AI workflows and agents
Multi-agent service desk
A coordinator routes requests between classification, knowledge and action agents, with approval before any effect.
Fictional scenario. An example situation that illustrates the approach. It is not a client project and describes no measured results.
Business problem
A fictional field-service company receives requests by e-mail, chat and phone. Classification is slow and knowledge is scattered across departments.
Assumptions
- There is a helpdesk with an API.
- The knowledge base is curated and current.
- Actions (e.g. scheduling) have APIs.
Constraints
- Agents do not act towards customers without approval.
- Every step is traceable.
- No new systems for the customer.
Reference architecture
Components by stage — an illustration, not a running system and not a certified design.
01Trigger
- New helpdesk request
02Orchestration
- Coordinator agent
- Routing rules
03AI and rules
- Classification agent
- Knowledge agent (RAG)
- Action agent
04Review
- Approval point
- Escalation to a person
05Output
- Reply or task
- Step log
The diagram as a list
| Stage | Components |
|---|---|
| Trigger | New helpdesk request |
| Orchestration | Coordinator agent, Routing rules |
| AI and rules | Classification agent, Knowledge agent (RAG), Action agent |
| Review | Approval point, Escalation to a person |
| Output | Reply or task, Step log |
Interfaces
| From | To | Via |
|---|---|---|
| Helpdesk | Coordinator | Webhook + REST |
| Knowledge agent | Knowledge index | Read-only search |
| Action agent | Calendar | OAuth with a limited scope |
Data flow
- The request arrives and is classified by type and urgency.
- The coordinator picks an agent and passes context.
- The knowledge agent drafts a reply with a source.
- An action is prepared and waits for approval.
- The result and the path are logged.
Security boundaries
- Each agent has its own minimal permissions.
- Customer input is treated as untrusted (prompt injection).
- Agents see no data beyond the request.
Where a person approves
- Replies to customers on sensitive topics.
- Every action with an effect in the calendar or CRM.
Delivery phases
01 · Discovery
Request types, volumes and handover rules.
02 · Pilot
Classification and draft replies, nothing sent.
03 · Expansion
The action agent behind approval.
04 · Operation
Evaluations before changes, escalation monitoring.
Limitations
- A multi-agent design is harder to debug.
- Routing rules need an owner.
Trade-offs
- Coordinator vs fixed flow — flexibility vs predictability.
- More agents — clearer responsibilities but higher cost.
Related services
Related assessments
Integration patterns used
Workshops and assessments
Facing something similar?
We start with an assessment of where you are and propose a realistic first phase.