How Perplexity gave its support team control of its Decagon AI agent
Perplexity's support team uses Decagon to run an AI agent named Sage across chat and email. Support operators write and refine its procedures in natural language, and the team reports a high deflection rate as volume grows.
The problem
Perplexity started with support handled through a single email channel. As the product grew, daily support volume rose 269% and users expected near-immediate replies across time zones. The team needed to decide which work could be automated consistently and which needed human judgment, without waiting on engineering for each policy change.
How they did it
- Deploy Decagon on chat and email support channels.
- Have support operators write policies and workflows as Agent Operating Procedures (AOPs) in natural language.
- Use Duet to turn an existing policy document into a first-draft AOP, then review, tighten and test it.
- Use Duet to surface gaps and edge cases in existing AOPs before updating them in production.
- Monitor user intent, escalations and sentiment in Decagon reporting, and ask Duet which AOPs have different deflection.
- Keep complex cases such as escalations, incidents, enterprise billing and privacy requests with human agents.
Results
- As reported by Decagon, Perplexity reached a 78% deflection rate over the last 30 days.
- As reported by Decagon, the agent supports 70+ user intents.
- As reported by Decagon, average daily support volume rose 269% compared with last year.
- Operators can change procedures directly instead of waiting for an engineering ticket, and Duet helps them spot edge cases.
As reported by the source (Decagon customer story); AgentGid did not measure these figures.
This suits a support team with steady multichannel volume, operators comfortable writing and testing procedures in plain language, and a sales process they can work through, since Decagon has no public pricing. Treat the 78% deflection and 269% volume figures as Decagon-reported, not independent. For a smaller or cost-sensitive team, Rasa (Free) is open source, though it is limited to 1,000 external conversations a month.
The agent used here
Similar use cases
How Substack automated Tier 1 reader and publisher support with Decagon
Substack used Decagon's AI agent to handle repetitive support requests such as cancellations and email imports. Decagon reports that most inquiries are now resolved without a huma…
Decagon reports that the agent resolves more than 90% of user questions without human intervention.
Triage support tickets with an n8n AI Agent and confidence-based routing
An n8n workflow cleans and validates incoming tickets with plain code nodes, uses an AI Agent only to classify urgency and type, and routes each ticket based on how confident the …
The suggested routing: above 0.85 processes autonomously, 0.6-0.85 is processed but flagged for review, and below 0.6 goes to a human.
How OpenTable resolved restaurant and diner questions with Agentforce
OpenTable launched Agentforce agents on its website for restaurants and diners, grounded in its existing knowledge articles. Salesforce reports that most restaurant cases were res…
Salesforce reports 73% case resolution within 3 weeks of launch for the restaurant agent.