Some content should not cross the border
Localization pipelines translate whatever they are given. That is usually acceptable. Sometimes it is a serious problem.
Market-specific pricing lands in a market where it does not apply. A claim that is legal in one jurisdiction is not legal in another. A case study features a customer who is a competitor to your best prospect in the target market. These are not translation errors. The translation is accurate. They are decisions the pipeline was never asked to make.
The judgment has three outcomes, not two: adapt it, pass it through unchanged, or withhold it. How far to adapt is a fourth question. Sometimes only the currency and the examples need to change. Sometimes the whole argument does, because the reason for buying varies from market to market more often than teams expect.
A person must always decide whether to withhold. The agent is not the one with authority to do that.
How the agent is built in Agent Studio
The per-market rule sheet is the foundation, and it belongs in prompt skills rather than in the conversation. It covers currency, date format, level of formality, regulated claims, and phrases to avoid. Regulatory constraints are rules, not judgments. The agent must read them rather than infer them.
The system prompt specifies the three possible outcomes and requires a reason for each, especially for withholding.
Because sessions last, a regional lead can review the content with the agent across several messages, correct as they go, and keep those corrections for the rest of the conversation.
To handle volume, an ActionFlow is used. The Loop node walks the content set. The Agent Chat node then asks the agent for a decision on each item. If an item is marked as adapt, it goes to a Generate Text node. If it is marked as withhold, it goes straight to a Slack node for a person.
Choose a model per market, because provider language coverage varies widely. This is one of the few cases where using the same model across all markets is worse than choosing different ones.
When a market's rules have settled, most of this work stops being judgment and becomes policy. At that point, put it in a flow.
What this agent is built from
- Prompt skills: the per-market rule sheet, including regulated claims, read rather than inferred.
- System prompt: the three outcomes and the requirement to justify each one.
- Service and model: set per market; provider language coverage varies more than most teams assume.
- Agent Chat (Util): a Loop-driven flow that asks for one decision per content item.
- Generate Text (AI Core): produces the adaptation for items the agent marked as adapt.
- Slack (Human in the Loop): every withheld decision reaches a person.
- When an ActionFlow is enough: If a market's rules are stable and written down, convert them into a workflow. The agent is most useful while you are still learning the market.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.