Your agent will translate your campaign line into French in 800 milliseconds, with total confidence, and it has no idea whether the result will land with an audience in Suisse romande or quietly embarrass your brand for a quarter. It returns the same confident string either way. The output is uniform; the risk isn’t.

That’s an architecture problem before it’s a language problem. Every agent stack has a translate call that returns instantly with no quality signal attached, while human review lives entirely outside the system – a ticket, a spreadsheet, a two-week loop with an agency, a person who is not in the room when the agent runs. So teams pick a failure mode: gate everything behind humans and ship nothing, or ship everything unverified and find out from a customer. What’s missing is verification as something you can call – with a latency budget, a state, and a policy for when it fires.

This workshop makes human expertise a callable primitive. We build a multilingual webshop live and wire it to an MCP server that exposes professional verification as just another tool: the agent translates everything in seconds, flags what it is uncertain about and what the business has marked as high-stakes, and calls in a qualified linguist. Corrections return asynchronously, within minutes, and update the live page. The webshop is the demo; the pattern applies wherever machine output is good enough ‘most of the time’ and ‘most of the time’ isn’t good enough – support tickets, help centres, and the legal sections nobody reads until they matter.

You leave with:

• The demo code – webshop plus MCP integration – to fork and reuse for whatever you're building.
• A set of triggering policies (page visibility, content class, model uncertainty, business rules) and the trade-offs of each.
• A state model for asynchronous quality: how to run a live surface where every string is machine-translated, pending verification, or verified, without confusing your users or your compliance team.
• A way to reason about the cost of selective verification against the cost of being wrong or outsourcing full pages to an agency.