Vendor bot or custom
Use the desk vendor's channel and widget, with retrieval and guardrails in your own orchestrator.
Why. Fast delivery with control over data and behaviour.
Trade-off. More integration work than a pure vendor bot.
CRM and business systems · Zendesk and ServiceNow
Service desk with chatbot deflection and a knowledge base RAG
Zendesk or ServiceNow as the system of record for tickets, a chatbot that answers from approved knowledge with citations, and a clean hand-over to a person with full context when it cannot.
A reference architecture from EITSAA, not a named client system. Service limits, prices and product features change, so confirm current details before you commit to a design.
- Voluntary AI Safety Standard
- Australian Privacy Principles
- ACSC Essential Eight
- WCAG 2.2 AA
Architecture
- People and clients
- Network and edge
- AI and ML
- Messaging, events and workflow
- SaaS and business systems
- Security and identity
- Monitoring, audit and analytics
- Request or data flow
- Asynchronous or optional
- Boundary: region, account, network or trust zone
Request and data flow
The user writes in the chat widget, Teams or another channel. The gateway authenticates signed-in users and limits abuse.
Personal information the bot does not need is redacted, and the conversation is tagged with channel and customer reference.
The orchestrator classifies intent and retrieves passages from the approved knowledge index only.
The model drafts an answer from the passages with citations. If retrieval is weak, the bot says it cannot answer rather than guessing.
On low confidence, a request for a person, a sensitive topic or repeated failure, the bot creates or updates a ticket through the API.
The ticket carries transcript, sources shown and customer details, routed by category and priority.
Knowledge changes are published in the help centre and synced to the index, with withdrawn articles retired.
Components and why they were chosen
| Component | What it does | Why this choice |
|---|---|---|
| Zendesk Support and Guide, or ServiceNow CSM and Knowledge | Tickets, SLAs, routing and the authoritative knowledge base. | Keep process and knowledge in the tool agents already use. |
| Chat gateway | Authentication, throttling and redaction. | A control point before any model call. |
| Bot orchestrator | Intent, retrieval, prompt, guardrails and escalation rules. | Testable logic outside the vendor widget. |
| Knowledge index (OpenSearch or pgvector) | Search over approved articles. | Hybrid retrieval handles product names and codes. |
| LLM via Bedrock or Azure OpenAI | Generates grounded answers. | Enterprise terms and regional options. |
| Ticket API integration | Creates tickets with context. | Customers do not repeat themselves. |
| Analytics and evaluation | Deflection, escalation reasons and answer quality. | Measures real value and shows knowledge gaps. |
| Accessible chat UI | Keyboard and screen reader support, plain language. | Needed for WCAG 2.2 AA in public services. |
Design decisions and trade-offs
Knowledge source
Only published, owned articles.
Why. Every answer traces to an owner.
Trade-off. Content owners need time to fix gaps.
Escalation
Always an easy route to a person, and escalate after repeated failure.
Why. Trust and fairness, especially for vulnerable customers.
Trade-off. Lower raw deflection, higher satisfaction.
Metrics
Resolution and satisfaction, not only deflection.
Why. Deflection can simply mean frustration.
Trade-off. Needs follow-up surveys.
Well-architected pillars
Security
Redaction, authenticated sessions for account data, no write actions without checks and prompt-injection testing.
Reliability
Fall back to search results or direct ticket creation if the model is unavailable.
Performance
Streaming answers, cached top queries and short context.
Cost optimisation
Tokens per conversation and desk licences. Deflection reduces ticket volume.
Operational excellence
Weekly review of unanswered questions with content owners, and release tests.
Sustainability
Reuse articles and use smaller models for classification.
Network, identity and data
Networking
The orchestrator runs behind API Gateway or Front Door with WAF. Desk APIs are called over TLS with scoped tokens. Model access is private.
Identity and least privilege
Signed-in customers pass a token so the bot can answer account questions through the API with that user's permissions. Anonymous mode uses public knowledge only. The desk integration uses an API user with ticket scope only.
Data and encryption
Minimise and redact. Define transcript retention. Tell users they are talking to an AI and how their data is used (APP 1 and 5). Do not train on customer conversations without a clear basis.
Security and compliance mapping
| Control | How this design addresses it |
|---|---|
| Transparency | Clear AI disclosure and an easy human option, in line with the Voluntary AI Safety Standard. |
| Privacy | Redaction and retention support APP 11. |
| Restrict administrative privileges | Scoped API user. Essential Eight. |
| Accessibility | WCAG 2.2 AA testing of the chat UI. |
Mappings show how the design supports each control. They are not certifications, and EITSAA does not claim IRAP assessment or compliance for any reference architecture.
Failure modes, targets and scaling
- Target RTOUnder an hour. The human channel remains available immediately
- Target RPOTranscripts near zero loss
- Scaling and limitsHundreds of concurrent conversations is typical. Check model quotas and desk API limits.
| Failure mode | Mitigation |
|---|---|
| Wrong or outdated answer | Only published articles, expiry dates on content and evaluation sets. |
| Bot loop frustrates the user | Escalate after two failed attempts. |
| Desk API outage | Queue the ticket and tell the user it has been recorded. |
| Injection through user-submitted content | Separate instructions from data and avoid side-effect tools. |
Observability and runbooks
- Resolution rate, escalation rate with reasons and customer satisfaction.
- Top unanswered questions by volume.
- Groundedness sampling reviewed weekly.
Cost drivers and levers
Tokens, desk licences, hosting and content upkeep.
- Route simple intents without the model.
- Cache answers.
- Improve the knowledge behind the top questions.
Implementation and fit
Phased approach
- Audit and fix the top 50 knowledge articles.
- Launch retrieval-only answers to staff.
- Add the customer channel with escalation.
- Add authenticated account lookups.
Anti-patterns to avoid
- Training on messy documents.
- Hiding the human option.
- Measuring only deflection.
- No content owner.
When to use it
- High volume of repetitive questions.
- Organisations with sound knowledge content.
- Councils and utilities with public enquiry volume.
When not to
- Sensitive or complex advice with legal consequences.
- No content to ground on.
More in crm and business systems
Entra ID and Okta
Identity and single sign-on across CRM and business apps
One identity provider for staff, partners and customers: SAML and OIDC federation, SCIM provisioning, Conditional Access, role mapping and a leaver process that really removes access.
Microsoft 365
Secure Microsoft 365 workplace, ready for Copilot
A Microsoft 365 tenant for councils and small organisations that is safe before AI arrives: Entra ID with Conditional Access, Intune-managed devices, Purview labels and DLP, oversharing cleaned up, and a backup copy outside the tenant.
Salesforce
Salesforce Sales and Service Cloud integrated with ERP through MuleSoft
API-led connectivity between Salesforce and a finance or ERP system: system APIs wrap the ERP, process APIs hold the business rules, Salesforce calls one stable contract, and events keep both sides in step.
Related insights
Discuss this architecture
Tell us your systems, constraints and timeline. We reply within one business day, and the initial conversation is free.