Loading page
About this tool
Define an ordered rule set for when your chatbot must hand a conversation to a human agent.
The Chatbot Escalation Rules Planner generates an ordered, first-match-wins rule set defining exactly when a chatbot must hand a conversation to a human agent. It builds on established conversation-design conventions — the two-strikes repair rule for repeated misunderstanding, an NLU intent-confidence floor (commonly around 0.7), unconditional handoff when a user asks for a human, and safety-first ordering for sensitive topics. It is built for support leads and conversation designers preparing a bot for launch or tightening one that traps users in loops.
Open Chatbot Escalation Rules Planner on AltFTool — it loads instantly in your browser.
Add your input to the workspace.
Adjust the options until the result looks right.
Copy or download the output and put it to work.
Self-harm, fraud and complaint topics outrank cost triggers, and the self-harm rule can never be switched off.
Two failed attempts and a 0.7 confidence floor reflect real conversation-design and NLU-platform conventions, and both are tunable.
If agents only work business hours, every rule gains a queue-and-callback path instead of a dead end.
At minimum: immediately when the user explicitly asks for a human, after two consecutive failed attempts to understand the same request, when intent confidence is too low to act safely on a consequential action, and always for sensitive topics like self-harm, fraud or formal complaints. High-value transactions and strongly negative sentiment are common additional triggers — this planner turns each into an ordered rule.
Around 0.7 is a common conservative floor for acting on an intent without clarification; NLU platforms such as Dialogflow ship classification thresholds in the 0.3–0.7 range and treat scores below the threshold as no-match. The right value depends on your platform's score distribution — set it too high and everything escalates, too low and the bot acts on guesses.
Because a third re-prompt almost never recovers the conversation and measurably increases frustration — conversation-design guidance (including Google's error-repair patterns) recommends at most one or two repair attempts before changing strategy. The two-strikes rule stops the loop, summarises what the bot understood, and passes that summary to the agent so the customer doesn't repeat themselves.
No. Hiding or refusing an explicit human request is a recognised dark pattern that damages trust and CSAT, and consumer-protection regulators have criticised obstruction of human contact. If no agent is available — for instance outside business hours — the right pattern is to say so honestly, collect contact details, create a priority ticket and commit to a response time.