Case study · Applied AI
Production AI agents at Zuora: Quick Assist and the Organization Admin agent
Two agents with two different jobs. Quick Assist helps people understand the product without leaving the page they are on. The Organization Admin agent handles access administration for IT, security and compliance teams, through the governed APIs the platform already trusts.
- Role
- Product manager
- Company
- Zuora
- Domain
- RAG, LLM agents, MCP
- Status
- Both agents in production
- ~35%
- of access changes now run through the admin agent instead of tickets
- ~9%
- fewer support tickets after Quick Assist launched
- 47%
- of active users engaged with proactive in-product guidance
The problem: the answer, and the access, were never where the work happens
A user mid-task, unsure how something would behave, usually had to leave the product to find out. Zuora almost always had the answer, in documentation, a community thread or a training video, but spread across separate destinations. Administrators faced the same distance in a different form: who has access to what, across every product suite, and how to change it, meant moving between screens or opening a ticket.
Before
- A help module of hyperlinks that sent users to external sites
- Answers spread across documentation, community threads and training videos
- Access changes made screen by screen, or by filing a support ticket
- Admins unsure who has access to what across Zuora's product suites
With agents in the product
- Help opens beside the workflow, with cited answers and the source one click away
- One search across documentation, community and training content
- Routine access work handled in plain language, through governed APIs
- Critical access routed to a human, with the agent's reasoning shown
Quick Assist: an in-product assistant that cites its sources
Quick Assist launched across all of Zuora's product suites. Support tickets fell about 9% after launch, and customers used in-product help instead of searching the web. It also ships as an MCP tool to Zuora's other help sites, so the experience is consistent wherever customers look for help.
Open help where the work is
Quick Assist replaced the legacy help module with a panel beside the page the user is on. The workflow stays in view, and what it suggests relates to that page.
Retrieve across everything Zuora has published
It combines RAG with keyword search across thousands of documentation, community and training resources, and reaches internal APIs and Zuora's support system through MCP, so customers can act in context rather than just read.
Answer with sources, learn from use
Users see a short answer with citations and can open the full article. Their feedback, and what they ask, shows where the content is thin.
Engage before they ask
User-behavior analytics let it flag a known issue, a relevant new feature, or a webinar for something the customer has purchased, which drove feature discovery and cross-sell.
A question, asked on the page
The page the user is on is the context
Retrieve
RAG plus keyword search
Across thousands of resources
- Documentation
- Community answers
- Training videos
- Internal APIs
- Support system via MCP
A short answer with citations
Every answer links back to its source
The reader
Opens the full article
When they need more
The product team
Sees where content is thin
From feedback and questions
The Organization Admin agent: access work, through governed tools
Built for IT, security and compliance teams, the agent manages human and AI-agent access across all Zuora suites. An admin asks who has access to what, by job function, location, department or other attributes, and the agent grounds its answer in the live access policy. When access needs to change, it makes the change through governed OneID APIs that clear the authorization layer, and it reaches external identity providers through MCP for automated user management.
Admins can also use it to create and scope service accounts and other non-human identities, with explicit approval for critical access. OneID remains the system of record. About 35% of access changes now run through the agent instead of tickets.
How the agent boundary keeps it in check →Product decisions
Ground answers in content we already trust
Quick Assist answers from Zuora's own documentation, community and training material rather than a model's general knowledge, and every answer cites where it came from. A user can always check the source.
Scope autonomy by how critical the action is
The admin agent handles routine requests on its own, assigned by criticality, and routes critical access to a person for approval. The line between the two is a product decision, written down, not left to the model.
Give agents tools, not keys
Neither agent has its own path into the system. They call governed APIs, and every call clears the same authorization boundary, so an agent can only do what those APIs and the requester's permissions allow.
Measure quality, not just usage
Evals, citations and user feedback decide whether an answer is good. Questions that return weak answers point to content that needs work, so the assistant improves the documentation as well as the experience.
Where it fits
These agents sit on the same foundation as the rest of my work. OneID decides who can act and what they may do, including AI agents through MCP authorization. The Observability Connector shows what they did. Agents like these are only trustworthy when both are in place.
Read about the agent boundary →Read the observability case study →Read the OneID case study →Building AI into an enterprise product?
I'm always glad to compare notes on grounding answers, evaluating agents, and giving them safe, governed ways to act.
Connect on LinkedIn