Tiledesk is an open-source, no-code platform for building AI agents and connecting them with human support. Its current product focuses on visual workflows, knowledge retrieval, prompt chains, multichannel conversations, APIs, tools, and human-in-the-loop escalation.
Affiliate disclosure: this article preserves the original Tiledesk creator URL. AI Tools Arena may receive a referral benefit if you use it, at no additional cost to you.
What Tiledesk Includes
The official Tiledesk site currently describes a visual agent designer, connected knowledge, full-text and semantic retrieval, prompt chains, multi-agent workflows, human handoff, APIs, MCP tools, web widgets, and channels including WhatsApp, Messenger, SMS, email, voice, and web chat. Exact access depends on the deployment and plan.
Plan the Agent Before Building
Start with one narrow support outcome, such as order-status questions or product documentation. Define what the agent may answer, which sources it may use, which actions it may take, and when it must transfer the conversation. Automation without these boundaries creates avoidable support and compliance risk.
Tiledesk Tutorial
1. Define Intent and Success
List the ten most common requests, the approved answers or data sources, and the success metric. Useful measures include resolved conversations, correct escalations, containment rate, response time, and reviewer-rated accuracy.
2. Create a Dedicated Project
Use a separate project or environment for development. Assign only required access, document the owner, and keep production credentials out of prompts and sample conversations.
3. Connect Curated Knowledge
Add current help articles, policies, product documentation, and structured data. Separate unrelated collections so retrieval sends each question to the correct source. Remove duplicate, expired, contradictory, or internal-only material.
4. Design the Conversation Flow
Use the visual builder to greet the user, identify intent, retrieve knowledge, collect necessary details, call approved tools, return an answer, and offer escalation. Smaller prompt blocks are easier to test than one oversized instruction.
5. Add Guardrails
Tell the agent to stay within the approved scope, cite or name its source when useful, avoid inventing policy, and refuse sensitive or unsupported actions. Require confirmation before any irreversible transaction.
6. Configure Human Handoff
Escalate when confidence is low, the customer asks for a person, authentication fails, policy exceptions are requested, sentiment deteriorates, or a regulated topic appears. Pass the transcript, detected intent, and collected details so the customer does not repeat everything.
7. Test in the Debugger
Test normal questions, spelling errors, multilingual variants, conflicting sources, prompt injection, missing order numbers, angry users, and unavailable back-end services. Confirm that failures are safe and understandable.
8. Deploy One Channel First
Begin with a limited web widget or internal group. Verify identity, consent, retention, notifications, and escalation before expanding to messaging, email, SMS, or voice.
9. Review Conversations Weekly
Sample resolved and escalated conversations. Label wrong retrieval, incomplete answers, policy violations, broken actions, and unnecessary handoffs. Update the source or workflow—not just the wording of one response.
Launch Checklist
- Approved knowledge with owners and review dates
- Clear scope and prohibited actions
- Fallback for missing or conflicting information
- Tested human escalation with transcript transfer
- Privacy notice, consent, and retention controls
- Least-privilege API credentials
- Monitoring for quality, latency, cost, and failures
- Rollback or disable procedure
Common Mistakes
- Connecting every document without curation
- Allowing the agent to answer beyond policy
- Launching every channel at once
- Measuring only containment instead of correctness
- Hiding access to a human agent
- Putting secrets or personal data in prompts
Frequently Asked Questions
Is Tiledesk open source?
Tiledesk currently presents itself as open source and also offers cloud and enterprise deployment options. Check current licensing and hosting requirements for your use case.
Can it transfer conversations to live agents?
Yes. Human-in-the-loop support and live-agent handoff are central current capabilities.
Does an AI support agent eliminate human review?
No. Knowledge, workflows, actions, permissions, and sampled conversations all require ongoing review.
Final Takeaway
Tiledesk is strongest when the agent has a narrow job, curated knowledge, constrained actions, and reliable escalation. Launch on one channel, measure answer quality as well as automation, and expand only after the workflow is dependable.