Govern · Watch word feeds
Five feeds that watch every conversation, so you don't have to read all of them.
A watch word feed is a standing rule an agent's conversations are screened against, continuously, without anyone re-reading a transcript to catch it after the fact.
Why a feed, not a spot check
The mistake that matters is the one nobody was reading for.
A support team can sample a handful of transcripts a week. An agent has one conversation with every customer who talks to it. A watch word feed applies the same rule to all of them, every time, and only surfaces the ones actually worth a person's attention.
The five feeds
Pre-built, not something you write from scratch.
- Prompt injection
- Attempts to override an agent's instructions from inside a message or a document it reads.
- PII
- Personal data (PII, short for personally identifiable information) appearing where it shouldn't: in a reply, a log, or a document about to be indexed.
- Toxicity
- Language a company would never want its name attached to, in either direction of a conversation.
- Profanity
- A lighter-weight companion to toxicity, for teams that want a stricter bar on tone alone.
- Starter set
- A general first feed for a new company still deciding what "wrong" looks like for its own agents.
What a feed does when it fires (warn, block, or hold) is set by Governance Mode.
Your own feeds
Add the words that matter in your business.
Anyone can build a feed of their own: product names that must never be promised, a competitor, a phrase your lawyers asked you to watch. It runs next to the pre-built ones, with the same choice of what happens when it fires. The pre-built feeds are maintained by Augela, so an update reaches every agent subscribed to them at once.
Example: a security team
Threat patterns in, alerts out, lessons kept.
- Their own feed
- Attack patterns and scam phrases the team tracks, screened in every conversation.
- Alert on the spot
- When it fires, the conversation is posted to the security channel in Slack.
- Lessons kept
- An analyst marks a new pattern worth keeping, and a reviewer adds it to company knowledge.
How the alert gets there: Agent pipelines. How a lesson becomes company knowledge: Insight elevation.
FAQ
Companies turning these on are asking
Can we write our own feed, or only use the five pre-built ones?
Both. Anyone can build a feed of their own, with the words and phrases that matter in their business, and use it on their agents next to the pre-built ones. The five pre-built feeds (prompt injection, PII, toxicity, profanity and a starter set) are maintained by Augela: when we update one, the change reaches every agent subscribed to it at once.
What actually happens when a feed fires?
Depends on the action you set for it: warn (logged, reply still sent), block (reply stopped), or hold (sent to human-in-the-loop review before it goes anywhere).
Do watch word feeds slow down normal conversations?
No. Every message is screened, but screening only costs time on the message that actually matches. Everything else ships at normal speed.
Screen every conversation, read only the ones that matter.
Twenty minutes with the founder. No slides. The real product.
