Put a check on every enrichment: Review Steps

October 20, 2026 AI enrichment can write thousands of values in minutes, which raises an old question in a new form: how do you keep the weak ones out without reading every single one? Approving values by hand does not scale, and letting everything through means an off-brand claim or an unsupported statement can reach your catalog before anyone notices. Review steps put the check next to the work. You add a check to an agent, describe in a prompt what a good value looks like, and choose what should happen when a value fails. From then on, every value that agent writes is inspected against your rule, and the ones that fail are held back or rejected before they go live.

​

A prompt that judges the output

A review step is a prompt, the same plain-language instruction you already write for agents, pointed at the value the agent just produced. For example: “Fail the enrichment if it claims the product is sustainable, recycled or eco-friendly without the product data saying so.” The step returns a simple verdict, Passed or Failed, on every value the agent writes.

​

You decide what a failure means

Each check sets its own consequence under If this check fails:

• Ignore it lets the value through and records the failed check for later.
• Hold for review sends the value to a person to approve or reject before it goes live.
• Reject stops the value from reaching your catalog at all.

An agent can carry up to three checks, and when a value fails more than one, the strictest consequence wins. The value itself is always kept with its history, so a held or rejected value is never lost, it just cannot go live until you say so.

​

The verdicts, where you review

On the Review surface, each value shows whether it passed or failed its checks, with a per-step breakdown and a note explaining the consequence that was applied. Open How the reviewer reached this to read the reasoning behind a verdict, or ask for it on demand when you want to understand a surprising call. Rejected values are tucked out of the way until you choose to see them.

​

Why it matters

Trust is what lets you run enrichment at scale. Review steps encode your standards once and apply them to every value automatically, so the cases that need a human eye are held for you and the clear violations never ship. You get the speed of AI enrichment and a guardrail you wrote yourself.