Enterprise use case · Zendesk + OpenAI + Microsoft 365
Drafted replies with human approval
Agents repeat the same policy and troubleshooting explanations while checking that each answer fits the actual customer's situation. Support agents receive a grounded response draft they can edit and approve before anything is sent.
Systems involved
What each system contributes
The model step runs against the AI provider you choose, including Anthropic, Google Gemini or a model you host yourself. Anything outside a pre-built action runs through an API, protocol or database call, or as Rules + Code in Python, Bash, Go, Java, .NET, PHP, PowerShell or JavaScript/TypeScript, with the same approvals and audit.
Architecture
One governed run
How it starts: A support agent requests a draft for an assigned Zendesk ticket.
Step by step
How the process runs
- 01
Assemble authorized context
Read the permitted Zendesk conversation and approved reference material from the configured Microsoft 365 email or document interface.
- 02
Retrieve relevant guidance
Select applicable passages for the ticket while retaining their version and access restrictions.
- 03
Draft the response
The selected model writes an answer grounded in those passages, identifying unsupported questions instead of inventing a policy.
- 04
Check the boundaries
Rules flag restricted content, unsupported commitments and missing response requirements before presenting the draft.
- 05
Release the agent's answer
After the named agent approves or edits the draft, a separate Zendesk or Microsoft 365 action sends through the designated channel.
Control
Human decision and exception path
Human decision
The assigned support agent approves the exact response before sending; refund offers, contract concessions or other monetary commitments need the relevant authority as well. An unapproved model draft remains unsent.
Exception path
- Missing or contradictory guidance stops the proposed answer and asks the agent to resolve the source conflict.
- Unsupported dates, promises or facts are flagged for removal or verification.
- A failed or uncertain send is checked against the destination conversation before another message is released.
Evidence
What gets audited
- Ticket and approved knowledge-source references.
- Model draft, instruction version and supporting passages.
- Agent edits, approvals and final response text under the agreed logging policy.
- Send request and destination confirmation.
Nexus runs on premises, in private or sovereign cloud, or air-gapped. Credentials stay in the vault, roles decide who can approve, and the execution record stays inside your deployment boundary.
Sovereignty and deploymentRelated
Patterns, connectors and next steps
- ServiceNow-Jira integration
Integration pattern · Incidents and changes reach engineering with context, and status returns to service.
- Zendesk connector
52 pre-built actions
- OpenAI connector
25 pre-built actions
- Microsoft 365 connector
45 pre-built actions
- Human-in-the-loop
Where approval belongs in a running workflow.
- Order to cash across CRM, ERP and billing
Another enterprise use case.
FAQ
Questions about this use case
- Does Nexus replace any of these systems?
- No. Each system keeps its own records. Nexus runs the process between them and records each decision.
- How does the run start?
- A support agent requests a draft for an assigned Zendesk ticket.
- What if a step is not covered by a pre-built action?
- The model step runs against the AI provider you choose, including Anthropic, Google Gemini or a model you host yourself. Anything outside a pre-built action runs through an API, protocol or database call, or as Rules + Code in Python, Bash, Go, Java, .NET, PHP, PowerShell or JavaScript/TypeScript, with the same approvals and audit.
- Can the draft be saved without being sent?
- Yes; drafting, approval and sending are distinct steps, so the agent can revise or discard the answer.
- Does the assistant make up an answer when the knowledge base is incomplete?
- The workflow treats unsupported questions as gaps and returns them to the agent rather than publishing a speculative answer.
Use case review
Discuss this use case
Bring your version of the process. We will map the systems, approvals and evidence it needs.