Agentic review
AI writes faster than your team can read.
Review steps check every value your agents write against your own rules, and only ask a person when something fails or needs a decision.
Face serum, 30 mlDescription agent · 3 review steps- 1Warm and plain, no superlativesHold for review
- 2No health claims the product data does not backReject
- 3Never use: miracle, flawless, guaranteedIgnore it
- Highlights
- Description
- Description, German
Of three values, one reaches your team.
Sound familiar?
The writing got fast. The checking did not.
Once AI writes the descriptions, the bottleneck moves to the people who have to read them. One wrong claim is worse than an empty field, so nobody dares to skip the check.
How review works in Cernel
You choose what goes live on its own, and what gets a review step.
You choose which values go live automatically and which get a review step; review steps check every value and only ask a person when something fails or needs a decision.
A person approves
Review and approval
With Review on, AI-written values wait as pending until someone approves, edits or rejects them, one at a time or in bulk. Set the default once and change it per run, per automation or per attribute. Every AI-written value shows the inputs it used, the reasoning and the prompt.
Agents check
Review steps
Checks on each agent, each one rule in plain words. After every run they read each value and return passed or failed, with the reasoning. You choose what a fail does: let it through, hold it for a person, or reject it.
Runs by itself
Automations
An automation runs the agents when a product is created or its data changes, so new products are filled and checked without anyone starting a run. Values then follow the review setting you chose.
Review by exception
Your team reads what was flagged, not the whole catalogue.
Your review steps read each value straight after the agent writes it, the way a person checking another agent's work would. What passes goes through on its own. Only what fails stops: with a person, or rejected.
1,000 new products, 3 text fields, 4 languages
- 0
- values checked by your review steps
- 0
- went through on their own
- 0
- wait for a person
- 0
- rejected, never live
Teddy bear, GreyDescription Hold for review
A soft grey bear with a knitted scarf, made from organic cotton. Safe from birth.
Claims against the product data failed. The product data says Material: Polyester.
Three outcomes
You decide what a flag does. When two disagree, the stricter one wins.
Each review step is one rule in plain words, tested on your own products in the Agent Sandbox before it runs. For each you choose what a fail does: note it and let the value through, hold it for a person, or reject it so it never goes live. When a value fails two steps, the stricter outcome wins, so a hard compliance rule is never softened by a lighter one.
Dining chair Description
A straight-backed dining chair in solid oak, made from FSC-certified wood, with a light seat pad.
- Material
- Beech
- Finish
- Oak stain
- Certifications
- None
- Warm and plain, no superlativesIgnore it
- No sustainability claims without proof in the dataHold for review
- No claims the product's attributes do not includeReject
- Ignore itNoted, goes through
- Hold for reviewWaits for a person
- RejectNever goes live
Two steps failed. The stricter one wins: rejected, it never goes live.
Written in your words
A check is one sentence. You test it on your own products first.
Each review step is one small rule, phrased as when a value should fail. You write them in the Agent Sandbox, run the agent on a handful of your own products, read every verdict and the reasoning behind it, and save the agent when the checks catch what you want caught.
Description agent, three review steps
Tone of voice
Fail if the text sounds casual, uses superlatives or exclamation marks. Our voice is warm and plain.
- If this check fails
- Hold for review
- Caught in the test
- The most stunning bag you will ever carry!
Banned words
Fail if the text uses any of these words: miracle, flawless, guaranteed, eco-friendly.
- If this check fails
- Ignore it
- Caught in the test
- A flawless glow in one drop.
Claims against the product data
Fail if the text states a material, certification or health benefit that the product's attributes do not include.
- If this check fails
- Reject
- Caught in the test
- Made from organic cotton. (Material in the data: Polyester)
Every market
Every language passes the same checks as the original.
When an agent writes in several languages, each translation is read by the same review steps, with the same outcomes. A superlative that slips into German waits for a person, just as it would in English. The same steps cover every language; where a rule differs by market, say so inside the step.
Molly & My
Molly & My builds every market on one foundation
- products enriched across 8 markets
- 20,000
- faster time to market
- 4x
- organic traffic to product pages
- +29.3%
Firtal Group
Firtal Group added a thousand extra hands without adding headcount
- products enriched
- +25K
- organic traffic to product pages
- +20%
- faster time to market
- +80%
Tiffany Lighting Direct
Tiffany Lighting Direct gets a new supplier range live in days, not months
- to get a new supplier range live
- Days
- products from 20+ brands
- 13,000+
- extra hires needed
- 0
FAQ
Questions about agentic review.
What is agentic review?
It is a review step that checks what Cernel's agents write. You add checks to an agent, each in plain words. After every enrichment an AI reviewer reads each value the agent wrote against those checks and returns passed or failed, with its reasoning. It does the work of a person checking another agent's output, so your team only sees what was flagged.
What can a review step check?
Anything you can say in a sentence about the text: tone of voice, words you never use, format and length, the language it is written in, or claims checked against the product's own attributes, such as a material, a certification or a health benefit the data does not include.
What happens when a value fails a check?
You choose per step. Ignore it: the failed check is noted and the value goes through. Hold for review: the value waits for a person to approve or reject it. Reject: the value never reaches your catalogue. The value is always kept with its history, so a held or rejected value is never lost. To replace a rejected value, sharpen the agent's prompt and run it again.
What if a value fails two checks with different settings?
The stricter outcome wins. If any failed step rejects, the value is rejected. If none reject but one holds, the value is held. It only goes through when every failed step is set to ignore it.
Do values that pass go live on their own?
They follow your normal review setting, which you choose per run, automation or attribute. If you approve values before they are published, passing values wait in Review as usual; if you let an agent publish on its own, they go through. A value a step holds always waits for a person, even on a run where approval was otherwise off.
Which values can be reviewed?
The values your agents write, such as titles, descriptions and highlights, in every language they write. Review steps belong to the agent, so every run of it is checked the same way.
How do I know a check works before it runs on the whole catalogue?
Review steps run in the Agent Sandbox. Test the agent on a handful of your own products, open each result's checks to see how the reviewer judged it and why, and adjust the step until it catches what you want. Nothing changes in your catalogue until you save the agent.
What happens if a check cannot run?
Cernel never lets an unchecked value through. If the review steps cannot run for an enrichment, the value is marked Failed instead of reaching your catalogue.
Where to next