Written on 6 November 2026. Pricing and versions: official pages read that day.
My editorial line
I have used AI since 2022 and started with ChatGPT, in November 2022, as soon as the conversational version came out. I use it to correct texts, quickly check information and write code.
Rules for this guide:
- I say what I tested and what I did not. An opinion without a trial is presented as such.
- Prices and versions are dated and come from official pages when they can be read. When I could not read them, I say so.
- An AI that gets it wrong is not a scandal, it is the starting point. Each article says where the tool tends to make things up.
- No partnership, no affiliate link.
The comparison in one table
| Assistant | Where it is comfortable | Where I stay wary | Article |
|---|---|---|---|
| ChatGPT | General use, many features (files, images, voice, tasks) | Plans and prices that move often, ads on Go | ChatGPT |
| Claude | Long texts, code, documents to rewrite | Usage limits that vary by plan | Claude |
| Gemini | Google ecosystem (Workspace, Drive, Gmail), large context | Many plans and tiers, names that change | Gemini |
| Microsoft Copilot | Work in Word, Excel, Outlook, Teams with the tenant’s data | Paid licence on top, results tied to existing rights | Copilot |
| Mistral (Le Chat, now Vibe) | French provider, general use, development | Product name in transition | Mistral |
| DeepSeek | Models available through an API at low cost | Data processed in China: rule out for work data | DeepSeek |
| Perplexity | Search with cited sources | Check every source; the answer is not the source | Perplexity |
This table is my reading at the time of writing. Day to day, I use Claude, Gemini and ChatGPT.
What I ask AI to do in an IT job
- Write or fix a script (PowerShell, Bash, Python): reliable for a skeleton, never for running without review. Always test with
-WhatIfor on a test account. - Point a troubleshooting session in a direction: an error message, a possible cause, the diagnostic commands to run. The AI suggests, the machine answers.
- Document: turn notes into a procedure, rephrase a reply to a client, proofread a report.
- Read logs: summarize an excerpt, spot a pattern. Provided the excerpt has been anonymized.
What you never paste into an assistant
My rule, on every assistant: never a password, bank details or IBAN, and never a first or last name. More broadly, unless the service is covered by a contract that governs data processing, do not paste:
- passwords, API keys, tokens, sign-in secrets, BitLocker keys;
- personal data of clients or colleagues (names, emails, numbers, IP addresses tied to a person);
- message contents, contracts, confidential documents;
- internal network diagrams, addressing plans, full firewall rule sets;
- raw, non-anonymized logs.
Anonymize: replace names with client-A, addresses with 192.0.2.10, domains with example.com. For professional use, prefer a provider’s business plan, which spells out how data is used, and read its terms. The French data protection authority (CNIL) publishes recommendations on AI: cnil.fr/fr/intelligence-artificielle (in French).
What to do
- Pick one main tool, and a second to check it. More than that is scattering.
- Write a one-page rule for the team: what is allowed, what is forbidden, who the contact is.
- Anonymize before pasting. Make it a reflex, not a guideline.
- Review everything that runs. Script, firewall rule, query: read and understood before it is launched.
- Date your choices: a price or a model from November 2026 will not hold in 2027.
Sources
- CNIL: artificial intelligence (French)
- The sources specific to each assistant are in its own article.
