AI platforms: products where AI does real work, with checks around it

What we run on AI today

• A content engine that improves 38 of our own websites each working day: it researches, writes, and automatic checks refuse any change with an unsourced figure, an unverifiable study, copied text or a broken page — about one change in four is refused.
• A video app (Mac and Windows, with automatic updates) that turns a brief into a sourced script and renders the video in four formats.
• Conversational agents on WhatsApp and on websites, with daily caps, dry-run mode, an off switch and human handover — stages from prototype to live, see WhatsApp AI agents.
• A document corpus of 1,226 documents prepared to serve as the memory of a CRM assistant.
• An MCP server that lets a whole team plug their own Claude into the company's operating system, every call logged — see Claude & MCP integrations.

Workflow or agent? Start simple

Anthropic, which builds the Claude models, distinguishes workflows — "systems where LLMs and tools are orchestrated through predefined code paths" — from agents, "systems where LLMs dynamically direct their own processes and tool usage". Its advice is to find "the simplest solution possible, and only increasing complexity when needed", which can mean not building an agent at all, and to test extensively "in sandboxed environments, along with the appropriate guardrails".

That is how we design. Most business features are workflows: a fixed sequence (fetch data, draft, check, ask for approval, publish) where AI does one step well. We use agents only when the steps genuinely cannot be known in advance, and we give them limits: a list of allowed actions, a budget, a dry-run mode and logs.

Quality: measured, not hoped for

Language models can produce confident, wrong text. Anthropic's own guidance lists ways to reduce this: allow the model to say it does not know, ground answers in direct quotes from the documents, make claims auditable with citations, and restrict it to the information provided. It adds that these techniques reduce errors but do not eliminate them. So we treat AI output like any input from an unreliable source: checked by code where rules can be written, reviewed by a person where they cannot, and measured on a test set every time something changes.

Rules that apply to AI features

In the EU, the European Commission explains that people using AI systems such as chatbots should be made aware they are interacting with a machine, with transparency rules applying from August 2026. When the AI processes personal data, the CNIL asks for a defined purpose, minimal data, clear information to the people concerned and a retention period fixed in advance. We build these in: AI features say they are AI, data sent to a model is limited to what the task needs, and we tell you which provider processes it and under which terms.

How a project runs

1. Define the job and the bar. What the AI does, what a good output looks like, what must never happen.
2. Collect real cases. Fifty to a few hundred examples become the test set.
3. Prototype. The simplest version, run against the test set.
4. Add checks and approvals. Automatic refusals for rule breaks, human approval where needed, caps and logs.
5. Ship through dev, uat, prod. With previews and builds on GitHub Actions; a person merges every change.
6. Monitor. Logs reviewed, the test set re-run after any model or prompt change.