (CRM, ERP, folders, emails)? Who’s allowed to see it? How will you know the AI saved time? These questions are the core of the AI audit: 10 days to rank use cases by payoff and difficulty, and to check access to the data.

Then you start small, with one use case and one team, and expand what works. A successful integration is measured by daily use, not by the number of licenses.

Many AI proofs of concept never reach production. The causes tend to repeat, and the tool is rarely one of them.

• No connection: the AI sits in a separate window, with no access to the CRM or ERP. People have to copy-paste, and the effort wears them down.
• No specific use: “use AI” isn’t an instruction. Without a defined task, everyone tries it, then gives up.
• No support: teams don’t know what to ask, or how to check an answer.
• No measurement: nobody knows whether the pilot saved time, so nobody champions it.

A well-designed integration handles all four from day one: a written use case, a connection to your software, training on your own tasks, and a before-and-after measurement.

Every new tool means one more account, one more habit, one more window. Adoption collapses within a few weeks. AI that shows up where your team already works (the CRM, email, Microsoft Teams), on the other hand, gets used without effort.

Technically, the integration runs through your software’s APIs (the same work as a standard API integration, with an AI model at the end): the AI reads the customer record, the orders, or the document at the moment it’s asked, then can write the result in the right place. The MCP protocol, adopted by several major AI providers, simplifies these connections and saves you from rebuilding everything if you switch providers.

One connection serves several uses: once the CRM is connected, meeting prep, call notes, and follow-ups all benefit. When the AI needs to go further and carry a task through to the end in your software, that’s an AI agent for business.

There’s no best model in absolute terms, only the right model for a given use. A few pointers:

• Microsoft Copilot is the natural choice when your company runs on Microsoft 365 (Outlook, Teams, SharePoint).
• ChatGPT and Claude are very strong at writing, analyzing long documents, and multistep tasks.
• Mistral AI, a French company, meets data sovereignty requirements or a preference for a European provider.
• Gemini fits naturally into Google Workspace.

I don’t resell any subscriptions: I advise you based on the use, the cost, and your data requirements. Accounts are opened in your name, and the integration is designed so you can switch models without starting over. Several models can also work side by side, each on the task it does best.

Connecting AI to your software means giving it access to sensitive data. Four rules govern every integration.

• Business plans: on the APIs and enterprise plans from OpenAI and Anthropic, your data isn’t used to train the models by default. That’s not always the case with consumer plans or personal accounts.
• Each person’s access rights: the AI only sees what the user can already see in the CRM or ERP.
• As little data as possible: only the fields that matter are sent to the model, and access keys stay in a vault.
• GDPR compliance: a data processing agreement with the provider, an up-to-date record of processing activities, and an EU region when the provider offers one.

A log keeps track of every request: you know who asked what, and on which data.

In France, public programs help small businesses. Bpifrance, for example, offers the Diag Data IA, partly state-funded and run by accredited experts. I’m not an accredited expert, but I can build the tools your assessment recommends. Depending on your industry, your OPCO (a French training fund) may cover part of your team’s training.

Training matters as much as the connection: it’s the heart of change management. It’s built around your team’s real tasks: what to ask, how to check an answer, when not to use AI.

The split of roles is simple. You choose the use cases, grant access, and name a point person. I design, connect, secure, train, and measure, as an AI integrator who doesn’t resell any licenses.