Freelance AI consultant
for SMBs in Paris.

Integration starts with identifying 1-2 simple use cases whose gain is easy to measure (email triage, quote generation, FAQ agent, RAG search) in **shared scope between your team and me**. First results in 4 to 8 weeks, as a guide: the timeline depends on the state of your data. **My scope (technical layer)**: LLM integration via API (Claude, GPT-4, Mistral), prompt engineering, RAG on your sources, technical guardrails, selection of a provider with EU hosting when possible. **Your scope (legal / GDPR layer)**: your DPO or lawyer validates DPAs, updates the processing register, informs data subjects, conducts DPIA if needed. Without an identified DPO or legal counsel, I may limit the mission to use cases not requiring complex GDPR validation.

Claude for long tasks and polished French writing, GPT for versatility and mature tooling, Mistral for EU sovereignty when data is sensitive. GPT-4 (OpenAI) remains the most versatile and tooling-mature (function calling, vision, audio). Claude (Anthropic) is better on long tasks, structured reasoning and polished French writing. Mistral (French, EU hosting) is best when data sovereignty is a priority (general-purpose law firms, accountants on standard B2B data). I often recommend a mix: Mistral on the backend for sensitive work, GPT or Claude for premium tasks. **Out of scope**: sectors requiring specific certifications (HDS healthcare, certified public sector, defense, OIV operators), those missions need certified hosting and regulatory expertise I don't have in-house.

Simple site FAQ agent (LLM + 10-50 document knowledge base): €2,500 to €5,000. Conversational agent with memory and CRM integration (24/7 client response, booking, human escalation): €5,000 to €12,000. Complex business agent (quote generation, document analysis, archive RAG search): €12,000 to €30,000. Recurring inference fees: €30 to €300/month depending on volume.

Yes, subject to four principles, but **legal validation remains the responsibility of your DPO or GDPR legal counsel**, not mine. **What I deliver** (technical layer): (1) selection of an LLM provider with EU endpoint when possible (Mistral La Plateforme, OpenAI Europe, Anthropic via AWS Bedrock EU); (2) technical configuration to avoid sending non-anonymized personal data; (3) documentation of LLM subprocessors used; (4) technical guardrails (human validation on critical actions, source citation). **What remains your responsibility** (legal validation): drafting and signing DPAs, updating your processing register, informing clients in your privacy policy, conducting DPIA if needed. Without an internal DPO or identified GDPR legal counsel, I may decline the mission or limit it to use cases not requiring complex validation. I **do not provide off-the-shelf DPA or legal notice templates**, that's the job of your DPO or lawyer. I can connect you with specialized partners.

RAG (Retrieval Augmented Generation) connects an LLM to your document base (CRM, drive, archives, internal FAQ) so it answers with your data rather than its generic knowledge. Essential when you want: an FAQ agent answering with your real pricing, intelligent search in your legal or accounting archives, an assistant drafting quotes with your standard clauses. Typical implementation: 4-8 weeks, vector embeddings (Pinecone, Supabase pgvector) + Claude/GPT/Mistral LLM.

Four cumulative techniques, but **no guaranteed quantified precision**: LLMs remain probabilistic by nature. (1) **Strict RAG**: AI only answers with your sources, not its general knowledge. (2) **Structured prompts** with explicit constraints and examples. (3) **Post-generation verification**: AI must cite sources, a second LLM can check coherence. (4) **Business guardrails**: human validation required for any critical action (sending quotes, CRM edits, payments, external email). Combined, these techniques significantly reduce errors but don't eliminate them. **I don't commit to a quantified error rate**: real-world feedback is used to tune prompts and guardrails, not to promise absolute precision. Truly critical actions always remain under human validation.

A simple site FAQ agent (with your knowledge base): 2 to 4 weeks after brief. A conversational agent with CRM integration: 4 to 8 weeks. A complex business agent (RAG, document generation, multi-step): 8 to 16 weeks. The limiting factor is almost never code, it's your knowledge base quality: if clean and structured, we move fast. If scattered across 12 tools, we consolidate first.