AI Solopreneur TrendDiscover the World's AI Solopreneurs

Nov 4, 2026 · sophiesbureau.com

A One-Person Operations Partner in France Who Lets AI Draft but Checks Everything: Sophie Kazandjian

Sophie Kazandjian runs a one-person digital operations business in the south of France, setting up systems, automations and websites for small business clients. She uses AI for transcripts, automations, dashboards, documents and drafts, and rebuilt her own website with it. She has also documented where it fails: confidently reporting a rise in email opens that was actually a fall, and reporting success when a third of the items were missing from one of its results. Her rule is that identifiable client information does not go into the AI tools she uses.

Virtual assistants and operations freelancers sell time and reliability. AI promises to multiply the first, but it can quietly undermine the second if nobody checks its work.

Sophie Kazandjian, who runs a one-person digital operations business in the south of France, has written openly about both sides. She uses AI throughout her work, and she has also kept a record of where it went wrong.

This article is for freelancers, virtual assistants and anyone who does operational work for other businesses. We look at how she uses AI, what she has seen it get wrong, the limits she sets, and then, in a separate section, at what her experience still has to do.

Who she is

Sophie Kazandjian describes herself as a digital operations partner. Her business, Sophie's Bureau, is based in Saint-Quentin-la-Poterie in the Gard department of southern France and works with consultants, coaches and small businesses in the UK and Europe. Her About page says she does all client work herself and has no associates or subcontractors. The business is registered in France as an entreprise individuelle, a sole proprietorship.

She sets up and tidies systems in tools such as Airtable and ClickUp, builds websites, manages inboxes, diaries and documents for clients, and helps them with EU-based tools, AI Act readiness and the UK's Cyber Essentials certification. Her website lists monthly retainers from €900 and sprints and projects from €1,600, excluding VAT, and carries testimonials from clients.

How she uses AI for clients

In articles on her website, she describes these uses:

  • Transcripts. She uses Otter for meeting notes and quick transcription, and Sonix for interviews and other audio where accuracy matters.
  • Writing. She uses AI to draft blog posts, social media copy and tricky emails. She wrote in 2025 that adapting one piece for LinkedIn, Instagram and an email subject line takes under a minute.
  • Documents. She uses Claude to create Word documents, slide decks and spreadsheets for work such as client reports and data presentations, and wrote in March 2026 that this gives her hours back every week.
  • A typical day. In an October 2026 account of her work, she described seven jobs from one day for a retainer client and a product demo video for another client. AI was part of five of the eight, including building dashboards, drafting a social media post and creating the voices for the demo video with ElevenLabs.

Claude is her main AI for writing and building, and she uses Mistral's Le Chat for research. Other AI tools on her published list for 2026 include NotebookLM for working across several documents and the AI features of ClickUp and Airtable.

How she uses AI for her own business

  • Automations. She wrote in 2025 that automations built with Make and n8n, combined with AI, had replaced several hours of manual work each week in her own practice, such as invoice summaries from Xero and project updates from ClickUp.
  • Her website. In May 2026, she rebuilt her website with Claude Design in two days, moving 37 articles across, and then spent three more days deploying it with the help of Claude's chat interface. Leaving her previous website platform saves her €336 a year.
  • Her own product. She built Skin Resonance, a small paid digital product priced at €15, with Claude in about a week, and wrote that Claude wrote all of its code while she made the other decisions.

What she has seen AI get wrong

She has written several articles about AI's failures:

  • Confident mistakes. In 2025, she wrote that some newer models make mistakes more often and with more confidence, so checking has become a bigger part of her work. In one case, Claude told her that a client's email open rates had risen by 23% when they had actually fallen.
  • False reports of success. In a case study she describes as a composite based on a workflow simulation, an AI reported that it had met every requirement, but one of its three results contained only 68 of 100 items. Her lesson: "AI does not run basic sanity checks unless you tell it to."
  • Problems it missed. For the demo video in her October 2026 account, she listed nine problems that the AI had not flagged and that she caught by watching the video as the buyer would.

The limits she sets

  • No identifiable client data. Her rule is that identifiable client information does not go into the AI tools she uses. Her About page says such information stays on a local model on her own computer.
  • Choosing tools on ethical grounds. In March 2026, she wrote that she had cancelled her paid ChatGPT subscription in February after OpenAI's deal with the Pentagon. Her updated list of tools says she has also dropped Perplexity for ethical reasons. She calls her use of Claude, from a US company, a trade-off she accepts.
  • Caution about AI agents. In October 2026, she warned against handing an email inbox to an AI agent, citing risks such as prompt injection and copies of mail going to the agent's company and its AI providers.
  • Concern for creativity. She has also written that reaching for generated answers before thinking for yourself can weaken the ability to create.

What experience gave her that AI could not

This section is our analysis, not a claim made by Kazandjian.

  • Knowing what the right answer looks like. She caught the false rise in open rates and the missing items because she checked the results against what she knew a correct answer should be.
  • Designing the system. Choosing which tools to connect and where AI fits in a workflow is operational judgment that comes from experience.
  • Protecting clients. Her rule on client data reflects a responsibility that no tool takes on for her.
  • Deciding what she is comfortable with. Her choice of AI tools is shaped by her own values, not only by features or price.

What you can copy

These are our suggestions, drawn from her story.

  1. Assume AI's summary of its own work may be wrong. Check the result, not the report.
  2. Tell AI to check itself. Ask it explicitly to run sanity checks, and then check anyway.
  3. Keep client data out. Decide what never goes into an AI tool before you start.
  4. Count the misses. Keeping track of what AI gets wrong shows clients, and you, where your value lies.
  5. Use AI for your own business too. A website rebuild or a small product can be a low-risk way to learn a tool's limits.

Notes on the sources

All facts in this article come from Sophie Kazandjian's own website, Sophie's Bureau, including articles published between 2025 and 2026 and her homepage, About and legal pages. Her sole proprietorship is also listed as active in France's public business register, with a creation date of April 2018. The time savings and error counts are her own accounts, and the 68-of-100 case is a composite she based on a simulation. Her website lists her prices but not her revenue or number of clients.

Sources

Related articles