Zero human validation, and that is the right call: the Convention Online AI chain

One document pipeline I delivered requires systematic human validation before anything gets written to the database. Auditability demands it: an error on a timesheet that feeds into payroll is expensive. Here is the other extreme, by deliberate design this time: convention-online.com, a worldwide directory of professional events whose entire content is produced by a fully autonomous AI chain, with no human involvement.

What runs every night, with nobody watching

Every night, four things happen with no human intervention at all:

• Professional events are collected from the web, with automatic deduplication and cleanup.
• A web-searching LLM enriches each entry: identifying the official site and organizer, writing descriptions in four languages (English, French, Italian, Spanish).
• Missing visuals are generated by an image model, on on-demand cloud GPU, with no dedicated infrastructure to maintain.
• Everything is published to a multilingual, SEO-optimized site, updated continuously.

The whole chain runs on a resilient scheduled service: if something fails, it resumes on its own, without intervention.

The real decision was not technical

The question that actually shaped the architecture was not “which model to use” but: what happens if the AI gets it wrong?

On timesheets that feed into payroll, a mistake is costly and propagates. The answer is simple: a human validates, always, no exceptions.

On an automatically enriched event listing, a mistake gets corrected on the next pass, with no serious consequence. The answer becomes: it is fine to automate 100% of it.

Same underlying technology. Opposite business stakes. Therefore opposite architecture.

Industrializing AI is not a single dial

This, to me, is where the real difference lies between an isolated AI project and a successful AI transformation: it is not about automating everything, nor about requiring validation everywhere on principle. It is about setting the dial correctly, business case by business case, based on the real cost of a mistake.

This is exactly the work I do ahead of every automation engagement: mapping the processes, identifying where a mistake is expensive and where it barely matters, then sizing the level of human validation accordingly. The outcome is never a binary choice between “fully automatic” and “fully validated,” but a precise calibration, process by process.